0% ont trouvé ce document utile (0 vote)

26 vues248 pages

These

La thèse de Florian Steinhilber, soutenue le 01/02/2024 à l'INSA de Lyon, explore l'utilisation de la tomographie par rayons X pour étudier l'influence de l'intégrité de surface sur les propriétés de fatigue du Ti64 fabriqué par ajout. Le jury était composé de plusieurs experts du domaine, et la thèse a été soumise à un contrôle de similitude pour garantir son originalité. Le document inclut également des remerciements à ceux qui ont soutenu l'auteur tout au long de sa recherche.

Transféré par

jeanlucdjobo4

Nous prenons très au sérieux les droits relatifs au contenu. Si vous pensez qu’il s’agit de votre contenu, signalez une atteinte au droit d’auteur ici.

Formats disponibles

Téléchargez aux formats PDF, TXT ou lisez en ligne sur Scribd

0% ont trouvé ce document utile (0 vote)

26 vues248 pages

These

Transféré par

jeanlucdjobo4

Nous prenons très au sérieux les droits relatifs au contenu. Si vous pensez qu’il s’agit de votre contenu, signalez une atteinte au droit d’auteur ici.

Formats disponibles

Téléchargez aux formats PDF, TXT ou lisez en ligne sur Scribd

No d’ordre NNT : 2024ISAL0014

THÈSE de DOCTORAT DE L’INSA DE LYON

membre de l’Université de Lyon

École doctorale N°34

Matériaux de Lyon

Spécialité de doctorat : Matériaux

Soutenue publiquement le 01/02/2024, par :

Florian Steinhilber

On the use of X-ray computed tomography to investigate

the influence of surface integrity on the fatigue properties

of additively manufactured Ti64

Devant le jury composé de :

Fabien Szmytka Professeur associé ENSTA Paris Rapporteur

Sabine Rolland du Roscoat Maître de Conférences 3SR Rapporteur

Stefano Beretta Professeur PoliMI Examinateur

Jean-Yves Buffière Professeur INSA LYON Directeur de thèse

Rémy Dendievel Professeur Grenoble INP Co-encadrant

Guilhem Martin Maître de Conférences Grenoble INP Co-encadrant

Victor Chastand Ingénieur Dassault Aviation Invité

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Département de la Formation par la Recherche
et des Études Doctorales (FEDORA)
Bâtiment INSA direction, 1er étage
37, av. J. Capelle
69621 Villeurbanne Cédex
[email protected]

Référence : TH1077_STEINHILBER Florian

L’INSA Lyon a mis en place une procédure de contrôle systématique via un outil de
détection de similitudes (logiciel Compilatio). Après le dépôt du manuscrit de thèse,
celui-ci est analysé par l’outil. Pour tout taux de similarité supérieur à 10%, le manuscrit
est vérifié par l’équipe de FEDORA. Il s’agit notamment d’exclure les auto-citations, à
condition qu’elles soient correctement référencées avec citation expresse dans le
manuscrit.

Par ce document, il est attesté que ce manuscrit, dans la forme communiquée par la
personne doctorante à l’INSA Lyon, satisfait aux exigences de l’Etablissement concernant
le taux maximal de similitude admissible.

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

SIGLE ECOLE DOCTORALE NOM ET COORDONNEES DU RESPONSABLE

CHIMIE DE LYON M. Stéphane DANIELE

C2P2-CPE LYON-UMR 5265
E D 206 https://www.edchimie-lyon.fr Bâtiment F308, BP 2077
Sec. : Renée EL MELHEM 43 Boulevard du 11 novembre 1918
CHIMIE Bât. Blaise PASCAL, 3e étage 69616 Villeurbanne
[email protected] [email protected]

ÉVOLUTION, ÉCOSYSTÈME, MICROBIOLOGIE, MODÉLISATION Mme Sandrine CHARLES

Université Claude Bernard Lyon 1
http://e2m2.universite-lyon.fr UFR Biosciences
E D 341
Sec. : Bénédicte LANZA Bâtiment Mendel
E2M2 Bât. Atrium, UCB Lyon 1 43, boulevard du 11 Novembre 1918
Tél : 04.72.44.83.62 69622 Villeurbanne CEDEX
[email protected] [email protected]

INTERDISCIPLINAIRE SCIENCES-SANTÉ Mme Sylvie RICARD-BLUM

Laboratoire ICBMS - UMR 5246 CNRS - Université Lyon 1
http://ediss.universite-lyon.fr Bâtiment Raulin - 2ème étage Nord
E D 205
Sec. : Bénédicte LANZA 43 Boulevard du 11 novembre 1918
EDISS Bât. Atrium, UCB Lyon 1 69622 Villeurbanne Cedex
Tél : 04.72.44.83.62 Tél : +33(0)4 72 44 82 32
[email protected] [email protected]

MATÉRIAUX DE LYON M. Stéphane BENAYOUN

Ecole Centrale de Lyon
http://ed34.universite-lyon.fr Laboratoire LTDS
E D 34
Sec. : Yann DE ORDENANA 36 avenue Guy de Collongue
EDML Tél : 04.72.18.62.44 69134 Ecully CEDEX
[email protected] Tél : 04.72.18.64.37
[email protected]

ÉLECTRONIQUE, ÉLECTROTECHNIQUE, AUTOMATIQUE M. Philippe DELACHARTRE

https://edeea.universite-lyon.fr INSA LYON
E D 160 Sec. : Philomène TRECOURT Laboratoire CREATIS
Bâtiment Direction INSA Lyon Bâtiment Blaise Pascal, 7 avenue Jean Capelle
EEA Tél : 04.72.43.71.70 69621 Villeurbanne CEDEX
Tél : 04.72.43.88.63
[email protected] [email protected]

INFORMATIQUE ET MATHÉMATIQUES M. Hamamache KHEDDOUCI

Université Claude Bernard Lyon 1
E D 512 http://edinfomaths.universite-lyon.fr Bât. Nautibus
Sec. : Renée EL MELHEM 43, Boulevard du 11 novembre 1918
INFOMATHS Bât. Blaise PASCAL, 3e étage 69 622 V illeurbanne Cedex France
Tél : 04.72.43.80.46 Tél : 04.72.44.83.69
[email protected] [email protected]

MÉCANIQUE, ÉNERGÉTIQUE, GÉNIE CIVIL, ACOUSTIQUE M. Etienne PARIZET

INSA Lyon
http://edmega.universite-lyon.fr Laboratoire LVA
E D 162
Sec. : Philomène TRECOURT Bâtiment St. Exupéry
MEGA Tél : 04.72.43.71.70 25 bis av. Jean Capelle
Bâtiment Direction INSA Lyon 69621 Villeurbanne CEDEX
[email protected] [email protected]

ScSo1 M. Bruno MILLY (INSA : J.Y. TOUSSAINT)

Univ. Lyon 2 Campus Berges du Rhône
E D 483 https://edsciencessociales.universite-lyon.fr 18, quai Claude Bernard
Sec. : Mélina FAVETON 69365 LYON CEDEX 07
ScSo Tél : 04.78.69.77.79 Bureau BEL 319
[email protected] [email protected]

1
ScSo : Histoire, Géographie, Aménagement, Urbanisme, Archéologie, Science politique, Sociologie, Anthropologie

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Cette thèse aura représenté pour moi, au-delà du projet scientiﬁque, trois années importantes
et une formidable source d’évolution sur le plan personnel. A toutes les personnes qui ont rendu
cette expérience unique, riche en apprentissages et un très beau souvenir pour toute la suite, je
tiens à leur dire de tout cœur merci.

Merci d’abord à mes formidables encadrants de thèse qui m’ont accompagné tout du long,
et qui m’ont aidé à faire mes premiers pas dans ce beau domaine qu’est la recherche. Merci de
m’avoir soutenu dans mes démarches et les voies que j’ai voulu emprunter, qui devaient pourtant
vous sembler assez obscures par moment... tout en m’aidant à me replacer parfois pour m’éviter
de m’égarer dans l’infinité des possibles.
Mon choix de faire cette thèse s’est décidé je pense que tu t’en souviens Jean-Yves dans
les combles d’une maison à la campagne en plein confinement Covid-19, pendant un échange
par visio. Quelques dizaines de minutes, cela aura été suffisant pour voir toute la simplicité,
bienveillance et autres qualités qui font de toi un excellent encadrant de thèse, en plus d’une belle
personne qui est source d’inspiration et de bonne humeur pour tout le monde. Cette simplicité
aura eu raison d’un bout de mon trop de sérieux et de rigidité, et ce n’est pas peu dire ! Merci
pour tous les échanges que l’on a pu avoir, au coin d’un tableau blanc ou d’une tasse de chicorée.
Guilhem, je pense que j’ai rarement eu l’occasion de travailler avec une personne aussi fiable,
constante et efficace (pense à prendre du temps pour toi aussi !). Que ce soit lors de notre
mésaventure sur les superalliages base Nickel ou plus récemment, tu as toujours répondu présent
et proposé ton aide quand il y en avait besoin... alors que ton emploi du temps bien rempli aurait
facilement pu t’en dissuader. Merci aussi pour tes retours détaillés et toujours pertinents, qui
ont forgé aussi mes qualités scientifiques durant ces trois années.
Rémy, cela fait un moment maintenant que l’on se connait, et j’ai apprécié de t’avoir autant en
tant que professeur (et responsable du département SIM :)), encadrant de projet qu’encadrant de
thèse. À y réfléchir, mes questionnements qui frôlaient parfois la métaphysique des matériaux
ont peut-être dû te faire poser quelques questions ou avoir des sueurs froides lorsque tu as
appris que tu allais m’encadrer pour trois ans au complet :) En espérant que ma ténacité qui
frôlait parfois l’entêtement n’aura pas dépassé votre limite d’endurance, et qu’au final ces trois
années ont été pour toi une bonne expérience comme elles l’ont été pour moi. Tes jeux de
mot aussi subtils que bien placés ont toujours su créer de beaux moments lors de nos réunions
d’avancement, desquelles je garde de très bons souvenirs. Vous formez une très belle équipe, qui
j’espère aura l’occasion de vivre encore d’autres aventures au cœur des mystères de la fatigue...

Remerciements | v

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Une thèse étant loin d’être une aventure solitaire, un grand nombre de personnes m’ont
soutenu et aidé à porter ce travail bien plus loin que je ne l’aurais pu seul. Au-delà des aspects
technique et scientifique, je suis heureux d’avoir pu échanger avec chacun de vous le long du
chemin.
Merci Quentin, cela a réellement été une chance de pouvoir travailler en complémentarité avec
toi. D’abord pour tes grandes qualités humaines ; à la fois posé et attentionné, tu as vraiment le
cœur sur la main. Ta rigueur scientifique, ton esprit d’initiative et ta capacité à marier milieux
académique et industriel ont été source d’inspiration et je pense autant de clés de réussite pour
la suite de ton parcours. Au plaisir de se refaire une randonnée au cœur du Pilat :)
Joël, je me demande bien comment je m’en serais sorti au beau milieu de mes lignes de code
si tu n’avais pas été là ! Merci pour tes explications, ta patience et les débogages sur lesquels
tu m’as aidé... pour certains c’est sûr que je n’en serais pas arrivé à bout seul ! Tu peux être
fier d’avoir initié un fervent utilisateur de Python :) Linux, ça reste à voir ;) Au-delà de l’aspect
technique, merci pour nos discussions philosophiques sur la société et tout le reste, toujours de
beaux moments source de réflexions.
Je ne pourrais pas parler de programmation sans mentionner aussi l’aide que tu m’as apporté,
David. Sans elle la mesure de courbure 3D serait peut-être restée une perspective lointaine, et
m’aurait sinon donné bien du fil à retordre ! Cela fait plaisir d’avoir pu faire cette collaboration
; un bon exemple de comment partager son savoir peut aider et éviter bien difficultés à une
personne qui a encore tout à apprendre dans le domaine !
Merci également sur cette thématique à Sylvain et Pierre-Antoine, qui m’ont apporté leur
aide et répondu à mes questions pour faire mes débuts dans le monde de la programmation.
Si coder a eu sa part de défis tout le long de ma thèse, je crois qu’on peut dire que les essais
de fatigue ne sont pas en reste non plus. Merci Justine Papillon, Théophile ou encore Pascal
Reynaud pour m’avoir aidé sur ce chemin. Merci Justine Papillon et Jérôme aussi pour votre
aide en tomographie, pour toute votre gentillesse et écoute, vous formez une super équipe et
l’équipe métal est bien chanceuse de vous avoir !

Plus largement, j’aimerais remercier l’ensemble des personnes que j’ai pu rencontrer au lab-
oratoire MatéIS, l’un des endroits les plus inspirants où il m’a été donné de travailler. Merci à
tous mes cobureaux : Masato et Amin, Sébastien et Maureen, Louis Lesage, Justine Taurines
et Théophile... Merci pour l’ambiance inspirante et la bonne humeur pendant ces trois années,
qui m’aidaient quand j’étais trop pris par les diﬃcultés techniques et que je pouvais en perdre
de vue l’essentiel. Une pensée en particulier à Louis H. avec qui j’ai eu la chance de partager le
même maître/directeur de thèse et la période de rédaction, avec aussi des expériences uniques
au synchrotron... preuves d’ailleurs que c’est tout à fait possible de bien vivre ses expériences
et de passer de très bons moments même quand les résultats ne sont pas au rendez-vous :) En
parlant de synchrotron, merci au passage à James Ball et Thomas Connolley pour l’opportunité
de faire de la 3D DRX et de la DCT à l’ESRF, ainsi que pour m’avoir proposé plusieurs scans
de tomographie X à Diamond.
Pour en revenir au laboratoire, merci à Xavier, Sophie et Sylvain pour vos discussions sur le
projet Aéroprint ; autant de personnes avec qui j’ai eu beaucoup de plaisir à partager pendant
ce projet et dont j’ai aussi apprécié toutes les qualités scientiﬁques. Merci à tous les doctorants,
post-doctorants, stagiaires, permanents pour les moments de partage et d’entraide, les parties
de babyfoot ou de tennis, les 16h de vendredi, les team buildings, la période aux algécos qui

vi | Remerciements

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
n’était finalement pas si mal que ça... : merci Camille, Hélène notre super gestionnaire, Samuel,
Adam, Julien, Théotime, Louis Lesage, Bassem, Mathilde, Quentin Saby, Laurabelle, Arnaud,
Maël, Michel, Laura, Patrice, Stéphanie, Eric, Carole... et bien d’autres encore !
Merci également aux membres du GPM2 pour leur accueil et leur aide lors de mes quelques
visites, je pense notamment à Charles et Mathis pour leur aide sur les essais de traction. Merci
enfin à tous les autres acteurs de cette thèse : Victor, Charles et Christopher pour les échanges
et les impressions d’échantillons, tous les autres collaborateurs du projet Aéroprint, les mem-
bres de mon jury de thèse pour s’être plongé dans mes travaux et pour les discussions qui en
ont suivi, et finalement Dassault Aviation et la région AURA pour le financement de ces travaux.

Ces trois années de thèse n’auraient évidemment pas été aussi riches et vivantes sans la
présence de toutes les personnes qui m’ont entouré au-delà du travail. Merci à ma famille
pour leur soutien, ainsi que pour tout ce qu’ils m’ont apporté et qui m’a permis d’en arriver là
aujourd’hui.
Merci à tous les amis de la Protection Civile : Stéphanie, Timon, Paul, Zachary, Lounès,
Clotilde, Thibault, Dimitri, Marc, Grégoire, Christopher... la liste est encore longue ! Cette
année et quelques avec vous aura été une expérience aussi intense qu’enrichissante, stimulante,
faite de partages et de belles rencontres, de galères et d’entraide... le complément idéal pour
redescendre les deux pieds sur terre dans les périodes d’intense réﬂexion, de code ou de rédaction
:)
Un profond merci enﬁn à l’association UCM et toutes les personnes qu’elle m’a permis de ren-
contrer. Je n’aurais probablement pas été capable de commencer cette thèse sans toute l’aide,
les compréhensions, le soutien et l’inspiration que vous m’avez apporté et qui continuent de
m’animer aujourd’hui. Merci à tous les professeurs qui ont su m’accompagner avec bienveillance
et profondeur : Anthony, Thibaut, Christiane, Alexis, Guillaume... Merci aussi pour les mo-
ments de partage avec tous les amis d’UCM en France, qui sont une profonde source de remise
en question, d’inspiration, de joie et de soutien dans les moments de grand questionnement.

Un beau merci à tous

Remerciements | vii

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Manufacturing defects are known to significantly impact the fatigue properties of additively
manufactured (AM) components. Notably, the large surface roughness, typical of AM processes,
leads to increased stress concentrations that promote crack initiation, thereby reducing fatigue
resistance.
Traditionally, surface roughness and related defects are evaluated using tools like white light
interferometers. However, these instruments offer a limited, single-perspective and only partially
three-dimensional analysis. Those limitations do not enable the thorough characterization of the
complex surfaces and hidden defects typical in AM components.
This study describes a methodology for performing a 3D surface analysis using X-ray Com-
puted Tomography (XCT) data. The method is illustrated on various samples, ranging from
simple cylinders to more intricate architected structures. It turns out to be very efficient at
detecting critical surface defects, such as notches hidden by partially melted powder particles.
The methodology is then applied to examine the effect of surface defects on the fatigue
properties of Ti64 produced by Laser Powder Bed Fusion (L-PBF). This analysis includes both
as-built surfaces and those subjected to post-treatments, specifically investigating the impact
of Plasma electrolytic Polishing (PeP) and surface oxygen contamination (presence of an α-case
layer) resulting from high-temperature heat treatment (860 °C).
Using XCT for 3D characterization, defects responsible for fatigue failure are identified,
the latter being predominantly surface valleys. The method’s ability to predict crack initiation
locations is also evaluated, as well as its potential to estimate the fatigue resistance of a specimen
before testing.

Summary | ix

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Les défauts de fabrication sont connus pour avoir un impact significatif sur les propriétés
de fatigue des composants élaborés par Fabrication Additive (FA). En particulier, l’importante
rugosité de surface, typique des procédés de FA, génère des concentrations de contrainte locales
qui favorisent l’amorçage de fissures, réduisant ainsi la résistance à la fatigue.
Traditionnellement, la rugosité de surface et les défauts associés sont caractérisés à l’aide
d’outils tels que les interféromètres à lumière blanche. Toutefois, ces instruments offrent une
analyse qui n’est que partiellement tridimensionnelle car limitée à une unique perspective.
Ces limitations ne permettent pas une caractérisation approfondie des surfaces de composants
élaborés par FA, qui peuvent avoir des géométries 3D complexes et des défauts de surface cachés
derrière un importante rugosité.
Ce manuscrit décrit une méthodologie permettant d’effectuer une analyse de surface en 3D à
partir d’une caractérisation par tomographie aux rayons X. La méthode est illustrée sur divers
échantillons, allant de simples cylindres à des structures architecturées plus complexes. Elle
s’avère efficace pour détecter les défauts de surface critiques, tels que les entailles cachées par
des particules de poudre partiellement fondues.
La méthodologie est ensuite appliquée pour étudier l’impact des défauts de surface sur la
tenue fatigue du Ti64 élaboré par fusion laser sur lit de poudre. Cette analyse porte à la
fois sur les surfaces brutes de fabrication ou après différents post-traitements, en étudiant en
particulier l’impact d’un polissage électrolytique plasma et de la contamination par l’oxygène
de la surface (présence d’une couche d’α-case) résultant d’un traitement thermique à haute
température (860 °C).
L’utilisation de la tomographie aux rayons X pour la caractérisation 3D des défauts a permis
d’identifier ceux responsables de la rupture par fatigue, ces derniers étant principalement des
vallées en surface. La capacité de la méthode à prédire le site d’amorçage des fissures est
également évaluée, ainsi que son potentiel à estimer la résistance à la fatigue d’un spécimen
avant essai.

Résumé | xi

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Remerciements v

Summary ix

Résumé xi

Contents xiii

Introduction 1

1 Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties 7

1.1 An introduction to fatigue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Defects inherited from L-PBF: application to Ti64 . . . . . . . . . . . . . . . . . 11
1.2.1 A classification of defects based on their spatial distribution . . . . . . . . 12
1.2.2 Surface defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.3 Internal defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Impact of different defects on fatigue properties . . . . . . . . . . . . . . . . . . . 16
1.3.1 Identifying the defects at the origin of crack initiation . . . . . . . . . . . 16
1.3.2 A quantitative comparison of the impact of surface vs internal defects . . 19
1.4 Modeling the impact of defects on fatigue properties . . . . . . . . . . . . . . . . 23
1.4.1 A brief exploration of theory: Kt , Kf , K... which difference? . . . . . . . 23
1.4.2 Estimation of defects harmfulness in the case of L-PBF Ti64 . . . . . . . 27
1.5 Towards a 3D local characterization of AM surfaces . . . . . . . . . . . . . . . . 34
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2 A methodology for the 3D characterization of surfaces using X-ray computed

tomography: application to additively manufactured parts 39
2.1 Materials and XCT data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2 3D surface characterization methodology . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.1 Surface segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 3D roughness calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.3 3D curvature calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2.4 Quantiﬁcation of the harmfulness of surface notches . . . . . . . . . . . . 54
2.3 Application to parts with complex geometries . . . . . . . . . . . . . . . . . . . . 56
2.3.1 Gyroid structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.3.2 Octet-truss lattice structure . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Contents | xiii

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
2.3.3 Impeller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4 Application to the prediction of crack initiation sites in fatigue . . . . . . . . . . 62
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3 Materials characterization 71
3.1 Samples manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.1.1 Processing conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.1.2 Fatigue and tensile specimens . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2 Microstructure characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.1 Microstructure of the bulk material . . . . . . . . . . . . . . . . . . . . . . 76
3.2.2 Microstructure near the surface . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 Tensile properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4 Defects characterization using XCT . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4.1 Acquisition setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4.2 Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4.3 Surface defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

4 Fatigue properties and crack initiation mechanisms of as-built and post-treated

specimens 103
4.1 Fatigue properties and crack initiation mechanisms for as-built specimens . . . . 104
4.1.1 Protocol for fatigue tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.2 Results of fatigue tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.3 Killer defects identiﬁcation: the need for a 3D characterization . . . . . . 111
4.2 Alpha-case formation and surface embrittlement induced by heat treatments . . . 120
4.2.1 Alpha-case formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.2 Impact on tensile properties . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.3 Impact on fatigue properties . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.3 Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) . . . . 131
4.3.1 Principle of PeP surface treatment . . . . . . . . . . . . . . . . . . . . . . 131
4.3.2 Conditions of PeP treatments applied to fatigue specimens . . . . . . . . 133
4.3.3 Roughness of specimens polished by PeP . . . . . . . . . . . . . . . . . . 138
4.3.4 Fatigue resistance improvement and change in crack initiation mechanism 146
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

5 Quantitative analysis of surface defects harmfulness 153

5.1 Surface defects segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.1.1 Detailed methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.1.2 Segmentation of valleys and pit-like defects . . . . . . . . . . . . . . . . . 158
5.1.3 Segmentation of protrusions/peaks . . . . . . . . . . . . . . . . . . . . . . 160
5.2 Relationships between fatigue resistance and parameters derived from the surface
characterization using XCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.2.1 Parameters selected to reﬂect surface defects harmfulness . . . . . . . . . 161
5.2.2 Correlation between the selected parameters and the fatigue lifetime . . . 163
5.3 Killer defect prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.3.1 Application to surface-treated specimens (PeP) . . . . . . . . . . . . . . . 169
5.3.2 Application to specimens with as-built surfaces . . . . . . . . . . . . . . . 178

xiv | CONTENTS

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Conclusion and Perspectives 185

A Comparison of diﬀerent distance calculation methods 197

B 3D Gaussian S-Filter using normalized convolution 203

C Near-surface microstructure after PeP treatment 207

References 211

CONTENTS | xv

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Additive Manufacturing (AM) includes many processes allowing to manufacture components

layer by layer from their Computer Aided Design (CAD). Among those, Laser Powder Bed
Fusion (L-PBF) is one of the most common processes to produce metallic components [VAF 21].
AM of metallic components has recently received increasing interest. The aerospace field is
one of the main drivers in this area and represented about 18.2% of the global additive manufac-
turing market in 2017 [NAJ 19]. Although its large-scale adoption requires further developments
in numerous fields and long qualification and certification processes [SEI 17], several successful
industrial applications have already been reported [THO 19, NAJ 19, Gen 18, Lie 17]. For in-
stance, Boeing already started in 2017 using at least four additively manufactured titanium
structural components for its 787 Dreamliner commercial aircraft [MEA 17]. Fig.1a-b show two
other examples of additively manufactured aircraft components used by General Electric (GE)
and Airbus.

Figure 1 | Examples of industrial components manufactured by AM. (a) Fuel nozzle

tip produced by GE for its LEAP engine [Gen 18]. GE had already produced such parts in more
than 30 000 units in 2018. AM enabled the reduction of the weight of the part by 25% compared
to its equivalent produced by conventional processes. Moreover, AM allowed the fabrication of
this component in a single part, whereas its predecessor was made of about 20 parts welded
together. (b) Additively manufactured valve block for a spoiler actuator made of titanium,
tested in ﬂight in an Airbus A380 airplane in 2017 [Lie 17]. The use of AM also enabled the
reduction of the component weight as well as alleviating assembly steps.

Introduction | 1

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
AM processes offer several advantages that explain their increasing popularity, even though
not all industrial components can be advantageously replaced by an additively manufactured
counterpart. Metallic AM is therefore often best suited for the manufacturing of high added-
value complex components produced in small series [THO 19, ATZ 12]. In such cases, some of
the advantages of AM processes are:
• A reduction of raw material usage, since AM parts are printed near-net-shape and require
only limited machining operations.
• A reduction of the component weight, which is always a crucial issue in the aerospace
industry (see the examples given in Fig.1a-b).
• The opportunity to simplify assemblies and avoid costly operations such as welding, which
can also affect the durability of the material (Fig.1a).
• The possibility of having more efficient supply chains, e.g. by producing spare parts on
demand rather than storing them over long periods.
However, some issues are still to be overcome to deploy metallic AM at a larger scale. Among
those, AM components are still prone to different manufacturing defects such as pores or an im-
portant surface roughness [CAB 18] associated with various surface defects, see e.g. Fig.2a-b.
Notch-like surface defects are known to strongly affect the mechanical resistance under cyclic
loading (i.e. fatigue) [PER 19, PLE 20a]. Yet, the fatigue resistance is a key factor in many
industrial applications. This is well illustrated in the review of [FIN 02], where 55% of the
aircraft components failures were found to be caused by fatigue. One of the challenges of AM
is therefore to better control these defects, understand their impact on fatigue properties, and
if possible, eliminate them.

Figure 2 | Surfaces inherited from AM processes. (a) Surface of a Ti64 thin strut
fabricated by Electron Powder Bed Fusion (E-PBF), adapted from [PER 19]. The surface is
reconstructed from an X-ray Computed Tomography (XCT) scan. (b) The surface of a Ti64
sample manufactured by Laser Powder Bed Fusion (L-PBF), observed using Scanning Electron
Microscopy (SEM) [NAK 19].

2 | Introduction

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The removal of surface defects by surface finishing methods is not a new topic and different
solutions can be considered. However, these surface finishing methods were mostly designed for
parts with simple shapes and/or limited surface roughness, typically inherited from machining.
On the opposite, AM components often show both complex geometries and high surface rough-
ness. Only a few surface finishing processes are possible in this case, and all have their own
drawbacks or limitations. For instance, chemical polishing was proven to give interesting results
[BAS 22, PER 20, BEZ 20], but requires the use of HF-based chemical solutions for Ti alloys
[BAS 22], causing health and safety issues. All post-processing operations that may be required
to remove and control defects from AM components can also represent an additional cost, reach-
ing sometimes nearly half of the price of a component [THO 19]. Some of those surface finishing
treatments are time-consuming, which may completely cancel the benefits of AM.
Post-processing operations should thus be avoided as much as possible. This means that
it would be advantageous to use, whenever possible, components with their as-built surface to
avoid expensive surface finishing operations. Doing so requires a good knowledge of the effect
of surface defects on the mechanical properties, in particular fatigue.
One of the defects characterization techniques that has recently drawn an increasing interest
is X-ray Computed Tomography (XCT) which shows several advantages. For example, it is the
only Non-Destructive Technique (NDT) that is sensitive to defects regardless of their location,
i.e. at the surface, in sub-subsurface or in the bulk material, see Fig.3a. Other techniques
such as white light interferometry only allow for surface characterization, whereas eddy current
or ultrasonic testing are sensitive to sub-surface defects only. XCT is also the most suitable
technique to characterize the surface of components with complex geometries (e.g. internal
channels), a key characteristic of AM components that often show complex geometries.

Figure 3 | (a) Comparison of some Non-Destructive Techniques (NDT), including XCT, in

terms of achievable spatial resolution and ability to capture defects at diﬀerent locations. From
[VIL 19]. (b) Typical spatial resolutions achievable by X-ray Computed Tomography (CT in
the ﬁgure) as a function of sample size. From [VIL 18a].

Introduction | 3

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Regarding spatial resolution, XCT provides generally significantly inferior resolutions than
most advanced surface characterization techniques such as white light interferometry (Fig.3a).
Nonetheless, achievable resolutions have been improved over the years and sub-micron resolu-
tions can now be obtained, even with commercial XCT scanners [KAS 17], provided that the
scanned object is small enough (Fig.3b). These recent advances make it an interesting option
to gain a further understanding of defects in AM components. However, the use of XCT for the
study of surface roughness is still limited, and further work is needed to take advantage of the
opportunities offered by XCT, especially for additively manufactured parts.
The aim of this PhD work is to develop and apply XCT techniques to investigate the effect of
surface integrity on the fatigue properties of additively manufactured components. This research
is part of the AEROPRINT research project, financially supported by the Auvergne-Rhône-
Alpes Region and Dassault Aviation (France). AEROPRINT’s primary goal is to promote the
industrial application of metallic AM in the aerospace sector. A pre-industrial demonstrator is
developed in Dassault Aviation, with the possibility to produce titanium and aluminum parts by
Laser Powder Bed Fusion (L-PBF). The project involves collaboration between various industrial
and academic entities, including Dassault Aviation, AddUp, and the French Commissariat à
l’Énergie Atomique et aux Énergies Alternatives (CEA).
These collaborations have enabled to cover a wide range of research topics, ranging from
powder characterization and optimizing L-PBF process parameters to exploring different surface
and heat treatments. This collaborative environment has also funded several PhD studies. In
this regard, it is worth mentioning the one conducted by Quentin Gaillard [GAI 23a] that is
complementary to the present PhD work and that primarily focuses on the influence of heat
treatments on the static mechanical properties of the Ti64 alloy manufactured by L-PBF.

4 | Introduction

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
This manuscript consists of 5 chapters that can be summarized as follows:
• Chapter 1 offers a short review of the current knowledge on how manufacturing defects
affect the fatigue properties of Ti64 alloy produced by L-PBF. It evaluates the extent of
fatigue property degradation due to defects and identifies the most detrimental, which
are typically surface defects in as-built samples. It also discusses some models from the
literature that assess defect harmfulness using geometrical characterizations, e.g. from
XCT.
Chapter 2 describes in detail the 3D XCT characterization methodology used throughout
this work to analyze surface defects. This method allows for the assessment of roughness
and curvature on complex 3D surfaces typically observed in parts fabricated by additive
manufacturing. A framework to quantify the impact of surface notch-like defects on fatigue
properties is also proposed, thereby gauging the influence of the inherent rough surfaces
from the L-PBF process.
Chapter 3 introduces the Ti64 mechanical specimens fabricated by L-PBF examined
in this study. The chapter covers defect characterization (including surface defects and
porosity), microstructure, and tensile properties.
Chapter 4 deals with the fatigue properties of L-PBF Ti64 samples after various post-
treatments: heat treatments (stress-relief and at higher temperature), or a surface treat-
ment (Plasma Electrolytic Polishing, denoted PeP). The objective is to quantify the impact
of defects on fatigue properties and to identify the defect(s) responsible for failure through
detailed 3D XCT characterizations, using the methodology developed in this work and
detailed in Chapter 2.
Chapter 5 explores the potential of XCT characterization of surface defects to estimate
the fatigue resistance and predict the most critical defects that may lead to fatigue failure.
Finally, a conclusion summarizes the main findings of this work and outlines future research
directions to build upon these results.

Introduction | 5

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Brief overview on defects

harmfulness on L-PBF Ti64 fatigue
properties

Contents
1.1 An introduction to fatigue . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Defects inherited from L-PBF: application to Ti64 . . . . . . . . . . . . 11
1.2.1 A classification of defects based on their spatial distribution . . . . . . . . . 12
1.2.2 Surface defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.3 Internal defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Impact of different defects on fatigue properties . . . . . . . . . . . . . . 16
1.3.1 Identifying the defects at the origin of crack initiation . . . . . . . . . . . . 16
1.3.2 A quantitative comparison of the impact of surface vs internal defects . . . 19
1.4 Modeling the impact of defects on fatigue properties . . . . . . . . . . . 23
1.4.1 A brief exploration of theory: Kt , Kf , K... which difference? . . . . . . . . 23
1.4.2 Estimation of defects harmfulness in the case of L-PBF Ti64 . . . . . . . . 27
1.5 Towards a 3D local characterization of AM surfaces . . . . . . . . . . . 34
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

This chapter aims to provide the required insights to aid in comprehending the work con-
ducted during this PhD. It primarily focuses on the core research topic − assessing how defects
aﬀect fatigue properties of Ti-6AL-4V (Ti64) alloy manufacturing by Laser Powder Bed Fusion
(L-PBF).
Ample literature exists on this subject, providing valuable insights that guide research and
highlight areas for improvement. Nonetheless, it seems also advisable to use previous studies
with caution. Results obtained on L-PBF Ti64 exhibit variability − e.g. due to the large choice
of machines and process parameters that have been used − and improvements in the L-PBF
technology have occurred over the years. Consequently, conclusions from earlier studies might
not directly apply today.

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Section 1.1 explains fatigue damage and the significant role of defects in fatigue properties.
Section 1.2 outlines key defects in Ti64 manufactured by L-PBF. In Section 1.3, the impact of
these defects on fatigue properties is discussed. Lastly, Section 1.4 presents methods proposed
to estimate defect influence on fatigue properties, in particular in the case of materials manu-
factured by L-PBF. These methods may have various purposes, e.g. predicting defect impact on
the fatigue limit (see Section 1.1 for a definition) or identifying the most critical defects within
a population.

1.1 An introduction to fatigue

Fatigue corresponds to the damage of a material caused by repeated loading. One particu-
larity of fatigue is that loading levels are often well below the yield strength of the considered
material. Nevertheless, the repetitive nature of the loading still enables the initiation and prop-
agation of cracks, ﬁnally causing irreversible damage and material failure. Fatigue is a key
property in many industrial applications such as aeronautics. This emphasizes the need for bet-
ter characterization and understanding of fatigue properties of materials, and the mechanisms
involved.
Loading sequences experienced by industrial components in real cases can be very complex.
To rationalize the study and facilitate interpretations, simple loading sequences and testing
conditions are generally used. To keep this part short, we focus only on uni-axial fatigue tests,
leaving aside torsion and bending fatigue, fretting fatigue, etc. In this case, a simple and widely
used approach to characterize fatigue properties is to apply a cyclic − e.g. sinusoidal − loading
to a specimen until failure. Repeating the same procedure for several specimens at diﬀerent
stress levels, it is possible to draw a so-called S-N curve − or Wöhler curve − which is a
plot of the applied stress as a function of the number of cycles to failure, see Fig.1.1.
The applied stress can be given in terms of maximum applied stress σmax or stress amplitude
σa “ σmax ´σ
2
min
, σmin being the minimum applied stress.

UTS
LCF

HCF
Run-out
σNf

Infinite life
log N
Nf
Figure 1.1 | Schematic of a typical Wöhler (or S-N) curve. Noteworthy stress levels
− the tensile strength UTS and the fatigue limit − are indicated on the stress axis. The three
main regions − LCF, HCF and inﬁnite life − are delimited.

8 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Three main regions can be distinguished on a Wöhler curve:
1. The Low Cycle Fatigue (LCF) regime, where the applied stress is sufficient to cause
macroscopic plastic deformation at every cycle. Resulting fatigue lives are usually lower
than 104 cycles.
2. The High Cycle Fatigue (HCF) regime, where the applied stress is significantly lower
than the yield strength of the material. Samples are loaded elastically at the macro-scale
and the measured fatigue lives are generally greater than 105 cycles.
3. The infinite life regime, where the stress is lower than the fatigue limit. In this case,
failure is considered to never occur. Strictly speaking, studies have actually shown that
such a limit does not exist in general [BAT 99], meaning that no plateau is visible on the
S-N curve. The concept of fatigue strength σN can therefore be used as a substitute. It
corresponds to the highest stress level for which no failure is observed until a pre-defined
number of cycles N − typically 2 ¨ 106 or 107 . Unbroken specimens are said to be run-outs.

Each Wöhler curve is measured under given conditions for the fatigue tests. These conditions
can have a strong impact on the results. The main parameters describing a fatigue test and that
have to be taken into account are the following:
σmin
• The stress ratio R “ σmax
• The mean stress σm “ σmin `σ
2
max

• The frequency of applied loading f

• The environment: atmosphere, temperature, etc.
• The shape, size and surface roughness of fatigue specimens

Fig.1.2 summarizes the deﬁnition of sigmamin , σmax , σa , σm and f .

stress

1/f
σmax
σa
σm time

σmin

Figure 1.2 | Graphical deﬁnition of minimum and maximum and mean stresses during fatigue
cycling σmin , σmax , and σm , of the stress amplitude σa and the frequency f .

An introduction to fatigue | 9

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Characterizing a material by its Wöhler curve provides relevant information, e.g. to design
industrial parts accordingly. However, to better understand the fatigue properties of a material,
it is needed to go deeper and look at the micro-mechanisms causing fatigue failure.
Broadly speaking, fatigue damage occurs in two main steps: crack initiation and
crack propagation, see Fig.1.3. At the end of these two steps occurs the ﬁnal sudden failure of
the material. Naturally, several cracks may coexist within a material and it is therefore possible
for one to be in the initiation phase while another would already be in the propagation phase.

Cyclic Crack Micro-crack Macro-crack Final

slip nucleation growth growth failure

Initiation Crack growth

period period
Figure 1.3 | Phases in fatigue life , adapted from [SCH 09, p.15].

The crack initiation period can be split into several steps. It begins with local cyclic slip,
i.e. local plastic deformation at each loading cycle. The regions where cyclic slip preferentially
occurs depend on the local stress field, which depends itself on the defects that act as stress
concentrators. It is also influenced by the local microstructure since cyclic slip will be facil-
itated in grains with slip systems oriented unfavorably with respect to the loading direction
[TUB 23, BOO 16, WU 23]. Cyclic slip eventually leads to the nucleation of a micro-crack,
which will continue to grow under cyclic loading. At this step, crack growth is however still de-
pendent on the environment of the initiation site − e.g. the local microstructure and stress field.
Grain boundaries, or α{β interfaces in Ti alloys, can for instance act as barriers for micro-crack
propagation. Depending on the local stress gradient, the crack may slow down at these obstacles
or even stop definitely. For these reasons, micro-crack growth is still considered to be part of
the initiation process. Only when crack growth becomes dependent on bulk properties alone, it
can be considered a macro-crack.
Nonetheless, the limit between micro-crack and macro-crack propagation − or in other words,
between crack initiation and growth − is highly dependent on the material studied. For instance,
for a given material without any defect, the majority of the fatigue life may be spent to initiate
the crack − see Fig.1.4. It might however be very different if defects such as inclusions, pores or
surface scratches are present [ZER 19]. In the particular case of Additive manufacturing (AM),
typical defects include pores, lacks of fusion or surface valleys − see next section for more details.
Although they are different in nature, they will all in a similar manner act as stress concentration
raisers and thus promote local plastic deformation. As a consequence, crack initiation may occur
at the very beginning of cycling, bypassing a wide percentage of the fatigue life that was spent to
initiate a crack. As a result, the fatigue strength will be drastically reduced. Thus, defects can
have a strong impact on fatigue strength by facilitating the crack initiation stage.

10 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

starting from
log (crack length) defect
micro-cracks

starting from
no defect
nucleation

0 % of fatigue life 100

Figure 1.4 | Diﬀerent scenarios of fatigue crack growth depending on whether crack
initiation occurs at a defect or not. If a suﬃciently large defect is present, almost all fatigue life
corresponds to macro-crack propagation, while it is the inverse in the case where no defect is
present. Adapted from [SCH 09, p.22].

1.2 Defects inherited from L-PBF: application to Ti64

A wide variety of defects can arise during additive manufacturing of metallic materials. Some
of them are harmless regarding fatigue properties, while others are known to be of major concern.
These defects range from dimensional deviation to cracks, pores and lacks of fusion to surface
valleys and unmelted powder particles [SAN 21]. The scope of this review is limited to the main
ones found in Ti64 manufactured by Laser Powder Bed Fusion (L-PBF). Furthermore, it focuses
on defects inherent in the L-PBF process, albeit optimized. Therefore, it does not consider those
that can be attributed to misuse of the process or an accidental event during manufacturing.

Defects inherited from L-PBF: application to Ti64 | 11

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.2.1 A classification of defects based on their spatial distribution
Surface defects vs. internal defects
A first method to classify defects is to distinguish surface defects from internal ones. The
two are very different regarding their location and shape, the means to characterize them, the
post-treatments to remove them or their impact on mechanical properties.

Sub-surface defects vs internal defects

Among the internal defects, we can also distinguish sub-surface defects, present in the out-
ermost layer of the material, from internal defects, located outside this area. They are similar
in terms of morphology and physical origin, but higher concentrations of pores are often
observed in the sub-surface region [BER 17]. This inhomogeneity can arise due to an ex-
cess of energy density near the surface which can lead to the formation of keyhole pores − see
Section 1.2.3 for details about keyhole pores. This excess energy can for example come from a
bad process parameters adjustment for the contour scan, an extra laser pass that is added at
the contour of the printed part, among others to improve surface roughness.
Thus, Fig.1.5 adapted from [KAR 23] shows that a reduction of laser scanning speed (and
therefore an increase in energy density) for the contour scan results in an increase in sub-surface
pore density. In a more subtle way, sub-surface pores can come from the deceleration and
acceleration of the laser beam at turning points (at the end of laser tracks). Those tend to
increase the time spent by the laser at those speciﬁc locations and therefore the local energy
density [MAR 19, AND 19a, MAN 16]. Strategies exist to alter process parameters at turning
points to compensate for this eﬀect [MAR 19, MAN 16].

Figure 1.5 | Increase in sub-surface pore density with a decrease in laser scanning
speed in the contour scan. The ﬁgure is adapted from [KAR 23]. Samples are fabricated by
L-PBF using pre-alloyed A20X aluminum powder.

Sub-surface defects are also of particular interest regarding their impact on mechanical prop-
erties. Indeed, the same defect will be much more harmful if it is at a suﬃciently
close distance from the surface than if it is in the core, as is discussed in Section 1.3.
Finally, sub-surface defects have the advantage of being the only internal defect that can be
detected by some Non-Destructive Testing (NDT) methods such as Eddy Current, or eliminated
by some surface treatments such as machining or chemical polishing. However, regarding this
last point, surface treatments involving material removal may also transform these defects into
surface defects, making them even more harmful [PER 20].

12 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.2.2 Surface defects
Surface defects in the case of L-PBF samples include:
• Partially melted powders stuck to the surface. Even if they contribute to surface
roughness, they do not have a significant impact on mechanical properties [KAN 18].
However, they may still cause other issues such as the detachment of powder grains in
hydraulic tubes [ZHA 21]. They are found to be most present in down-skin regions, while
it is much less the case in up-skin regions [PYK 13, CAB 18]. Fig.1.6 shows partially
melted powders on the surface of an L-PBF sample characterized by SEM.
• Spatters, which are particles ejected from the melt pool during manufacturing [LI 22,
ZHA 17, AHM 20, LAD 16, PAU 23, GUN 18]. When not evacuated by the airflow set
up in the build chamber, spatters contaminate the next powder layer. Since their size
is sometimes much larger than powder particles used for manufacturing, the laser may
be unable to completely melt them. This mechanism can lead to the presence of large
particles associated with lack of fusion defects at the sample surface as well as in the bulk
material [PES 20, LI 22, WAN 22]. Fig.1.7a-c show examples of fatigue failure caused by
the presence of spatters at the specimen surface.
• Dross, which is characterized by the presence of large protuberances most often in down-
skin regions. Down-skin regions refer here to the ones that are facing the build plate (and
therefore the powder bed), in opposition to up-skin regions which face away from the build
plate. Dross formation is a consequence of the low thermal conductivity of the powder bed,
which causes heat accumulation in overhangs and thus an excessive melting of powders
[CHA 22, FEN 21, LIN 23, DEP 18, VIA 22, CAL 14]. Furthermore, the melt pool tends
to sink into the powder bed in down-skin regions, which results in large protuberances.
Fig.1.8 gives an example of a sample fabricated by Electron Laser Powder Bed Fusion
(E-PBF), where dross formation can occur similarly to L-PBF.
• Valleys, which will be called equivalently notches. Valleys can have multiple origins. In
up-skin and vertical regions, they can arise due to the superposition of fused powder layers
− see Fig.1.6 for an example. The resulting surface often looks like a stack of plates or
a staircase, which is why this effect is sometimes called plate-pile like or staircase effect
[PYK 13, ZHA 19, PER 19]. Similarly, dross formation induces the presence of valleys in
down-skin regions. Regardless of their origin, these valleys generate significant stress
concentrations that are detrimental to mechanical performances, especially in
fatigue [PER 19, KAN 18, CHA 15].

Defects inherited from L-PBF: application to Ti64 | 13

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 1.6 | SEM observation of an L-PBF as-built surface highlighting partially melted
powders and valleys. Adapted from [BEN 17].

Figure 1.7 | Fatigue failure from two surface spatters in IN625 manufactured by
L-PBF [PES 20, Fig.3.28]. It can be seen that the spatter at the origin of the largest crack
caused the formation of a lack of fusion.

Figure 1.8 | XCT slice showing an example of dross formation in the down-skin region
of a 2 mm cylinder fabricated by Electron Powder Bed Fusion, adapted from [PER 19].

14 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.2.3 Internal defects
The main internal defects in L-PBF Ti64 are pores. Several types can be distinguished
according to their size, morphology and physical origin, leading to the following categories:
• Lacks of fusion (LOF) which can form when the energy density provided by the laser
source is insufficient to completely melt the powder [GUN 18, p.45]. This is for instance
the case when the hatch spacing (distance between adjacent tracks) is too large [ABO 14],
resulting in bad overlapping between adjacent laser tracks. Lacks of fusion may also occur
because of irregularities in the powder bed, spatters or exceptionally large powder particles
[DEP 18, WAN 22, PAL 18].
Lacks of fusion often present complex and elongated geometries − see Fig.1.9) − with the
major axis orthogonal to the build direction. Furthermore, they are usually larger than
other types of porosity [HU 20a]. They can also be identified by the presence of unmelted
powder particles, as can be seen in Fig.1.9.
• Keyhole porosity, formed on the contrary in cases where the energy density is too high
[BAY 19]. They are generally smaller than lack of fusion defects („60 µm according to
[VAY 20]), their morphology is often spherical [LIU 20] as in Fig.1.10a but can sometimes
be more elongated and angular as illustrated in Fig.1.10b.
• Pores originating from powders. During powder atomization, some gas can be trapped
inside the solidified powder particles [BIA 11]. Those gas pores may remain even after L-
PBF, resulting in some additional porosity. They are generally of small size („5 µm in
[VAY 20]) and have a very spherical morphology. Their quantity can be greatly reduced
by using a technique such as plasma atomization to manufacture the powders.

Figure 1.9 | Lacks of fusion in L-PBF Ti64 observed using SEM by [PAL 18]

Figure 1.10 | Keyhole porosity observation in L-PBF Ti64. (a) XCT slice showing
keyhole porosity, from [BAY 19]. Keyhole pores are visible all along the last laser track, at the
top of the image. (b) SEM image of a non-spherical keyhole pore, from [PAL 18].

Defects inherited from L-PBF: application to Ti64 | 15

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.3 Impact of different defects on fatigue properties
The defects described above have unequal impacts on fatigue properties. Some of them are
critical, while others can be considered to be harmless. This section presents the ranking of the
most harmful defects, suggested by various results from the literature. It also described several
attempts that were made to quantify their impact on fatigue properties.
A first approach to finding the most harmful defects is to identify the ones that cause crack
initiation on fractured surfaces. The more frequently a defect is found to initiate a crack, the
more harmful it is considered to be. This analysis is addressed in Section 1.3.1.
It is also possible to study the relative influences of different defects by comparing the me-
chanical properties of samples including a controlled defect population. For example, additive
manufacturing was employed to generate deterministic defects [LES 22]. More simply, it is also
possible to compare as-built samples − i.e. without any post-treatment − to others subjected
to Hot Isostatic Pressing (HIP) and/or machining and polishing. In this way, the impact of
surface and internal defects can be studied relatively independently to determine which ones are
the most harmful. This analysis is detailed in Section 1.3.2.

1.3.1 Identifying the defects at the origin of crack initiation

1.3.1.1 Prevalence of surface notches
For an as-built surface, crack initiation occurs most often at a surface notch
[KAS 15, CHA 18, CAR 19, YU 19] as shown in Fig.1.11. In the study of [CHA 18], a surface
defect is for example found to be at the origin of failure in 10 cases out of 13 for both L-PBF
and E-PBF Ti64 samples with an as-built surface. For [CAR 19, KAH 17, MIL 22], it is even
the case for 100% of the observed initiation sites.

Figure 1.11 | Fatigue crack initiation at surface notches in L-PBF Ti64 [KAS 15]
.

16 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Several reasons can explain this prevalence of crack initiation at surface defects. The results
of FEM calculations given in Fig.1.12 show for example that a surface pore generates a higher
stress concentration than a pore located in the bulk material. In the y-axis of the graph, σ is
the applied stress whereas σzz max is the maximum stress component in the direction of applied
max
stress. Therefore, the ratio σzz {σ quantiﬁes the stress concentration induced by the pore. In
the x-axis, r is the pore radius whereas d is the distance of the pore center to the surface. As
shown on the small schematics, the ratio r{pr ` dq tends to 0 for a pore deeply embedded in the
bulk material and equals 0.5 when the pore is in direct contact with the surface. Above 0.5, it
is an open pore, i.e. a spherical notch. From this graph, it can be seen that a spherical notch of
radius r (i.e. r{pr ` dq “ 1) causes the same stress concentration as a spherical pore of diameter
2r in the core material (in the case r{pr ` dq Ñ 0). In other words, a spherical surface hole causes
the same stress concentration as an internal spherical hole of twice its size.

Figure 1.12 | Evolution of stress concentration induced by a spherical pore in func-

tion of its distance to the surface. The ﬁgure is drawn from [BOR 02] who obtained these
results using FEM. σ is the applied stress whereas σzz max is the maximum stress component in

the direction of applied stress. r is the radius of the pore, whereas d is the distance of the center
of the pore to the surface. Therefore, the pore touches the surface for r{pr ` dq “ 0.5.

Another particularity of surface defects is that they are exposed to the external environment.
Conversely, internal defects are in contact with vacuum or inert gas. It has been shown that
the presence of air can have a deleterious eﬀect on the fatigue life of the Ti64 alloy and in
particular on the threshold stress intensity factor Kth , which is the minimum stress intensity
factor required to induce a signiﬁcant crack growth [XU 21, JUN 21, IRV 74].

Impact of diﬀerent defects on fatigue properties | 17

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Crack initiation on an internal defect can sometimes be seen, even on samples with an as-built
surface [LIU 14]. This was for example observed in 3 cases out of 13 in the study of [CHA 18]
on the fatigue of L-PBF and E-PBF Ti64.
In such cases, the defect is often located in the sub-surface region [LIU 14, AND 19b]. As an
example, [TAM 17] observed in Ti64 manufactured by E-PBF that for 10 out of 11 cases where
initiation was found to occur from a pore, the pore was located in the sub-surface region. Like
surface defects, sub-surface defects are therefore particularly harmful. The reasons are
also similar. FEM calculations introduced previously in Fig.1.12 show that the stress concentra-
tion induced by an internal defect increases when its distance to the surface decreases. Such a
stress concentration promotes crack initiation and can provoke the rapid failure of the ligament
that separates the pore from the surface. This creates a large defect at the surface, which will
be particularly harmful for the reasons invoked in the previous paragraph.

1.3.1.3 Comparing internal defects: lack of fusion defects vs spherical pores

When a specimen is machined, surface defects − which are the most harmful − are removed.
Consequently, the remaining internal defects become the most critical ones and are often found
to be the cause of fatigue failure [KAS 15, Fig.15c]. This effect is more pronounced for samples
with polished surfaces since in the case of machined surfaces, grooves are still present at the
surface and can promote crack initiation [YU 19].
Among all internal defects, lacks of fusion are generally identified as the most
critical ones [GON 15, PLE 20c, SAN 20, LAR 18]. This is due to their elongated shape and
larger size. For example, [CHA 18] observed in the case of machined L-PBF Ti64 specimens 25
initiations on lack of fusion defects against 12 on smaller internal defects such as spherical pores.
The effect is even more pronounced when specimens are printed with their axis parallel to the
build direction, since lacks of fusion are in this case perpendicular to the loading direction.
Similarly, by comparing the mechanical behavior of L-PBF Ti64 manufactured with different
process parameters (excessive vs insufficient energy input), [GON 15] concluded that keyhole
pores − that are small and spherical − do not significantly impact mechanical properties even
for volume fractions of the order of 1%. On the opposite, lacks of fusion generated by an
insufficient energy input − that are larger and elongated − already lead to a substantial decrease
of mechanical properties for volume fractions of 1%. This effect is visible for both tensile and
fatigue properties.
According to the study of [AND 19b] dedicated to the fatigue properties of 316L stainless
steel fabricated by L-PBF, the first-order factor remains the proximity of the defect to the
surface. The defect responsible for the failure, located in the sub-surface region, had thus in
several cases a size much smaller (up to 3.6 times) than that of the largest defect observed in
the core of the sample.
To summarize, it is finally possible to give a first qualitative classification of the types of
defects by order of increasing harmfulness in fatigue for Ti64:
Lacks of fusion > Spherical pores
Surface notches >
Sub-surface > Core

18 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.3.2 A quantitative comparison of the impact of surface vs internal defects
To estimate the relative influence of internal and surface defects more quantitatively, the
impact of the following different post-treatments can be compared:
• No Hot Isostatic Pressing (HIP) or surface modification (denoted as-built here)
• HIP only (no modification of surface defects while internal ones are healed)
• Machining/polishing (suppression of surface defects while internal ones are still present)
• Machining/Polishing + HIP (suppression of both internal and surface defects)

Since machining or polishing enables the removal of surface defects and HIP allows the
removal of internal ones, the combination of these conditions enables decorrelating the influence
of these two families of defects.
Of course, the conclusions must be nuanced for several reasons:
• Machining/polishing without HIP will bring defects that were previously located in the
sub-surface directly at the surface. Even if surface notches present in as-built surfaces are
removed, new surface defects can therefore be created [CHI 21, CAR 19].
• Machining leaves residual stresses at the surface that can have a strong influence on the
mechanical behavior. Typically, if compressive stresses are generated, they will inhibit
crack initiation at the surface.
• HIP treatments − carried out at very high temperatures („900 °C) − will cause significant
modifications in the microstructure such as broadening of α-laths, which may influence
fatigue properties.
• Not all pores are necessarily closed after HIP [PLE 20b, LI 19] and it has therefore been
observed in some cases that crack initiation occurs on a partially closed pore [CHA 17].

Nonetheless, many studies report the inﬂuence of defects on fatigue properties by comparing
the above conditions and most of them show good agreement with each other. Among others,
the work of [MAS 18] is representative of most of the published results. The fatigue performance
of the as-built, HIP, machined, and HIP+machined conditions was compared in rotary bending
fatigue tests. Comparing these diﬀerent conditions, [MAS 18] arrived at the following ranking
in terms of performance:

AB surface w/o HIP ă AB surface with HIP ! Polishing w/o HIP ! Polishing with HIP
where AB is an abbreviation for as-built.

Impact of diﬀerent defects on fatigue properties | 19

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.1.13 shows the Wöhler curves obtained by [MAS 18] in more detail.

Figure 1.13 | Wöhler curves obtained with or without polishing and HIP of L-PBF
Ti64 [MAS 18] (60Hz rotational bending tests on vertically built specimens)

Before comparing the fatigue performances of the different conditions, we can already notice
in Fig.1.13 that the differences in fatigue performance between conditions are more pronounced
for longer fatigue lives. It is in this zone that crack initiation represents the most important
fraction of fatigue life (compared to the crack propagation regime); the influence of defects
therefore tends to be exacerbated.
The Wöhler curves also generally show more scattering on machined or polished specimens
than on as-built surface specimens. This is slightly noticeable in Fig.1.13, but the effect can
be even more pronounced − see for example [WYC 13]. This difference in behavior can be
attributed to the fact that in the as-built condition, there is a very large statistical representation
of surface defects in each specimen. Conversely, on machined or polished specimens, the critical
defect density is lower and therefore there may be statistically more variation in fatigue life from
one specimen to another.
We can see that the beneficial effect of HIP is limited on specimens with an as-built
surface: the fatigue strength at 107 cycles σ107 is 155 MPa with an as-built surface without HIP
and 195 to 220 MPa in the as-built + HIP condition − which represents an improvement of
25-42%. Once again, we see the most deleterious effect of the surface defects present for
an as-built surface in comparison with the effect of the internal defects suppressed by HIP. This
effect can be partly attributed to the fact that tests are performed in rotational bending, thus
favoring crack initiation at the surface. However, the same tendency can be found in several
other studies with tests not carried out by rotary bending (see e.g. references given in Tab.1.1):
greater harmfulness of surface defects is observed.

20 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Compared with the application of HIP without any surface modification, polishing without
HIP allows a much more significant increase of σ107 from 155 MPa to 370 MPa (139% improve-
ment) because the surface defects have been removed. Finally, the application of a HIP
treatment together with surface polishing yields the best results. The first-order de-
fects once the polishing is done are indeed the pores and their elimination by HIP makes it
possible to obtain a fatigue strength of 610 MPa − i.e. an improvement of 294% compared to
polishing without HIP. For comparison with Ti64 obtained by conventional processes, [VAY 19]
measured on rolled Ti64 a fatigue limit of 676 MPa at R “ ´1, and [KAH 17] measured a fatigue
strength at 5 ˆ 106 cycles of 800 MPa at R “ 0.1 for forged Ti64.
The same trends are found by many authors, whether for fatigue tests at R “ ´1 [CHA 18,
VAY 19, KAS 15] or R “ 0.1 [WYC 13, GRE 17, KAH 17, BAG 17]. [PER 18] also came to the
same conclusions on E-PBF Ti64.
[EDW 14] on the other hand finds little influence of polishing (see Fig.1.14a), contrary to
what is generally observed in the literature. In this particular case, the poor performance in
the polished state (200 MPa at 106 cycles) could be explained by the high porosity observed −
see Fig.1.14b-c. Under such conditions, the influence of porosity could eventually become non-
negligible compared to that of the surface state. It is also possible that polishing has brought
these pores to the surface, making them even more harmful.

Figure 1.14 | (a) Wöhler curves obtained for varying surface states ("Net" = as-built) and
build directions by [EDW 14] (R “ ´0.2). (b) and (c) Optical microscope characterization of
the corresponding material, showing a high pore density [EDW 14].

Impact of diﬀerent defects on fatigue properties | 21

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To conclude this section, Tab.1.1 gathers estimations of the fatigue limit of L-PBF Ti64
found in several studies from the literature. In most cases, the fatigue limit is estimated by the
fatigue strength at 107 cycles.

Surface Build
Study Heat treatment R σf estimate (MPa)
state direction

as-built Z 750 °C 1h Ar + HIP -1 223

[VAY 19]
Machined Z 750 °C 1h Ar + HIP -1 513

as-built Z HIP + 700 °C 1h -1 „ 200

[KAS 15]
Machined Z HIP + 700 °C 1h -1 „ 350

Machined Z HIP -1 620

[LEU 13]

as-built Z SR -1 155

as-built Z SR + HIP -1 195-220

[MAS 18]
Polished Z SR -1 370
Polished Z SR + HIP -1 610

Polished Z None 0.1 350

[GON 15]

as-built Z 710 °C 2h (vac.) 0.1 200

[GRE 17] as-built Z HIP 0.1 170
Machined Z 710 °C 2h (vac.) 0.1 470

as-built Z 540 °C 4h (vac.) + HIP 0.1 300

[BAG 17]
Machined Z 540 °C 4h (vac.) + HIP 0.1 775

as-built Z 650 °C 3h (Ar) 0.1 300

[KAH 17] as-built Z 650 °C 3h (Ar) + HIP 0.1 300
Polished Z 650 °C 3h (Ar) + HIP 0.1 500-600

as-built XZ 800 °C 2h 0.1 230

[REK 15]
[RA 13] Machined - None 0.1 550

as-built 45° 650 °C 3h (vac.) 0.1 210

[WYC 13]
Polished 45° 650 °C 3h (vac.) 0.1 510

Table 1.1 | Fatigue limit σf estimations for L-PBF Ti64 in several studies , for varying
stress ratios R, surfaces states, heat treatments (SR = Stress Relieving) and orientations (Z K
build plate). Most of the time, the fatigue limit is estimated by the fatigue strength at 107 cycles.

22 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.4 Modeling the impact of defects on fatigue properties
Many studies have focused on modeling the impact of defects on fatigue properties. Their
aim is often, for a given population of defects, to estimate the resulting fatigue life [ROM 18,
PEG 19, XU 21]. When trying to understand more deeply the mechanisms at the origin of the
influence of defects, it can also be a goal in itself to find a model that quantifies the relative
harmfulness of the different defects [TAM 17, PLE 20a]. Such a model can help to understand
which parameters are the most relevant to defects harmfulness − e.g. its size, proximity to the
surface, etc. It therefore provides information about which types of defects have to be most
avoided. In the best case scenario, one can even hope to predict before a fatigue test the killer
defect − i.e. the one which will cause failure. Such a prediction tends to be complex and
attempts made so far often shown limited success.
Several physical quantities are used in the literature when trying to quantify defects’ harm-
fulness. Common ones are the stress concentration factor Kt , the fatigue notch factor Kf
and the stress intensity factor K . These parameters describe different physical phenomena.
It is thus interesting to understand the physical meaning behind the choice of one or another to
quantify defects’ harmfulness.

1.4.1 A brief exploration of theory: Kt , Kf , K... which diﬀerence?

Fig.1.15 completes Fig.1.3 by adding the physical quantity which is theoretically relevant to
predict the material behavior at each step of the overall fatigue life according to [SCH 09]. One
can see that according to this author, the relevant parameter for crack initiation is the stress
concentration Kt .

Cyclic Crack Micro-crack Macro-crack Final

slip nucleation growth growth failure

Initiation Crack growth

period period
Fracture
Stress concentration Stress intensity toughness
factor Kt factor K KIC

Figure 1.15 | Diﬀerent stages in fatigue life and dedicated physical quantities.
Adapted from [SCH 09, p.15].

1.4.1.1 The stress concentration factor Kt

The stress concentration factor Kt is employed to quantify stress increases at specific points
within a component due to variations in geometry. Specifically, it represents the ratio of the
maximum stress σpeak experienced by the material to the nominal stress σnominal that the mate-
rial would encounter without stress concentration. It is worth noting that Kt does not encompass
stress increments resulting from a reduction in cross-sectional area within a component. The
definition of Kt according to [SCH 09] is presented in Fig.1.16 and Eq.1.1.

Modeling the impact of defects on fatigue properties | 23

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

σpeak

σnominal

σpeak
Kt “ (1.1)
σnominal

Figure 1.16 | Stress concentration induced

by a central hole in a plate , adapted from
[SCH 09, Fig.3.1].

Elevated stress concentration triggers unusually high localized stresses. This, in turn, can
activate cyclic slip and potentially lead to crack initiation. Hence, Kt can be regarded as a useful
parameter for assessing crack initiation. From this perspective, the inﬂuence of defects
can be quantiﬁed by considering their role as stress and strain enhancers.

For simple geometries, analytical formulas expressing Kt as a function of geometrical param-

eters can be found in the literature. For instance, [ING 13] showed that the stress concentration
generated by an elliptical notch in a semi-inﬁnite panel is given by Eq.1.2 − see Fig.1.17 for an
illustrative scheme.
σ

σpeak
d
ρ
d
d
σpeak Kt “ 1 ` 2
ρ
(1.2)

where d is the depth of an elliptical notch in a

semi-inﬁnite panel, ρ is the radius of curvature
σ and Kt is the stress concentration at its root.
Figure 1.17 | Stress concentration in-
duced by an elliptical notch in a semi-
inﬁnite panel submitted to a stress σ in the
direction d⃗σ perpendicular to the ellipse ma-
jor axis. d and ρ are respectively the depth
of the notch and the radius of curvature at its
root.

24 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.4.1.2 Kf and the Theory of Critical Distance (TCD)
However, it has long been known that the diﬀerence in fatigue strength between a
smooth (without any notch) and a notched specimen is generally not well predicted
by the simple use of Kt . This led to the deﬁnition of the fatigue notch factor Kf , which is
simply the ratio between the fatigue strengths of smooth and notched specimens [WEI 95] −
see Eq.1.3.

def Fatigue strength of smooth specimen

Kf “ (1.3)
Fatigue strength of notched specimen
Various expressions have then been proposed to describe the Kf ´ Kt relation. The most
common and easiest one is probably the one in Eq.1.4. It considers that the diﬀerence between
Kf and Kt is given by a material constant, the notch-sensitivity factor q [MAJ 10].

def Kf ´ 1
q “ (1.4)
Kt ´ 1
However, such an expression does not explain the physics behind Kf : it somehow hides the
question behind another parameter, q. Moreover, Kf actually depends not only on the material
but also on other parameters such as the stress gradient −i.e. how fast the tensile stress decreases
in the material near the defect −which, unlike Kt , depends on the notch size.
Again, many expressions have been proposed to reﬁne the Kf ´ Kt relation so that it can
account for these eﬀects. One can for example cite the work of [PET 59] which considers the
stress that has to be compared to the fatigue limit (of smooth specimens) is not the one at the
notch root, but the one at a certain distance beneath the surface. Using some approximations,
he arrives at Eq.1.5.

Kt ´ 1
Kf “ 1 ` (1.5)
1 ` α{ρ
where ρ is the radius of curvature at the notch root and α a material constant.
Works in this field gradually led to the definition of a broader theory, the Theory of Critical
Distances (TCD) [TAY 08]. The main idea is simple: to induce crack initiation, a stress
raiser has to increase stress not only at a surface hot spot but also over some critical
distance/volume beneath the surface. A possible justification for this is that for a crack
to initiate, it needs to overcome microstructural barriers such as grain boundaries or interfaces
with other phases during the early stages of crack propagation (i.e. small crack propagation).
If the stress fades away too fast from the stress raiser, the remaining stress may not be high
enough to overcome these obstacles [Val 00].
Several methods are proposed in the TCD to calculate this stress. The simpler is the point
method, which goes back to the idea of Peterson [PET 59]. It consists of taking the stress at
a given distance L{2 from the hot spot generated by the stress raiser. It is also possible to
calculate an average stress by integrating stress along a line, over an area or a volume. The
characteristic distance L is a material constant and can be calculated from the threshold stress
intensity factor Kth and the fatigue strength of a smooth specimen σ0 using Eq.1.6.
ˆ ˙2
1 Kth
L“ (1.6)
π σ0

Modeling the impact of defects on fatigue properties | 25

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
where Kth is the threshold stress intensity factor and σ0 is the fatigue strength of a smooth
specimen. Note that these two parameters should to be measured using conditions representative
of the loading experienced by the studied sample (stress ratio R for instance). This expression
of the L parameter is discussed in more detail in [TAY 08].
The TCD has proven to yield satisfying results in many applications, including on engineering
components and for various fatigue loading. Among others, [GIL 22] studied the ability of TCD
to predict the impact of surface notches of as-built AM surfaces on fatigue properties. However,
this method requires to know the precise stress distribution inside the material.
Although FEM is now widely used and can provide such data, its use can be prohibited by
high computation times in cases of large and complex geometries. This is typically the case for
as-built surfaces obtained with additively manufactured components. In such cases, calculations
are done mostly on 2D proﬁles or small surfaces to have reasonable calculation times [BAR 21,
VAY 19, HAM 20], which is a strong limitation.

1.4.1.3 The stress intensity factor K

The stress intensity factor (SIF, or K) is a parameter that describes the stress state at a
crack tip. It is therefore a parameter related to the propagation of an already existing crack:
high stress intensity factors will induce a fast crack propagation. The relation between SIF and
crack propagation rate can be modeled using laws such as the one proposed by [PAR 61].
Although the stress intensity factor SIF is a parameter related to propagation and not crack
initiation, it is very often used to model the harmfulness of defects. In other words,
the defects are assimilated to cracks of equivalent size. This can be justified by the fact
that for a sufficiently large defect, the crack initiation step can be reduced to a very short time,
the majority of fatigue life being spent on long crack propagation (which is governed by the SIF
at the crack tip). This effect was depicted in Fig.1.4.
Moreover, as soon as a crack initiates from a given defect, the resulting stress intensity factor
is close to the one of a crack whose length is the sum of the defect and the crack one. Even though
this is an approximation and the SIF will be over-estimated at the beginning of propagation,
when the size of the crack emanating from the defect gets larger, its SIF becomes equivalent to
the SIF of a crack of similar size but with no defect. This effect is shown in the case of a circular
hole in a finite plate in Fig.1.18.
In other words, if the defect is large enough, it may be suitable to consider the defect to be a
crack from the beginning, and therefore to define an associated SIF that will reflect the severity
of the defect. This approach has been widely used in many applications in the literature. The
equivalent SIF is commonly calculated using the Murakami parameter, which is the
square root of the projected area of the defect in the loading direction [MUR 83].
This parameter is then often used to estimate the fatigue limit σω of a specimen containing a
defect using Eq.1.7 [MAS 18, YIN 23].

C ¨ pHV ` 120q
σω “ ? (1.7)
area1{6
?
where σω is in MPa, HV is the Vickers hardness in kgf/mm2 , and area is the highest Murakami
parameter measured for the defects contained in the specimen, in µm. The constant C is equal
to 1.43 for a surface defect and 1.56 for an internal one.

26 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 1.18 | Comparison of the stress intensity factor of a central crack and of a
hole with two edge cracks. Taken from [SCH 09, Fig.5.8]. As the cracks propagate (i.e. as
2a{W increases), the stress intensity factor in the case of the two edge cracks initiated on the hole

rapidly reaches the same value as the one of the central crack.

1.4.2 Estimation of defects harmfulness in the case of L-PBF Ti64

?
1.4.2.1 The Murakami parameter area

The impact of defects on the fatigue properties is very often quantified using the Mu-
rakami parameter, which is the square root of the area of the defect projection along the
?
loading direction. It is usually noted area. [MUR 02] has shown that there is a good corre-
lation between this parameter and the SIF of a crack. In cases where it seems reasonable to
?
assimilate the defect to a crack, the area parameter seems therefore a good option to estimate
its harmfulness.
Thus, the Murakami parameter has been used in several materials to predict the impact of
various defects on the fatigue limit: inclusions in steels [MUR 02], but also artificial surface
defects in conventionally processed Ti64 [MAT 03]. Since then, many studies questioned its
validity on additively manufactured materials [PER 20, MAS 18, ROM 17, BAR 23, BER 17,
TAM 17].
Although the simple definition of the Murakami parameter as the square root of the projected
area is straightforward in the case of isolated core defects, some adaptations are needed for
particular cases. It is for instance the case to take into account the proximity of a defect to the
surface, the proximity of two neighboring defects or even the irregular morphology of defects.

Modeling the impact of defects on fatigue properties | 27

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
?
Fig.1.19a-e introduce the deﬁnition of this eﬀective parameter areaef f used by [MAS 18].

?
Figure 1.19 | Definition of the effective Murakami parameter area eff in different
configurations [MAS 18]. The figures correspond to the projection of the defects in the plane
perpendicular to the stress. (a) Internal non-spherical defect. (b) Surface defect. (c) Sub-surface
defect. (d) Interaction between two adjacent defects. (e) Inclined defect in contact with the
surface
.
It can be noticed that this parameter allows − at least partially − to take into account the
greater harmfulness of internal defects located near the surface. Indeed, it can be seen that
the whole ligament connecting such a defect to the surface is considered mechanically inefficient
and part of the defect. This effect is taken into account when the distance of the defect to the
surface is smaller than its size. This is consistent with Fig.1.12 which showed that the stress
concentration induced by a spherical pore is higher when its distance to the surface is smaller
than its radius.
The Murakami parameter is often measured post-mortem on a fracture surface.
The defect responsible for failure can then be identified and its projected area measured (by
?
SEM for example). Typical area values of (internal and surface) defects responsible for failure
in materials manufactured by L-PBF are given thereafter. They give an order of magnitude of
the size of the defects that can cause failure, even if the studies in question concern machined
or polished specimens:
• [HU 20b] found values between 30 µm and 72 µm on 21 machined Ti64 specimens, with an
average of 61 µm.
• [MAS 18] measured values between 25 µm and 138 µm on 12 polished Ti64 specimens, with
an average of 49 µm.
• [AND 19b] found values between 42 µm and 109 µm on 10 polished 316L specimens, with an
average value of 69 µm (only the specimens manufactured with the best process parameters
are considered).

28 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.1.20 shows the estimations of fatigue life for polished L-PBF and E-PBF Ti64 specimens
using the Murakami parameter [MAS 18]. We observe a rather good estimation of the fatigue
limit by the model in the case of L-PBF specimens. All the failed specimens were cycled at a
stress higher than the predicted fatigue limit σω (taking into account the presence of defects),
while several specimens cycled at the fatigue limit did not break after 107 cycles. On the other
hand, the model is not working well for the E-PBF specimens, since several specimens broke
under the predicted fatigue limit.

Figure 1.20 | Fatigue limit σw prediction for L-PBF and E-PBF polished specimens
using the Murakami parameter and Eq.1.7 [MAS 18]. The fatigue limit prediction from
the model corresponds to the black line, whereas experimental data is in red and blue. Note
that in this ﬁgure, the ordinate does not correspond directly to the stress amplitude during the
σa test but to the amplitude normalized by the factor (HV+120) found in the σω formula. The
acronym DMLS (Direct Metal Laser Sintering) refers here to L-PBF process and EBM (Electron
Beam Melting) refers to EBM.

In addition to estimating the impact of defects on the fatigue limit, it can be interesting to
evaluate their impact on fatigue life reduction in the finite life domain. Several studies have
for example evaluated the possibility of predicting the killer defect by characterizing the defect
?
population. [PER 20] proposed a prediction using the Murakami parameter area. For this
purpose, E-PBF Ti64 chemically polished specimens were scanned using XCT before testing and
?
the defects with the largest area were identified. Fatigue tests in uni-axial tension (R=0.1)
were then performed and the defects responsible for failure were identified.

Modeling the impact of defects on fatigue properties | 29

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
It can be found in Fig.1.21 that for the 13 tested specimens, the defect responsible for
the failure was systematically among the most critical defects identiﬁed using the
?
parameter area. However, Fig.1.21 also shows that the killer defect is rarely the one
predicted by the model, which shows its limits. [AND 19a, Fig.III-38 p.160] followed the
same methodology on L-PBF 316L and came to similar conclusions.

Figure 1.21 | Attempts to predict the killer defect in E-PBF chemically polished
?
Ti64 fatigue specimens (R “ 0.1) using the area parameter [PER 20]. For each spec-
?
imen, the main defects detected using XCT are ranked according to their size − i.e. the area
parameter. The killer defects identiﬁed after fatigue testing are highlighted in red. Ten speci-
mens have undergone only chemical etching (in blue) while three did also undergo HIP treatment
(in green).

Another important limit of this approach is that it is limited to specimens whose defects
can be clearly delineated as shown in Fig.1.19a-e, as is the case for internal or surface pores.
In the case of L-PBF Ti64, it means that it can only be used for machined or polished specimens.
Indeed, as discussed in Section 1.3.1, these are the only cases where most of the cracks will initiate
on pores. Conversely, for specimens with an as-built surface, the defects at the origin of failure
are most often notch-like defects. These have more complex shapes and are diﬃcult or even
impossible to delineate since they tend to be interconnected with each other (see Fig.1.22a-b for
an example of interconnected valleys). Other methods may therefore have to be used to
characterize the mechanical severity of valley-type defects.

30 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 1.22 | Highlight of complex surface valleys geometry on a vertical as-built
surface of a Ti64 sample manufactured by L-PBF, adapted from [BEN 17]. (a) SEM-
SE observation of the surface. (b) Delineation of valleys in white.

?
The study of [NAK 19] also suggests that the area parameter as deﬁned in Fig.1.19a-e
is not suited to characterize specimens with as-built surfaces. In this work, the fatigue limits
of L-PBF and E-PBF Ti64 specimens with as-built surfaces have been estimated. They were
found to be of 155 MPa and 140 MPa respectively (for non-HIPed specimens). According to the
observed fractured specimens in [NAK 19], failure occurs at the surface for as-built surfaces.
Furthermore, no obvious and clearly delineated defects could be found on the SEM fracture
?
surface. Eq.1.7 has then been used to estimate the area parameter of the hypothetical defect
leading to the experimentally measured fatigue limit. The obtained values are 16 505 µm and
9635 µm respectively, which was very far from reality.
To overcome this issue, the authors used semi-analytical formulas to estimate an equivalent
? ?
area (denoted areaR ) parameter based on roughness parameters measurement − see Eq.1.8
and Eq.1.9. These formulas were proposed by [MUR 02] for surfaces with periodical notches of
given depth and separated by a given distance. To use these formulas in the case of as-built sur-
faces, [NAK 19] proposed to replace the notch depth and the notch-to-notch distance by Ra and
?
RSm roughness parameters respectively. By doing so, they could ﬁnd more consistent areaR
values (264 µm and 203 µm for E-PBF and L-PBF specimens respectively). Such approaches,
based on roughness parameter measurements, are detailed in the next section.

? ˆ ˙ ˆ ˙2 ˆ ˙3
areaR Sz Sz Sz Sz
“ 2.97 ´ 3.51 ´ 9.51 , for ď 0.195 (1.8)
RSm RSm RSm RSm RSm
?
areaR Sz
« 0.38, for ě 0.195 (1.9)
RSm RSm
where Sz is the maximum surface height [Int 21] and RSm the mean width of proﬁle elements.
?
To conclude, the area parameter (with its initial deﬁnition) yields interesting results −
even if still perfectible in several cases − to account for the severity of internal or surface pores
which can be clearly delineated. However, samples with as-built surfaces require other solutions.

Modeling the impact of defects on fatigue properties | 31

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.4.2.2 Estimation of surface defects impact on fatigue by roughness measurements

Since surface defects are of primary importance in the case of as-built L-PBF Ti64 specimens,
many studies have been published on the relation between surface state and fatigue properties.
Most often, the surface state is characterized using standard roughness parameters, whose def-
inition can be found in standards such as [Int 21]. In many cases, the only parameter used is
the arithmetic average roughness (Ra in 2D and Sa for areal measurements).
For instance, [MIA 21] tried to correlate the fatigue limit of E-PBF Ti64 specimens from
several studies with Ra , using a simple curve fitting. Even though a reasonable fit is found,
such relations tend to be valid only in particular cases and do not provide tools that can be
used in a broader context (different materials, manufacturing processes, build directions, etc.).
Furthermore, only a few studies are compared in [MIA 21], which makes the fit questionable.
In general, the Ra parameter is therefore not considered to be a good candidate
to correlate with fatigue performance [ZAB 18, GOC 19], especially for as-built surfaces
of AM materials which often show much more complex geometries than more usual machined
surfaces. This is understandable since Ra only reflects the average surface state, while fatigue
properties are rather determined by the few most critical defects present in the material.
Several studies − concerning both AM and conventional machined materials − conclude
therefore that a parameter such as the maximum valley depth (Rv /Sv ) is better suited
[TAY 91, GOC 19]. However, even in the case of [GOC 19] who found a correlation between Sv
and fatigue life, it can be seen that a lot of scatter remains − see Fig.1.23a-d.

Figure 1.23 | Correlation between roughness parameters Sa /Sv and fatigue life of
L-PBF IN718 specimens, taken from [GOC 19]. Fatigue tests were carried out at R=0.1
and σmax “ 900 MPa with a frequency of 5 Hz. (a) Sa measured by Structured Light scanning
(SL) (b) Sa measured by XCT (c) Sv measured using by SL) (d) Sv measured by XCT.

32 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To go a step further, several models have been proposed to combine several rough-
ness parameters to get a better prediction of the fatigue properties [LE 21, ARO 99,
ARO 02, BHA 23, PEG 19]. As mentioned previously, [MUR 02] proposed for example a semi-
?
analytical formula to calculate an equivalent area parameter associated with surface roughness.
For its part, [ARO 99] proposed to compute an eﬀective stress concentration Kt using Eq.1.10.
Note that this equation is derived from Eq.1.2 shown previously. The fatigue notch factor Kf
can then be computed as discussed in Section 1.4.1. Using this methodology, [PEG 19] and
[LEE 20] obtained reasonable S-N curve predictions for L-PBF Ti64 and 304L stainless steel
specimens. However, more studies are needed to estimate the relevance of such an approach.

ˆ ˙ˆ ˙
Ra Ry
Kt “ 1 ` n (1.10)
ρ Rz
where Ra , Ry and Rz are the arithmetic average roughness, peak-to-valley height and 10-point
roughness respectively. ρ is the average radius determined from the dominant profile valleys.
n “ 2 for uniform tension and n “ 1 for shear loads.
[POM 20] used similarly an estimation of Kf using roughness and curvature measurements,
but in a more local approach. The objective was to try predicting, on top of the fatigue limit, the
area of crack initiation. To achieve this result, the surface of the specimen is first segmented into
1 mm squares. For each of them, Kf is calculated from the maximum depth of the valley in this
zone and the radius of curvature at the bottom of the notch, determined using Focus Variation
microscopy (FV). Fig.1.24 shows that with this methodology, the area of crack initiation −
marked by a red cross) − could be effectively identified.

Figure 1.24 | Kf ,s map computed using focus variation microscopy on a cast alu-
minum bending fatigue specimen [POM 20]. The crack initiation site is highlighted by
the red cross and is found to occur in the area of maximum Kf,s .

Such an approach seems promising in the particular case of as-built AM surfaces

but has been deployed in a limited number of studies. To go further, it would also be
interesting to make an even more local characterization. Indeed, the segmentation in Fig.1.24
of the surface using 1 mm squares is somehow arbitrary and one can suppose that a different
segmentation could have led to different results. Furthermore, the defect causing failure is not
precisely identified in this case.

Modeling the impact of defects on fatigue properties | 33

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1.5 Towards a 3D local characterization of AM surfaces
The previous sections discussed how to use roughness measurements to quantify surface
defects’ severity. However, it is also worth noting that the way the roughness is measured can
be of primary importance. This is especially true for as-built surfaces inherited from L-PBF, and
more generally from AM processes. Indeed, since AM allows greater design freedom, additively
manufactured parts often show more complex shapes than those obtained with conventional
processes. Architected structures are a typical example [FOR 16, SOR 21]. In such cases,
standard contact-based roughness instruments or optical systems provide incomplete information
due to limited access to the surface. Additionally, at the micro-scale, surfaces inherited from
AM exhibit a large variety of surface defects such as notches, protrusions or unmelted powders
[SAN 21, PER 19, CHA 22]. The interaction between all those defects can make the analysis
even more challenging. For instance, overhanging surface features or unmelted powder
particles can hide underlying notches, as depicted in Fig.1.25. This complexity highlights
the need for both instruments and analysis tools that can accurately characterize
complex 3D structures and surface features1 .

Notches hidden
Perpendicular by a spatter or
notch Oblique notch powder particles

2.5D
3D
by XCT
Figure 1.25 | Schematic showing the interest of characterizing AM as-built surfaces
in 3D using XCT. Several examples are shown, corresponding to different surface defects. The
orange lines show surfaces as seen by a 2.5D characterization tool which probes the surface from
a single point of view and without the ability to look through matter. The blue lines correspond
to the true surfaces, i.e. the one which would ideally be obtained through XCT using a high
resolution, with no noise nor artifacts. In two of the three examples shown, the 2.5D surface
characterization fails to account for notches because they are hidden by other surface features.
1
Note that the term surface feature is used here because it is very common in the field of surface characterization
and has more general meaning than surface defects. Surface features can thus refer to specific elements of the
surface that are to be characterized because they may be beneficial for a given application, whereas surface defects
refer only to features considered as undesirable.

34 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
While many studies report 3D roughness analyses, stricto sensu most of them are not fully
3D measurements. These approaches have limitations in terms of sample geometry and surface
features they can characterize. These limitations can originate from the instruments used for raw
surface acquisition, such as interferometers or confocal microscopes. The data representation
used for subsequent analysis can also lead to information loss. To clarify this, key concepts will
be introduced hereafter to differentiate existing approaches for 3D surface characterization and
their actual capability to characterize complex 3D structures and surface features.
Surface characterization often involves optical instruments like white light interferometry
or confocal microscopy [MIN 14, HAM 20, FOX 18, TOW 16, ALI 20]. These instruments use
unidirectional light sources, characterizing the surface from a single perspective. Consequently,
surface topography is projected onto a single plane − the one perpendicular to the
light source. This yields height maps with a single z-coordinate value for each (x,y) point on
a regular grid. As the third (z-axis) dimension is only partially used, this data representation
can be referred to as 2.5D [HAI 13, THO 18]. Accordingly, instruments like interferometers or
confocal microscopes will be termed 2.5D instruments hereafter.
2.5D characterization is feasible for surfaces with simple geometries like planes or cylinders.
For complex surfaces, irregularities often prevent the measurement tool from accessing certain
areas. This is obvious in AM parts with sophisticated geometries. Even when aiming at charac-
terizing flat and accessible regions, limitations persist if the surface topography includes hidden
features. This effect is depicted in Fig.1.25. Orange lines in Fig.1.25 show that a 2.5D surface
characterization manages to describe a notch only if the latter is perpendicular to the surface. It
fails if the notch is oblique or hidden by partially melted powders and/or spatters.
In order to overcome such limitations, developments have recently been made in the field of
free-form metrology [JIA 20a, ABD 13], which aims at characterizing surfaces of complex and
arbitrary shapes. Raw surface acquisition can now be achieved through instruments like struc-
tured light scanners [WAN 21, SHA 21] or commercial X-ray tomographs [LOU 19b, ZAN 19,
FOX 18, THO 18]. These tools offer omnidirectional characterization, thereby facilitating the
characterization of complex surfaces. To fully harness this omnidirectional capability, alternative
data representations are required. The 2.5D representation limits each (x,y) point to only one
z-coordinate, imposing a substantial constraint. To mitigate this, free-form metrology adopts
surfaces represented as meshes or point clouds composed of three-dimensional points (x,y,z).
Multiple points with identical x and y coordinates can thus exist.
While the free-form surface representation is versatile, it requires novel analysis techniques
different from 2.5D methodologies, e.g. for surface filtering. Although some studies have
employed free-form representations [LOU 19b, GRA 23, ABD 13, MCB 20], this approach is
rarely used. Even when surfaces are captured using methods like XCT, they are often reduced
to a 2.5D height map for subsequent analysis [FOX 18, THO 18, PAG 18, KER 12, LIF 21].
This trend stems from the fact that surface analysis operations, such as filtering of roughness pa-
rameters measurement, can be much more complicated when using a complete 3D representation
instead of a simpler 2.5D one [JIA 20a].
Thus, some technical issues still need to be overcome to obtain more robust results. Even
though the existing solutions may be sufficient, they are often sophisticated and therefore not
necessarily user-friendly. They require advanced computer programming skills to be imple-
mented and important computational resources. Further work is thus needed to address these
challenges and promote the broader adoption of free-form metrology.

Towards a 3D local characterization of AM surfaces | 35

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Another distinction can be made between instruments that are able to look through matter
(e.g. XCT) and those that are not (e.g. structured light scanning). Although 3D light scanning
enables the characterization of complex geometries, it will be limited for parts with inaccessible
areas (e.g. lattice structures or internal channels) or too complex/small surface features. On the
contrary, the ability of XCT to look through matter and have access to internal features enables
− at least in theory − the characterization of any shape and account for any hidden surface
features. Hence, as depicted by blue lines in Fig.1.25, XCT succeeds in properly accounting
for complex 3D surface features such as notches hidden by unmelted powder, provided that
the spatial resolution is suﬃciently high. This makes XCT an interesting option for the
characterization of surface defects inherited from AM processes such as L-PBF.

Tab.1.2 summarizes the diﬀerent factors that can inﬂuence the capability of the various in-
struments to properly characterize surfaces with intricate 3D macroscopic shapes and/or hidden
microscopic surface features.

2.5D instruments Structured light

XCT
(e.g. interferometers) scanning
Source directionality Unidirectional Omnidirectional Omnidirectional
Data representation 2.5D 3D possible 3D possible
Ability to look
No No Yes
through matter
Possible to characterize
No Yes Yes
complex 3D components
Possible to account for
No No Yes
hidden micro-features

Table 1.2 Factors inﬂuencing the ability of diﬀerent instruments to properly characterize surface
with complex macroscopic shapes and hidden surface micro-features (e.g. notches hidden by
unmelted powder particles).

36 | Chapter 1 – Brief overview on defects harmfulness on L-PBF Ti64 fatigue properties

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary
Defects act as stress concentration sites, facilitating crack initiation and consequently
inﬂuencing fatigue resistance. In Ti64 manufactured by Laser Powder Bed Fusion
(L-PBF), defects can manifest as internal anomalies (e.g., pores) or surface irregularities
(e.g., valleys). Recent research has revealed that surface defects exert a more pronounced
impact on fatigue resistance than internal ones, particularly for specimens with as-built
surfaces. Notably, the presence of surface defects can lead to a substantial reduction in
fatigue resistance. For example, the fatigue limit σf is typically 2 to 2.5 times higher for
machined or polished specimens (where surface defects have been eliminated) compared to
as-built specimens with surface defects − see Tab.1.1. This underscores the signiﬁcance of
developing surface treatments suitable for Additive Layer Manufacturing (AM) that can
eliminate these defects.

Traditional methods employed to gauge the impact of defects on fatigue properties

?
(e.g., the area parameter) may not be well-suited for accounting for surface defects in
as-built samples. This is because surface defects in these cases often exhibit intricate and
challenging-to-deﬁne morphologies such as interconnected valleys. Alternative strategies
have been suggested, involving the coupling of surface roughness and curvature measure-
ments to account for the mechanical severity of as-built surfaces. While these approaches
show promise, further investigation is required to validate their eﬃcacy.

In the context of surface characterizations, it seems relevant to question whether con-

ventional tools and ensuing analysis workﬂows are equipped to address defects that have
the most critical implications for fatigue properties. Traditional surface characterization
primarily takes place in 2D or 2.5D, which may not adequately capture hidden features
commonly found in as-built surface features (e.g. notches concealed by unmelted powder
particles). Recent advancements in X-ray Computed Tomography (XCT) and free-form
metrology (which pertains to 3D surface characterization) oﬀer promising perspectives in
this regard. However, such approaches are still relatively uncommon, and solutions for an-
alyzing free-form surfaces are often intricate and relatively unknown. Further investigation
is thus necessary to assess the applicability of such tools and enhance their accessibility.

Towards a 3D local characterization of AM surfaces | 37

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

A methodology for the 3D

characterization of surfaces using
X-ray computed tomography:
application to additively
manufactured parts

Contents
2.1 Materials and XCT data acquisition . . . . . . . . . . . . . . . . . . . . . 40
2.2 3D surface characterization methodology . . . . . . . . . . . . . . . . . . 41
2.2.1 Surface segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 3D roughness calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.3 3D curvature calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2.4 Quantiﬁcation of the harmfulness of surface notches . . . . . . . . . . . . . 54
2.3 Application to parts with complex geometries . . . . . . . . . . . . . . . 56
2.3.1 Gyroid structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.3.2 Octet-truss lattice structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.3 Impeller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4 Application to the prediction of crack initiation sites in fatigue . . . . 62
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

This chapter introduces a workﬂow for 3D surface characterization using XCT, mea-
suring both roughness and curvature. By combining these two measurements, we propose a
parameter representative of the severity of surface features with respect to mechani-
cal properties. This parameter is subsequently applied to the prediction of crack initiation sites
on chemically polished Ti64 samples produced by Electron beam Powder Bed Fusion (E-PBF)
and subsequently polished. Note that such a methodology could be applied to a large variety
of other applications not related to mechanics, e.g. to study the impact of surface roughness on
the osseointegration of biomedical implants [DIA 22, ALL 11].

| 39

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The aim of Chapter 2 is also to demonstrate the beneﬁt of characterizing surfaces inher-
ited from AM as free-form surfaces, taking into account hidden micro-features and opening
new avenues to characterize components showing intricate macroscopic shapes. Hence, several
application cases are shown with samples of complex shapes, such as architected structures.

2.1 Materials and XCT data acquisition

Different samples were used to demonstrate the application of the developed workflow. Most
of them were manufactured by Electron Powder Bed Fusion (E-PBF) using an ARCAM
A1 machine and Ti64 powders. To characterize their as-built surface, no surface treatment was
applied. Powder grain size distribution ranged from 60 µm to 100 µm and the layer thickness
was set to 50 µm. More details about the processing conditions can be found in [PER 19]. To
characterize their surface, samples were scanned using laboratory XCT.
A 2 mm vertically built cylinder was first used as a simple example. It was scanned
using a cone beam phoenix | x-ray V | tome | x laboratory tomograph with a voltage of 90 kV,
a current of 240 µA, an exposure time of 333 ms and 720 projections. No physical filter was
used during scan acquisition. Reconstruction was performed using a standard filtered back
projection algorithm (phoenix datos x software). The voxel size used was 2.5 µm, but volumes
were downscaled by a factor 2 before further analysis1 . The resulting voxel size is 5 µm.

Figure 2.1 | Ti64 architected structures fabricated by E-PBF. (a) Gyroid structure.
(b) Octet-truss lattice structure.

Two Ti64 architected structures were also characterized: a gyroid structure [SCH 70,
YAN 19, SOR 21] and an octet-truss lattice structure, see Fig.2.1a-b. Surfaces were kept
as-built in both cases.
1
The data were retrieved from the work of [PER 18], who performed this downscaling to reduce data size.

40 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Sample Gyroid Gyroid Octet truss

Voxel size
5 10 10
(µm)
V (kV) 230 230 230
I (µA) 35 70 58
Cu Filter (mm) 1.4 1 1
Number of
3616 3616 2240
projections
Exposure
2000 333 500
time (ms)

Table 2.1 | Acquisition parameters for XCT scans performed using the RX Solutions laboratory
tomograph.

Finally, a Ti64 impeller manufactured by L-PBF was scanned at the BM18 beamline at ESRF
synchrotron2 , using a voxel size of 18.5 µm.
All XCT scans were converted to 8-bit after reconstruction to reduce data size. After con-
version, volumes size were 158 Mo, 4.6 Go, 5.1 Go and 8.9 Go for the cylinder, the gyroid (10 µm
scan), the octet-truss lattice and the impeller respectively.

2.2 3D surface characterization methodology

2.2.1 Surface segmentation
The first step of surface characterization is the segmentation, i.e. the extraction of the surface
from XCT scans. Since it can be affected by noise, a noise-reducing filter is applied to the
reconstructed volume before further calculations. An edge-preserving filter is used to preserve a
detailed description of the surface [JAI 16]. A median filter is used because it provides satisfying
results while keeping computing time reasonable for large volumes.
After filtering, the sample is segmented from the volume through thresholding. Here,
thresholding aims at separating the two main peaks of the grayscale histogram, which will be
referred to as dark (voids) and bright (sample) peaks − see Fig.2.2. Several methods can be used
to automatically determine an optimal threshold. One of the most popular is Otsu’s method
[OTS 79, VAN 22]. The latter assumes that all voxels are separated into two classes based on
their gray level while minimizing the inter-class variance. It yields in most cases consistent and
robust results.
2
The data could be obtained thanks to the help of Dr. Wolfgang Kitsche (design and sizing; testing; Institute
of Space propulsion), Katia Artzt, Tarik Merzouk, Ahmet Turak, Jan Haubrich, Guillermo Requena (LPBF
building, materials characterization, machine operation etc., Institute of Materials Research).

3D surface characterization methodology | 41

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Dark Bright
peak peak

Maximum Maximum
abscissa abscissa

Black White
1
Figure 2.2 | Deﬁnition of the TTBH, illustrated based on the E-PBF cylinder normalized
grayscale histogram.

However, it may not always be the best choice if one is looking for some specific surface
features such as surface defects that will be prone to initiate failure during mechanical loading.
In this specific case, partially melted particles are unlikely to be of much interest. They can even
hide more severe surface defects such as notches. Contrariwise, deep and sharp notches are very
often the most critical defects and are therefore of great interest. However, the sharpest ones are
often difficult to segment from XCT data, because the corresponding voxels show intermediate
grayscale values. This leads Otsu’s method to consider many of them as foreground voxels,
erasing notches from the surface although they are the most interesting defects when questions
related to crack initiation and failure must be tackled.
This effect can be seen in Fig.2.3a and Fig.2.3b. Fig.2.3a shows an XCT radial slice of a
2 mm as-built E-PBF cylinder, where two deep notches can be seen. Fig.2.3b shows in orange
the pixels considered as foreground using Otsu’s threshold for segmentation. It can be seen that
both notches are not properly accounted for when using Otsu’s threshold, although
they are clearly visible on the grayscale image in Fig.2.3a.
A threshold corresponding to a higher grayscale value than Otsu’s one will be certainly more
adapted to capture such severe surface defects. Indeed, both partially melted particles and sharp
notches typically show intermediate grayscale values because they are at the interface between
matter and air. Thus, selecting a threshold that stands at the end of the plateau − just at
the left border of the histogram’s bright peak − will enable to discard some parts of unmelted
powder particles while better capturing sharp notches.

42 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Notches

200 µm 200 µm 200 µm

Figure 2.3 | (a) XCT 8-bit radial slice showing two sharp notches and a powder particle (voxel
= 5 µm). (b) Otsu’s threshold application (orange area) (c) TTBH application (purple area).

The thresholding method proposed is inspired by the triangle threshold introduced by [ZAC 77].
It is thereafter referred to as the Triangle Threshold for Bimodal Histograms (TTBH). Its princi-
ple is described schematically in Fig.2.2, and ImageJ/Python implementations have been made
available on an online repository [STE 23].
First, if the histogram is too noisy, it may be useful to smooth it. In the present work, a
moving average of size 10 is applied. Second, the histogram is normalized so that both the bright
peak maximum and the distance between the two histogram peaks are equal to 1. Finally, the
desired threshold is simply the gray value which maximizes the distance d as defined in Fig.2.2.
As required, the obtained threshold is located just at the left edge of the bright peak. Fig.2.3b-c
show that the two sharp notches are better captured using the TTBH than by Otsu’s
method. The powder grain is cropped, which can be both an advantage and a drawback,
depending on which surface features one aims to characterize.
It may be worth mentioning that this thresholding method is particularly sensitive to noise
and artifacts (e.g. beam hardening). This can be at least partially compensated by the use of
the proper noise-reducing filter beforehand. It may also be relevant in some cases to make a
compromise between the TTBH and Otsu’s threshold. A simple and convenient way to do that
can be to calculate both and take an intermediate value.
Following segmentation, the volume undergoes a cleaning process to remove all internal pores
or tiny objects caused by measurement noise. This cleaned binary volume is subsequently
referred to as the sample mask, illustrated in Fig.2.4b. In our case, the sample surface is defined
as the surface mask depicted in Fig.2.4c. It is the binary mask whose foreground is composed
of all surface voxels. These surface voxels are the foreground voxels of the sample mask which
have at least one background voxel as a first neighbor. A connectivity of 1 is used to determine
neighbors.

3D surface characterization methodology | 43

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Intensity Foreground Back- Foreground Back-

(Sample voxel) ground (Surface voxel) ground

Figure 2.4 | (a) Schematic example of a grayscale XCT slice. (b) Corresponding sample mask.
(c) Corresponding surface mask.

2.2.2 3D roughness calculation

The topography of a surface is often divided into 4 components, each one corresponding as a
first approximation to a range of spatial frequencies: the object’s form, the waviness, the rough-
ness, and the micro-roughness [POM 19, Int 12]. Most commonly, only the roughness component
is considered significant for characterizing surface micro-features such as notches or unmelted
powder particles. The form and waviness typically represent geometric deviations, while micro-
roughness is often viewed as measurement noise. The objective is therefore to discriminate
roughness from these other components to achieve a complete surface characterization.
In the case of a conventional 2.5D surface characterization, the shape of the characterized
surface is necessarily simple (plane, cylinder...). Form removal, which consists of subtracting the
component geometry, can thus be done rather easily using least-squares optimization. Filters
are then used to discriminate roughness from waviness (L-filter) and micro-roughness (S-filter).
In both cases, the Gaussian filter is the default option [Int 12].
However, 2.5D characterization is only possible for parts with simple geometry, as discussed
in Section . To take full advantage of XCT, a 3D surface representation must be used − e.g.
a mesh or a point cloud made of points with arbitrary (x,y,z) coordinates. In this case, the
roughness characterization workflow can be more challenging. In particular, the form removal
step is complex when no analytical expression is adapted. For 3D printed parts, a first approach
consists of using as a reference the CAD file or a nominal geometry derived from analytical
formulas [?, PAG 18]. This is sometimes done to measure geometrical deviations between the
manufactured part and its CAD model [SAV 07, JIA 10, VIL 18b]. However, this requires the
CAD file or an analytical model to be available, and the geometrical deviations to be small
enough to get an accurate measurement.
If the CAD file cannot be used, it is possible to smooth the raw surface and use the result as
a reference. In this case, the roughness is defined as the distance between the surface
and a smoothed version of it. Many studies already addressed this topic, see the work
of [JIA 20a] for an extensive review. In such studies, surfaces are generally meshed and may
be smoothed using morphological filters [LOU 13b, LOU 19b], diffusion-based filters [JIA 11],
anisotropic diffusion-based filters [TAS 02], or wavelet decomposition [JIA 20b].

44 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Sample
voxels

Sample
Background mask
Erosion
1 + Logical XOR
Convert to ﬂoat/int
2 + 3D gaussian ﬁlter
Surface
Voxel voxels
intensity
Max

Smooth volume Background

Min Surface
mask
3 Thresholding
Smoothed
sample
voxels

Smooth
Background mask

4 SEDT Binary
Signed volumes
distance
Max
Smooth
mask SEDT
Min

Float/int volumes
5 Apply surface mask to SEDT
Height
Max

Min
Roughness
+
Background
Micro-roughness
Height
Max
6 S-ﬁlter
7 Remove borders if presence of edge eﬀects Min

8 Height distribution zero-centering Background

Roughness

Figure 2.5 | Workflow proposed for the 3D roughness computation. The sample mask
and surface mask are defined in Fig.2.4b-c. The images shown correspond to transverse cross-
sections of a 2 mm E-PBF Ti64 cylinder at different steps of the calculation.

3D surface characterization methodology | 45

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The proposed workflow only uses as input data the sample mask, whose computation is
detailed in Section 2.2.1.
The first step is the computation of the surface mask − also defined in Section 2.2.1.
The latter will be used at the end of the workflow to extract data from volumes only at the
voxels of interest − i.e. the surface voxels. The surface mask can be obtained by applying an
erosion to the sample mask and then computing the logical XOR operation between the sample
mask and the result of the erosion.
The second step of the workflow is the smoothing step. It is done here by converting the
sample mask to float or integer values and applying a 3D Gaussian filter3 . The chosen degree of
smoothing will determine the limit between roughness on the one hand and waviness/form on
the other hand. It is quantified by the cut-off wavelength λc [Int 15], which is the wavelength
considered to discriminate form and waviness from roughness. From an image processing point
of view, a Gaussian filter is more often adjusted using the standard deviation σ, which can be
easily derived from the cut-off wavelength following Eq.2.1.
c
ln 2 λc
σ“¨ « 0.187 ¨ λc (2.1)
2 π
where σ and λc are the standard deviation and the cut-off wavelength of the Gaussian filter
respectively.
ISO standards provide rules to appropriately set the value of the parameter λc
[Int 12]. Although they are meant for 2D filters applied to height maps, they are used by
extension in the developed workflow to adjust the 3D Gaussian filter. Ideally, the choice should
be made by observing the surface and identifying the features that need to be characterized as
roughness. It is then advised to set λc to be five times the size of the largest feature of interest,
choosing among a list of predetermined values (0.25 mm, 0.8 mm, 2.5 mm, etc.) [Int 12]. This
choice is somehow arbitrary, especially for surfaces obtained by AM with complex topologies.
This can explain the large variations of values used in the literature on AM materials [LOU 19a].
[GRA 23] used for example a cut-off wavelength of 0.25 mm while [VET 14] used several values
up to 2 mm.
In the third step of the workflow, the smoothed volume obtained after filtering
is segmented by thresholding. This results in a smoothed version of the sample mask
denoted hereafter as the smooth mask. The latter is then used as a reference for roughness
calculation. The most straightforward choice for the threshold value is the mean value between
the values corresponding to background and foreground voxels (e.g. 127.5, if the background
is 0 and foreground voxels are 255). However, this leads in general to some volume shrinkage,
especially for large λc values. To avoid this problem, the threshold value is instead determined
automatically so that the smooth mask volume equals the volume of the original mask.
The fourth step aims at computing the distance between the sample surface and
its smoothed version. For this purpose, the Signed Euclidean Distance Transform (SEDT)
is computed from the smooth mask4 . This results in a float volume where the value at each
3
An alternative solution, not developed here, would be to perform smoothing by applying a 3D morphological
filter [LOU 13a, JIA 20a] on the sample mask.
4
Several methods can be used for the distance computation. The SEDT is given as an example here because
it is a very common operation that can be easily found on software such as ImageJ. However, it may lead to some
bias in the measured roughness for a smooth surface scanned with a large voxel size. Other methods may in such
cases lead to more accurate results. This topic is discussed in Appendix A.

46 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

tomography: application to additively manufactured parts
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
foreground voxel is the distance to the nearest background voxel, and contrariwise for the values
at the background voxels. The distances are signed, e.g. positive for background voxels and
negative for foreground voxels.
At the fifth step, the surface mask is applied to the smooth mask SEDT. This results in
a sparse volume where most of the voxels have the same background value (e.g. 0), except the
surface voxels which hold the distances to the smooth mask, i.e. the sought roughness values.
The sixth step consists in applying an S-filter. This aims at removing the highest frequen-
cies, in other words, the micro-roughness. The standard filter used for this purpose is again a
Gaussian filter [Int 12]. The cut-off wavelength used for this filter is commonly denoted λS . A
procedure to calculate the S-filter using 3D Gaussian filters only based on normalized convolu-
tion [KNU 93] is described in B. This filtering step becomes particularly relevant if the voxel size
set for XCT scans is not small enough to properly describe the surface topography. When the
voxel size approaches the dimensions of surface features, the surface tends to be oversimplified
and discretized, yielding a staircase appearance. Such oversimplification can result in sharp
fluctuations in the roughness measurement. Using the S-filter helps to moderate such abrupt
changes.
The seventh step addresses edge effects that arise due to Gaussian filtering [Int 16]. When
computing the filtered value for a particular voxel using the Gaussian kernel, surrounding voxels
are taken into account. However, for those voxels near the XCT scan boundary, the convolution
radius might extend past this boundary. For these voxels, the XCT scan has to be extrapolated,
for instance using an arbitrary constant value. The most common choice is zero. This value
aligns well with our needs since 0 represents the background value that is assumed to envelop
the sample. However, this method can result in biased values for surface voxels too close to the
scan boundary. For example, in the context of a cylinder scan, the topmost and bottommost
surface voxels may contain such biased values.
Various solutions have been suggested to address this challenge for 2D or 2.5D roughness
measurements [JAN 12]. The most straightforward solution is to discard roughness
values for points located too close to the edges, typically within a distance of the cut-
off value or half of it. In the described workflow, the approach adopted involves cropping the
upper and lower borders by a width of λ2c wherever necessary. However, if this method results
in discarding too much data, other strategies, based on extrapolation or the use of normalized
convolution [KNU 93], can be employed in the Gaussian filtering step. This would allow the
user to address edge effects without the need to truncate volumes later on.
Lastly, the eighth step ensures that the height distribution is zero-centered, meaning
the average height is zero. This condition is intuitive and is assumed when computing certain
roughness parameters like Ssk and Sku . To achieve this, the average height is subtracted from
the height of each surface voxel.
An example of roughness calculated on a 2 mm diameter cylinder fabricated by E-PBF is
shown in Fig.2.6. The initial sample’s surface and the smooth reference surface are displayed, as
well as the derived roughness. Note that points closer than λ2c to the upper and lower boundaries
were discarded at the end of the computation. The triangle threshold for bimodal histograms
presented in Section 2.2.1 was used for surface segmentation and a cut-off wavelength λc of
0.8 mm has been chosen for roughness calculation. Regarding the S-filter cut-off λs , a value of
three times the voxel size (0.015 mm) was used.

3D surface characterization methodology | 47

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 2.6 | Surface roughness measurements, as described in Fig.2.5, applied to a
2 mm diameter as-built E-PBF cylinder. Computations were made from an XCT scan with
a voxel size of 5 µm. Cut-oﬀ values were set to λc “ 0.8 mm and λS “ 0.015 mm “ 3 voxel size.
To avoid edge eﬀects, points closer to the boundaries than λ2c were discarded.

The most conventional roughness parameters can be computed easily from the obtained 3D
roughness measurement. Formulas that have been used to compute roughness parameters in
this chapter and subsequent ones are given in Eq.2.2.

1 ÿ
Sa “ |h| (2.2a)
n surface voxels
d
1 ÿ
Sq “ h2 (2.2b)
n surface voxels
´ ¯ ´ ¯
Sz “ max h ´ min h (2.2c)
surface voxels surface voxels
Sv “ ´ min h (2.2d)
surface voxels
m3
Ssk “ 3{2
(2.2e)
m2
m4
Sku “ ´3 (2.2f)
m22

where
1 ÿ
mi “ ph ´ h̄qi (2.2g)
n surface voxels

is the biased ith central moment, n is the number of surface voxels, h is the height of a given
surface voxel and h̄ is the mean height over all surface voxels.

48 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

tomography: application to additively manufactured parts
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Sa , Sq , Sz , and Sv parameters have intuitive definitions. However, additional explanation
may be required to understand the concepts behind Ssk , Sku .
The skewness parameter (Ssk ) reflects the asymmetry of the height distribution. When Ssk
is positive, the distribution leans towards positive height values, indicating that the surface is
primarily composed of peaks. Conversely, when Ssk is negative, the distribution leans towards
negative height values, suggesting that the surface is dominated by valleys. The kurtosis pa-
rameter (Sku ) signifies the sharpness of the height distribution. A high Sku value characterizes
a distribution with elongated tails and which is sharp near its mean value. This can imply the
presence of sharp peaks or valleys on the surface.
All these definitions are mostly identical to the ones provided in ISO standards for 2.5D
roughness measurements [Int 21]. However, it is worth noting that using the same formula with
2.5D or 3D will not necessarily lead to the same value for all parameters (even when ignoring the
ability of 3D measurements to capture features invisible to 2.5 measurements). Subtle differences
may arise, as discussed in [JIA 20a].
As far as computational performance is concerned, the use of 3D operations makes the pre-
sented roughness calculation methodology demanding in terms of Random Access
Memory (RAM). This can be a limitation for large volumes, which can be overcome to a cer-
tain extent by dividing the volume and performing computations in several steps. Processing
time is nonetheless reasonably low since efficient implementations exist for the operations
used. As an example, the computation for the 158 Mo XCT scan of the cylinder in Fig.2.6 took
around 1 min on a conventional laptop with 8 CPU cores and using the ImageJ macro provided
in [STE 23].
A Python implementation is also available in [STE 23], with some improvements to limit
RAM usage and significantly increase computation speed. It also proposes an alternative algo-
rithm for the distance calculation step (4th step), based on a so-called Nearest Neighbor Search
(NNS). The NNS-based method is not readily available in ImageJ but has the advantage of
requiring much less RAM than the SEDT-based method. Furthermore, it provides more accu-
rate results when the surface to characterize is very smooth, typically when the Sa roughness
parameter yields a value close to the voxel size. For this reason, roughness computations are
carried out using the NNS method from Chapter 3 onwards, although they are carried out us-
ing the SEDT method for the whole of Chapter 2. For the interested reader, a more detailed
presentation of the NNS-based method and a comparison of the measurement errors obtained
by SEDT and NNS methods for different surface states are provided in Appendix A.

2.2.3 3D curvature calculation

The curvature κ is a measure of how a curve (in 2D) or a surface (in 3D) bends at a particular
point. For a curve, it corresponds to the inverse of the radius ρ of the osculating circle at this
given point, as schematically shown in Fig.2.7a.
Curvature is a more complex concept for surfaces and several definitions exist. The closest
3D equivalent of the curvature in 2D would be the directional curvature, which measures how
much the surface bends along a particular direction. When the surface bends outward, the
directional curvature is positive; conversely, it is negative. This requires a prior definition of
which side of the surface is considered to be the outside. Although directional curvature can
be calculated in an infinite number of directions, two are of particular interest. These are the
direction of minimum curvature d⃗min and the one of maximum curvature d⃗max (the sign being

3D surface characterization methodology | 49

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
considered), also known as principal directions. The corresponding curvatures are called the
principal curvatures κmin and κmax .

b 3D
a 2D
Osculating
circle

Figure 2.7 | (a) Deﬁnition of the curvature κ and the radius of curvature ρ for a curve in 2D.
(b) Schematic representation of a saddle-like surface. The normal ⃗n and the principal directions
are indicated at the saddle point. In this particular case, κmin ă 0 and κmax ą 0.

Fig.2.7b shows an example of a saddle-type surface, where d⃗min and d⃗max are respectively
displayed in purple and orange.
Knowing principal directions and curvatures, the directional curvature in any direction can
be derived using Eq.2.3 [TAU 95].

¨ ˛J ¨ ˛ ¨ ˛
vn 0 0 0 vn
κ p⃗v q “ ˝ vmin ‚ ¨ ˝0 κmin 0 ‚¨ ˝ vmin ‚
(2.3)
vmax 0 0 κmax vmax
2 2
“ vmin ¨ κmin ` vmax ¨ κmax

where κ p⃗v q is the directional curvature in the direction ⃗v , ⃗v “ vn ¨ ⃗n ` vmin ¨ d⃗min ` vmax ¨ d⃗max
is an arbitrary vector and ⃗n “ d⃗min ˆ d⃗max is the surface normal.
Finally, two other common curvatures are often used: the mean curvature κmean and the
Gaussian curvature κgauss (see deﬁnitions in Eq.2.4 and Eq.2.5):

κmin ` κmax
κmean “ (2.4)
2

κgauss “ κmin ¨ κmax (2.5)

50 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Point where
Convolution curvature
kernel (rcurv=2) is measured

Foreground
pixel

Background
pixel

Figure 2.8 | Principle of integral invariant based curvature measurement from a

digital shape. The example is given in 2D for clarity, but the principle remains the same for
volumes. Here, the intersection area between the convolution kernel and the object is measured
by counting the number of pixels whose center falls into the kernel of radius rcurv .

For example, κmean is directly related to the volume of the intersection between the convo-
lution kernel and the object. Hence, an estimation of κmean at a given boundary point can be
computed using Eq.2.6 [COE 14].
˜ ¸
16 1 Vinter prcurv q
κmean prcurv q “ ¨ ´ 4 3
(2.6)
3 ¨ rcurv 2 3 ¨ π ¨ rcurv

where rcurv is the radius of the spherical convolution kernel and Vinter prcurv q is the portion
of the convolution kernel’s volume that resides inside the surface’s boundary. By computing the
covariance matrix of the intersection instead of its volume, it is possible to design estimators
for the complete curvature tensor and thus estimate likewise principal curvatures and directions
[LAC 17].
Since computations are made on volumes, this oﬀers the opportunity to perform compu-
tations using software such as ImageJ. Using Eq.2.6, κmean can for instance be computed using
a linear convolution. Since an optimized implementation of this method is already available in
the open-source C++ library DGtal, it has been used in the present work [DGt 22].

3D surface characterization methodology | 51

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The only parameter needed to perform the calculation is the radius of the convolution
kernel rcurv , which is the scale at which the curvature is computed. Although this
parameter is of great importance and has a quite intuitive meaning, it can be hard to properly
set in practice. A first constraint that limits the possible values for rcurv is the resolution of
the XCT scan, since a minimum of a few voxels are necessary to limit noise in the measured
curvature. If one aims to minimize the noise while keeping the curvature measurement at a
fine scale, it was found useful to apply a denoising filter after curvature computation. For this
purpose, we used the same Gaussian filter that was used for the S-filter in Section 2.2.2.
The choice of the convolution radius can also be driven by physical considerations and depends
on the scale of the surface features of interest. This is the case in the example shown in Section
2.2.2, where curvature at a large scale can be used to detect biased roughness measurements
near sharp geometrical features. Section 2.4 also discusses the relevance of curvature measured
at different scales to predict crack initiation sites in fatigue applications.
The various definitions of curvatures introduced previously are complementary because they
carry different information, see e.g. Fig.2.9a-e. Depending on the objective, one definition can
be more relevant than others. Mean curvature is for example a relevant parameter concerning
surface tension and wetting issues [BLU 19]. Triply Periodic Minimal Surfaces (TPMS) such as
gyroids, which can be manufactured using AM processes [YUA 19, TIL 23, POL 22, KHR 21],
are also characterized by a zero mean curvature. Such structures were for example found to
achieve interesting energy absorption properties [ZHA 18]. Gaussian curvature provides com-
plementary information. For instance, a surface characterized by a zero Gaussian curvature is a
developable surface (e.g. a cylinder). In the present work, curvature is computed to characterize
surface notches and distinguish them from other surface features. The maximum and Gaussian
curvatures do not seem to be suited for this purpose, as they will have the same value (zero)
both on a perfectly flat surface without any notch and at the root of a linear notch in a plane.
The choice between the other curvatures being less straightforward, Fig.2.9a-e are helpful to
guide our final choice. Fig.2.9b displays an artificial object showing ideal notches with different
geometries. Each of those notches represents a configuration that can be found locally on a real
surface, see Fig.2.9a. Notches A are cups, i.e. areas where κmin ă 0 and κmax ă 0. Notch B
exhibits a saddle-like geometry, with principal curvatures of opposite signs. Notches C are linear
ones in a plane, i.e. κmin ă 0 and κmax “ 0. Two orientations are shown to demonstrate that
it is possible to discriminate different notches based on their orientation.

´Three
¯ curvatures are computed on this artificial object, namely κmin , κmean and κσ . κσ “
κ dσ is the directional curvature along the direction d⃗σ (vertical in this case). The comparison
⃗
of the three curvatures in Fig.2.9c-e makes it possible to identify which one highlights best the
different notches.
The first one is the minimum curvature κmin , which successfully captures all notches.
Furthermore, all have the same κmin value. The second curvature is κmean . Cups (type A) have
the lowest κmean value whereas linear notches (C) show intermediate values. This difference is
not necessarily desirable, since cups are not expected to reduce the mechanical properties more
than linear notches. Even worse, κmean tends to 0 for the saddle-like notch (B). Although this
notch was chosen as an example for the sake of clarity and seems far from a real case, there are
many regions of the surface where κmin ă 0 and κmax ă 0. Thus, κmean seems less relevant
than κmin to detect notches in general because it is somehow biased by the κmax contribution.

52 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Finally, the third curvature is the directional one, κσ . As illustrated in Fig.2.9e, all notches
are well identified except the one parallel to d⃗σ . High κσ values are also more concentrated at
notches roots in comparison with κmin . κσ can thus be considered the most appropriate
choice when one aims at characterizing surface features with a specific orientation
with respect to a loading direction.
Based on these considerations, both κmin and κσ were thus computed on the same cylindrical
sample used in Section 2.2.2, see Fig.2.10a-b. Since the sample is a fatigue specimen meant to be
loaded along its axis, κσ was computed along this direction. A radius of 30 µm (= 6 voxel size)
was chosen here for the convolution kernel. An additional filter was used similarly to what has
been done for roughness, using λS “ 0.025 mm “ 5 voxel size. These choices were made to
provide a sufficiently detailed curvature measurement without being too much affected by noise.
Fig.2.10a shows that κmin underlines the presence of notches, which correspond to the lowest
values. κσ greatly attenuates the vertical notches, which can be seen when comparing enlarged
views in Fig.2.10a and Fig.2.10b.

3D surface characterization methodology | 53

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 2.10 | (a) κmin and (b) κσ measurements on an as-built cylindrical sample
fabricated by E-PBF. Computations were made from an XCT scan with a voxel size of 5 µm,
using rcurv “ 30 µm “ 6 voxel size and λS “ 0.025 mm “ 5 voxel size. The magnifying windows
provide an enlarged view on a notch parallel to d⃗σ . It shows that κσ manages to ignore it,
whereas notches perpendicular to the loading direction are kept.

2.2.4 Quantiﬁcation of the harmfulness of surface notches

The objective here is to combine both roughness and curvature to derive a parameter ac-
counting for the mechanical severity of surface features. For surfaces derived from AM, the
surface features that are expected to have the most significant impact on mechanical properties
are notches [KAN 18, PER 19, PLE 20b]. To account for this notch effect, the simple analyt-
ical formula given in Eq.2.7 is used. It gives the stress concentration factor at the root of an
elliptical notch in a semi-infinite panel, in the absence of plastic deformation [ING 13]. The two
parameters required to estimate the stress concentration factor Kt are the notch depth d and
the radius of curvature at its root ρ (see Fig.2.11).
d
d
Kt “ 1 ` 2 (2.7)
ρ

In the previous sections, d is estimated by the local height, which corresponds to roughness,
while ρ was deﬁned by the inverse of the curvature κσ . Some approximations still have to be
made to apply Eq.2.7 using those parameters.
First, it is important to emphasize that both the roughness and curvature, as calculated
in Section 2.2.2 and 2.2.3, are estimates. Therefore, their values may depend on the method

54 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

σpeak
d
ρ

σpeak

σ
Figure 2.11 | Elliptical notch in a semi-inﬁnite panel, submitted to a tensile stress
σ in the direction d̃σ perpendicular to the ellipse major axis. d and ρ are respectively
the depth of the notch and the radius of curvature at its root. This notch generates a stress
concentration given by Eq.2.7, which means that the local stress at its root σpeak is higher than
the nominal stress in the section.

and parameters used for their measurement. The curvature values, for instance, are notably
inﬂuenced by changes in the convolution radius rcurv and in the voxel size. While there are
theoretical proofs, such as the one presented by [COE 14], showing that integral invariant-based
calculations approach exact curvature values as voxel size reduces, the resolution required to
observe this might be very high. Keeping this in mind, one should treat these estimated values
as semi-quantitative ones, that can for instance be used for ranking the severity of notches.
Another approximation to consider is that Eq.2.7 is valid at the notch tip, which might be
hard to detect automatically. As a result, the formula has been generalized and applied to every
point on the surface where both κσ ă 0 and d ă 0. This means it is used across the entire
notches, not just at their root. While Eq.2.7 might not be applicable for numerous points, the
highest values will still be found at the notch roots where the depth is maximal and the curvature
is minimal. Therefore, the values derived are still relevant, especially when identifying areas with
the highest stress concentration. In other words, it provides a semi-quantitative parameter
that reﬂects the mechanical severity of surface notches. Finally, an approximate value
of Kt , called Kt˚ , can be computed at the sample surface using Eq.2.8.
a
Kt˚ “ 1 ` 2 height ¨ κσ where height ă 0, κσ ă 0 (2.8)
where Kt˚ is the estimated local stress concentration, height is the local surface height
obtained from roughness measurements and κσ is the directional curvature along the loading
direction.
Fig.2.12a-b show an example of a Kt˚ map computed from the same cylindrical sample used
in the previous sections. Roughness and curvature (κsigma ) values used for computations are
the ones presented in Section 2.2.2 and 2.2.3.

3D surface characterization methodology | 55

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 2.12 | Kt˚ maps of a 2 mm diameter cylindrical sample fabricated by E-PBF.
The Kt˚ formula is given in Eq.2.8 and makes use of the roughness and curvature measured in
the previous sections. The curvature κσ is computed in the direction of the cylinder axis, which
also corresponds to the build direction. The points where Kt˚ is not deﬁned − because d ě 0
or height ě 0 − are by default colored in black. The parameters used for computations are
λc “ 0.8 mm | λS “ 0.015 mm for roughness, and rcurv “ 50 µm | λS “ 0.025 mm for curvature.
(a) Surface as seen from the exterior of the sample. (b) Surface as seen from the interior of the
sample. The internal point of view is obtained from the surface extracted from the XCT scan.

Different views of the surface can be provided. The first one in Fig.2.12a is the usual external
view, which corresponds to what can be seen using a conventional 2.5D surface characterization.
The area shown is the same as in Fig.2.6 and Fig.2.10a-b. The two other views shown in
Fig.2.12b, are internal views which can be generated after extracting the surface from the XCT
scan. They offer a unique way to identify deep and sharp notches that would very often be
hidden using conventional characterization methods, see the comparison between Fig.2.12a and
Fig.2.12b. Once again, this illustrates the interest in characterizing the surface as a 3D free-form
one obtained by XCT.

2.3 Application to parts with complex geometries

In order to test the ability of the methodology developed in this work to characterize complex
geometries, three examples were studied: a gyroid structure and an octet-truss lattice structure
manufactured by E-PBF, as well as an impeller manufactured by L-PBF. All would be impossible
to characterize thoroughly using conventional 2.5D characterization tools and methodologies.

56 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Figure 2.13 | 3D characterization of an E-PBF Ti64 gyroid. The 3D roughness maps

shown were obtained from the XCT scan made with a 10 µm voxel size. (a) Picture of the gyroid
sample. (b) 3D roughness map with magnifying windows showing a down-skin and an up-skin
region (λc “ 0.8 mm and λS “ 0.05 mm). c) 3D minimum curvature map with magnifying
windows showing a down-skin and an up-skin region (rcurv “ 50 µm and λS “ 0.05 mm). ()d)
Cumulative distribution function of Kt˚ for both voxel sizes. The same area was used for the
comparison of both scans.

Fig.2.13a shows a picture of the gyroid. Two local tomography scans [STO 08] were acquired
at the center of the structure. Large artifacts were observed on the XCT scans, which made the
use of the presented threshold for bimodal histograms inappropriate. Otsu’s threshold was found
to be more eﬃcient in this case. To examine the inﬂuence of voxel size on roughness, curvature,

Application to parts with complex geometries | 57

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
and Kt˚ measurements, two distinct voxel sizes were employed: 5 µm and 10 µm. In either case,
roughness and curvature calculations were done using λc “ 0.8 mm and rcurv “ 50 µm.
Fig.2.13b shows the measured roughness 3D map obtained from the lowest resolution scan.
Two magnifying windows are also displayed to illustrate a down-skin and an up-skin region.
One can notice, quantitatively, the higher roughness in the down-skin area. The roughness
parameters measured for both voxel sizes are summarized in Tab.2.2. The average roughness Sa
is slightly lower for the lower-resolution scan, which is consistent since a lower resolution tends
to smooth the surface.
The higher maximum height Sz and maximum valley depth Sv are respectively 52 µm and
45 µm higher for the higher resolution scan, compared with the lower resolution one. Since
the diﬀerence is roughly the same for Sz and Sv , it can be concluded in the present case that
sharp and deep notches are slightly better captured using a smaller voxel size.
The improvement is less signiﬁcant for surface features that have positive height values such as
unmelted powder particles.
The skewness Ssk (= asymmetry) and kurtosis Sku (= sharpness) values are also consistent
with this conclusion. The increase of Sku with a higher resolution means that the surface height
distribution contains more extreme values. The decrease in Ssk may also be related to the fact
that notches are better taken into account. Thus, both Sku and Ssk suggest that using a smaller
voxel size allows for a better capture of notches, in particular the deepest ones.

Voxel sizes, Sa , Sv and Sz in µm

Voxel size Sa Sv Sz Ssk Sku
5 37.3 420 812 0.42 3.6
10 36.6 375 760 0.55 3.1

Table 2.2 Roughness parameters measured from the gyroid XCT scans with two voxel sizes
using λc “ 0.8 mm and λS “ 0.05 mm. To ensure consistency between the values measured for
the two voxel sizes, the parameters were evaluated using only the roughness values present on
the surface available in both scans.

Fig.2.13c shows the 3D κmin map on the same surface as Fig.2.13b. Fig.2.13d illustrates
how curvature evolves with respect to the convolution radius rcurv . The mean curvature is used
here since it is known that it should equal zero for an ideal gyroid, i.e. with no roughness.
The average value across the entire surface is represented by the average of absolute values
mean|κmean |, analogous to the role Sa “ mean|height| plays in roughness metrics. As observed
in Fig.2.13d, mean|κmean | tends to zero with increasing rcurv values. This is because a larger
rcurv measures curvature at a larger scale, corresponding more to the gyroid shape (with zero
mean curvature) than microscopic surface features, which have pronounced curvatures.
Finally, Fig.2.13e-f show the Kt˚ measurements derived from roughness and curvature. Note
that since the gyroid could be mechanically loaded in any direction, the minimum curvature is
used instead of κσ to compute Kt˚ . Fig.2.13e shows the 3D Kt˚ map obtained from the scan
performed with a 10 µm voxel size. Meanwhile, the cumulative distribution functions of the Kt˚
parameter for both voxel sizes are given in Fig.2.13f. The obtained Kt˚ values are slightly higher
for the 5 µm voxel size scan. This can, once again, be attributed to the better ability at high
resolution to properly capture sharp and deep notches.

58 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

2.3.2 Octet-truss lattice structure

While the gyroid shows the ability to characterize a complex 3D structure, it is a favorable
example regarding the roughness measurement because it does not have sharp features like
corners where roughness measurements will tend to be biased. To account for this eﬀect, the
second example chosen is an octet-truss lattice structure which, contrary to the gyroid case
presents sharp features. Fig.2.14a shows the κmin 3D map of the octet-truss. Here again, one
has a good perception of the volume and surface topography details when the curvature is used
for the 3D rendering.

Figure 2.14 | 3D characterization of an E-PBF Ti64 octet-truss lattice structure.

(a) 3D minimum curvature map (rcurv “ 50 µm and λS “ 0.05 mm). (b) 3D roughness map
(λc = 0.8 mm and λS “ 0.05 mm) with magnifying windows showing the region near the lattice
node where roughness measurements are biased. (c) 3D mean curvature map computed at large
scale (rcurv = 1000 µm and λS “ 0.05 mm). Nodes, where roughness measurements are biased,
can be identiﬁed thanks to their low mean curvature. The region highlighted in purple in b
corresponds to the one with a mean curvature lower than −0.4 mm−1 .

Application to parts with complex geometries | 59

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.2.14b shows the 3D roughness map using a cut-off wavelength λc “ 0.8 mm. It can be
seen that roughness is successfully measured everywhere on the octet-truss lattice,
except at nodes where points’ height is significantly lower than it should be (see the regions
delineated in purple). This shows the limit of the Gaussian filter to derive the reference surface
and roughness measurements near sharp features.
It is possible to get around this issue by discarding the roughness values in
those regions. This can be done manually or with the help of a mean curvature computation.
In the case of the octet-truss lattice structure, the mean curvature was measured using the
integral invariant approach described in Section 2.2.3 and a large convolution radius of 1 mm,
see Fig.2.14c. It can be seen that nodes are characterized by a very low mean curvature at this
scale. For example, the regions delimited in purple in Fig.2.14b were obtained by thresholding
the mean curvature using a manually chosen value of −0.4 mm−1 .
The limitation of this method is that it does not enable the estimation of roughness at
sharp features. Those may, however, be areas of particular interest. For example, it is the case
if the aim is to quantify the impact of surface roughness on mechanical properties. Indeed,
sharp features will have a stress concentration effect that will add to that due to the presence
of surface notches. Even though this problem is complex and has no ideal solution, more
advanced smoothing methods such as anisotropic diffusion of normals [TAS 02] may
be relevant in such cases as a replacement for the conventional Gaussian filter. They are
indeed used to smooth surfaces while keeping sharp features of the objects form. This could
lead to less biased roughness values at sharp features.
Concerning the Kt˚ values obtained for complex structures, it is worth noting that the com-
puted values estimate only the stress concentration generated by the surface topography at a
micro-scale. For a more complete characterization, one may want to add the contribution of
the macroscopic geometry. The latter may be computed using Finite Element Modeling (FEM)
on the ideal part geometry. The same calculations could also be used to estimate locally the
direction of maximum principal stress. These directions could then be used to compute the
curvature instead of using a single direction for the whole part.

2.3.3 Impeller
Finally, the methodology has been applied to an impeller manufactured by L-PBF to show
a real-case example. To keep a reasonably low voxel size (18.5 µm), only about one-half of the
impeller has been scanned. Fig.2.15a-b and Fig.2.16a-b show the 3D characterization of the
component from two points of view. Roughness was computed using λc “ 0.25 mm and λS “
0.09 mm. The relatively low λc value allows for minimizing the occurrence of biased roughness
values at sharp features such as corners. The curvature is measured using rcurv “ 100 µm, this
value being larger than in the previous cases because of the larger voxel size.
Fig.2.15a-b show well the complex internal structure of the component. Magnifying windows
make it possible to see that even with a 18.5 µm voxel size, the as-built surface roughness can
still be distinguished.
Fig.2.16a-b show the outside of the impeller. First, it is worth mentioning that although
most of the surface seems to appear in red in Fig.2.16a, this is only a matter of point of view.
Indeed, the small square window showing an internal view shows that valleys (in blue) are for
the most part hidden in the external view.

60 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

The magnifying windows show some speciﬁc features such as regularly spaced marks, charac-
teristic of surfaces that were machined after printing. A piece of scrap is also visible on the left
part of the window and may have been generated during the machining step. Because this scrap
is very thin, it leads to very high roughness values („450 µm). This shows the diﬃculty that
can arise when characterizing large real-case components: some unexpected features may
always lead to abnormal values, requiring a cleaning step to be conducted after calculations.

Application to parts with complex geometries | 61

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 2.16 | 3D characterization of a half Ti64 impeller manufactured by L-PBF
(view from the outside). The XCT scan was obtained at the BM18 beamline at ESRF
synchrotron with a voxel size of 18.5 µm. (a) 3D roughness map (λc = 0.25 mm and λS “
0.09 mm). (b) 3D minimum curvature map (rcurv “ 100 µm, without subsequent denoising
ﬁlter).

2.4 Application to the prediction of crack initiation sites in fa-

tigue
The methodology we developed was assessed for its ability to predict crack initiation sites
during fatigue testing. Several studies attempted such predictions [PER 20, AND 19a, PLE 20b,
?
TAM 17]. In many cases, the Murakami parameter area is employed to quantify the defects’
harmfulness. However, its applicability to complex surfaces, like those from AM processes, is
limited. Instead, characterizations based on local roughness and curvature measurements may
offer a more versatile approach for these surfaces.
The method proposed, initially designed for complex surfaces, was first tested on a simpler
scenario to predict crack initiation sites. Specifically, we looked at cases where defects leading to
?
failure were easily delineated, allowing for area measurements. This facilitated a compari-

62 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

tomography: application to additively manufactured parts
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
?
son between our proposed method and the conventional area parameter, evaluating
which approach yielded better, comparable, or less conclusive results.
?
As a reference, we used the results from [PER 20] who employed the area parameter, and
applied our method using the same XCT data to measure roughness, curvature, and Kt˚ .
The method from [PER 20] is summarized below. For more details, the reader is redirected
to the original publication. The samples used are 2 mm diameter E-PBF cylindrical fatigue
specimens which were chemically etched. Only a few surface defects remain after etching, which
?
generally enables their unambiguous delineation. Still, the area measurement had to be done
manually for each defect, which is a time-consuming process. Therefore, for each specimen,
?
the area measurement was done only on a pre-defined set of defects whose depth was below
a certain threshold. These critical defects were chosen based on depth estimates from 2.5D
roughness measurements using XCT scans. Note that a 2.5D approach is not necessarily a
strong limitation in such a case since the surfaces obtained after etching show much less re-
entrant features than more complex as-built surfaces.
After the etched sample was characterized by XCT, they were tested in fatigue (R=0.1). The
fracture surfaces were observed using SEM to identify the initiation site on each specimen, see
Fig.2.18a. An XCT scan of the fracture surface was also acquired. It was possible to identify
the same initiation site that was found based on SEM observations, see Fig.2.18b. Finally, by
comparing the XCT scans before and after fatigue testing, the killer defect that led to failure
was identified and compared to all others.
In the present study, roughness, curvature, and Kt˚ measurements were computed instead of
?
the area parameter. Samples’ surfaces were segmented using the triangle threshold for bimodal
histograms presented in Section 2.2.1. Roughness and curvature were studied individually (even
if Kt˚ is theoretically more relevant) to evaluate their ability to capture the most critical defects.
To study the influence of the cut-off wavelength λc and the convolution radius for curvature
measurements rcurv on the quality of the prediction, a parametric study was conducted using
λc P t0.25, 0.8, 2.5, 8u mm and rcurv P t30, 50, 75, 100, 125, 150, 200u µm. Both roughness and
curvature values were subsequently filtered using a cut-off λ “ 0.025 mm. For the sake of clarity,
only the most relevant results are discussed.
?
Several scores can be used to gauge the efficacy of a given metric (like area or Kt˚ ) in
predicting crack initiation sites. Two complementary approaches are used in this work.
The first method consisted of ranking defects based on their predicted harmfulness.
Thus, if the actual killer defect was given a rank of 1, it means that it was successfully identified.
On the contrary, a rank of 10 means that 9 defects were considered to be more harmful than the
actual killer defect. To determine this rank in our case, the surface voxels belonging to the killer
defect were first identified manually. Afterwards, the most critical value (in terms of height,
curvature or Kt˚ ) of the killer defect was computed. For height and curvature, it corresponds
to the lowest one, whereas for Kt˚ it corresponds to the highest one. Then, the surface was
thresholded using this value, so that only surface voxels considered to be more critical than
the killer defect were kept. The number of remaining voxel aggregates was then counted to
determine the rank of the killer defect (e.g. one remaining aggregate corresponds to a rank of
2). This clustering step can be done using algorithms such as labeling − available in ImageJ
− or DBSCAN [EST 96]. Here, the DBSCAN algorithm was used with a neighbor distance of
10 voxels (= 50 µm) and no threshold on the number of points needed to form a cluster (which
means that no point is considered as noise).

Application to the prediction of crack initiation sites in fatigue | 63

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Our secondary approach for evaluating crack initiation site predictions involved comparing
the metric value of the killer defect against its most severe counterpart on the
entire surface. Specifically, we calculated the ratio between these values, which will be referred
hereafter as the percentage score. This complements the ranking approach. For instance, one
can consider a killer defect ranked at 20 using the Kt˚ parameter. While this rank might suggest
Kt˚ is not ideally suited for pinpointing the crack initiation site, the Kt˚ value at the killer defect
might still be 4.5, with the peak value across all defects being 5. In this scenario, the killer
defect’s severity is 90% of the maximum, implying that other factors like microstructure might
account for the 10% gap. Consequently, the prediction may not be as off as the rank suggests. In
this case, it could mean that the parameter gave a good quantification of the impact of surface
topography on crack initiation, but that many defects actually have similar harmfulness. In
such a case, other factors than surface topography such as microstructure should probably be
taken into account and could explain why a certain defect caused crack initiation rather than
another.
Still, this percentage score should always be interpreted with caution. The first
reason is that several defects may seem to have comparable harmfulness because of an insufficient
resolution of XCT scans. Indeed, with an insufficient resolution, the sharpest notches’ roots will
tend to be smoothed. As a consequence, curvature measurements will tend to saturate to a
certain threshold, and notches depth may also be sometimes underestimated. As a consequence,
all the most severe notches may lead to similar roughness, curvature and Kt˚ values, although
this does not reflect the reality. This effect of XCT scan resolution was discussed to some
?
extent in Section 2.3.1. A second reason is that area, curvature, roughness and Kt˚ estimate
different physical quantities. Therefore, direct percentage comparisons between these metrics
can be misleading. An example of this is that it may seem just as relevant to consider the
?
stress intensity factor associated with the area parameter rather than this parameter itself,
?
using Eq.2.9. This amounts to assimilating the defect with a crack of the same area [MUR 02,
p.16-21]. Due to the square root presence in Eq.2.9, the percentages are substantially altered,
despite evaluating the same defects. As such, when analyzing percentage values, especially
across different physical properties, it is paramount to proceed with caution and compare them
alongside ranks.

?
b
KImax « 0.65 ¨ σ ¨ π area (2.9)
where KImax is the maximum stress intensity factor along the front of the crack that would
emanate from the surface defect considered and σ is the uniform tensile load applied, in our
case, along the vertical d⃗σ direction).
Tab.2.3 and Tab.2.4 showcase respectively ranks and percentages scores from ten samples,
corresponding to samples 1-10 in [PER 20]. Note that in the second column giving the rank
?
obtained from the area parameter, the second number corresponds to the number of defects
that were pre-sectioned by [PER 20] for the manual measurement. Colors give information
about the quality of the prediction: a green cell indicates a good prediction and the color tends
to red as the quality of the prediction decreases.
From what has been said up to now, the most natural parameter in our study would be
Kt˚ (λc “ 0.8 mm | rcurv “ 30 µm). The choice of λc is dictated by the estimated size of
the relevant surface features (see Section 2.2.2) and rcurv is chosen as small as enabled by the
scan resolution. Although for most samples the killer defect has a rank inferior to 10, none is

64 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Percentages
?
area κσ Kt˚
λc “ 0.8 mm
[PER 20] rcurv “ 100 µm
Sample rcurv “ 100 µm
1 48 84 71
2 86 97 89
3 74 100 82
4 78 94 92
5 61 98 89
6 53 89 91
7 85 94 89
8 97 100 97
9 100 100 100
10 61 66 83
Median 76 96 89
?
Table 2.4 | Killer defect predictions percentage scores for area, κσ , and Kt˚ pa-
rameters. Scores are computed for samples 1 to 10 from the study of [PER 20]. The median
percentages over all samples are also computed for each parameter.

Application to the prediction of crack initiation sites in fatigue | 65

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
successfully identified and two are clearly mis-estimated. The obtained median rank over all
?
samples is 7, which is not as good as the one achieved using the area parameter (4). Besides,
Kt˚ measured with these parameters is hardly more relevant than roughness alone, and both
show in the end similar results. As for κσ (rcurv “ 30 µm), although it shows good results for 4
samples (rank inferior to 3), it is irrelevant for other samples.
It may not be particularly surprising that κσ (rcurv “ 30 µm) alone leads to poor predictions
since it lacks information at a larger scale, which is given by roughness. However, the fact it
does not bring significant improvement when combined with roughness could also be due to a
lack of resolution. Indeed, with a 5 µm voxel size, notch roots tend to be overly discretized
and therefore smoothed. It may thus be difficult to obtain relevant curvature measurements at
such a small scale. Another interpretation is that curvature at a small scale may not be
the most relevant parameter when trying to predict fatigue crack initiation. This is in a
?
sense one of the assumptions made when using the area parameter since it already assimilates
the defect as a crack and therefore does not consider the curvature at its root.
However, when measured at a larger scale, it is observed that curvature provides
interesting predictions. Fig.2.17 presents the evolution of the median values and standard
deviations of ranks obtained for all samples for a given rcurv . A good prediction capability is
characterized on this graph by a small median rank and a small standard deviation as well. An
optimum can be seen around 100-125 µm. For rcurv “ 100 µm in particular, the obtained median
rank is 3. As can be seen in Tab.2.3, only three killer defects got a rank higher than 4 − namely
for samples 1, 4 and 10.

Median (Rank) Std (Rank)

14 30
Median(Rank)
12 Std(Rank)
25
10
20
8
15
6

4 10

2 5
50 100 150 200
rcurv (µm)
Figure 2.17 | Quality of initiation sites prediction by κσ , depending on the rcurv used
for curvature computation. The quality of predictions is quantiﬁed through the median and
the standard deviation of ranks obtained over all samples. The green vertical line corresponds
to the mean of maximum depths Sv measured for all samples with λ “ 0.8 mm.

66 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Figure 2.18 | Some illustrations of the samples used for the killer defect prediction
study. Samples and XCT volumes used are the same as in [PER 20] (E-PBF Ti64 chemically
etched cylinders). (a) Fracture surface of sample 8, observed by SEM-SE. (b) Fracture surface
of sample 8, extracted from XCT volume. (c) Surface of sample 8 before fatigue testing, where
the killer defect is highlighted with a red arrow. (d-f) Curvatures and Kt˚ maps computed for
sample 8. (g-j) Curvature, Kt˚ and roughness maps computed for sample 1.

One can also notice that the percentage score for rcurv “ 100 µm is very high for most
samples, with a median percentage of 96%. In other words, it is observed that in most cases,
crack initiation occurs on a surface feature whose curvature κσ (rcurv “ 100 µm) is close to
the minimum (maximum in absolute value) reached on the whole surface. This highlights the
relevance of κσ (rcurv “ 100 µm) parameter, but as discussed above, it could also suggest that
most severe notches root tend to be smoothed, leading to similar value for many of them.

Application to the prediction of crack initiation sites in fatigue | 67

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The given results raise the question of why curvature provides much better results for rcurv “
100 µm than for rcurv “ 30 µm. A possible explanation is that in this case, curvature is not
measured at the scale of notches roots anymore but rather at the scale of the entire defects
themselves. Therefore, it can better take into account the severity of the whole defect, by taking
into account its sharpness at a large scale but also its size − a contribution that is normally
provided by the depth measurement in Eq.2.7. Thus Fig.2.19b shows that by using a radius rcurv
of the order of magnitude of the depth of the deepest notch, it is possible to efficiently make the
distinction between a deep notch and a shallow notch using curvature measurements. On the
contrary, Fig.2.19a shows that using a small radius provides more local information and thus
will not be able to make the difference between two notches of different depth but with the same
curvature at their roots. On this topic, it is interesting to note that the average maximum valley
depth Sv measured over all samples (using λ “ 0.8 mm) is 115 µm. As can be seen in Fig.2.17,
this corresponds well to the optimal rcurv , which is consistent with the proposed interpretation.
This also suggests that Sv could be a first guess to set rcurv parameter appropriately,
although more tests on more diverse surface states would be required to confirm this hypothesis.

a b Shallow notch

Deepest notch Shallow notch Deepest notch

Sv Sv
κdeepest = κshallow

rcurv κdeepest ≠ κshallow

rcurv << Sv
rcurv
rcurv ≈ Sv

Figure 2.19 | Schematic example showing curvature measurements on two notches

with identical curvatures at their root but diﬀerent depth. One of the notches is the
deepest notch on the whole surface (i.e. its depth equals to Sv roughness parameter) while the
other is a rather shallow notch. The curvature measurement is done using (a) a small convolution
radius rcurv taking into account notches roots only, (b) a large rcurv of the order of magnitude of
Sv . A comparison between (a) and (b) shows that computing curvature with the largest radius
enables the diﬀerentiation of both notches, thus taking into account the notch depth. On the
contrary, a more local curvature measurement using a small radius will not enable to distinguish
the two notches.

Since rcurv “ 100 µm brings better results than rcurv “ 30 µm − when curvature is considered
alone, it seems interesting to see if the predictions obtained with Kt˚ are also better when
computing curvature at this scale. Tab.2.3 shows that it is not really the case, the results
obtained being similar to what was obtained with rcurv “ 30 µm. It can still be noted that
for two of the three "worst-case scenarios" (Samples 1, 4 and 10), the prediction is largely
improved by the the addition of roughness. This may come from the fact that in this case,
roughness still better takes into account the topology at a larger scale than curvature, thus
providing relevant complementary information in some cases. Furthermore, it is interesting to

68 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Application to the prediction of crack initiation sites in fatigue | 69

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary
This chapter presented a methodology for the 3D characterization of surface topography
using XCT data, focusing on measuring both roughness and 3D curvature. While this
methodology has broader applications, we emphasized its usefulness on samples produced
by additive manufacturing and dedicated to mechanical applications.

Key ﬁndings can be summarized as follows:

• The proposed methodology could effectively account for hidden surface features on
surfaces inherited from E-PBF. For example, the 3D method can identify notches that
traditional 2.5D techniques would miss.
• The method effectively captures roughness and curvature in complex geometries such
as architected structures. While the results are encouraging, certain geometric fea-
tures, like sharp edges, present challenges in roughness assessments. In such instances,
more sophisticated metrological tools might offer deeper insights.
• We have tailored the methodology to make it accessible, even for those unfamiliar
with advanced data analysis tools and programming. For example, we deliberately
used standard image analysis techniques, such as 3D Gaussian filtering, to extract
roughness from the raw surface. This approach can be applied using popular software
like ImageJ. With the Gaussian filter being a standard operation for the analysis of 2D
and 2.5D roughness, guidelines from ISO standards can be adapted for the presented
3D workflow.
• Several tools have been introduced to leverage this 3D characterization in understand-
ing the mechanical implications of surface notches. For instance, it was found that the
standard Otsu’s thresholding missed some of the sharpest notches. We proposed an
alternative, the triangle threshold for bimodal histogram, which yielded better results
in this case. Additionally, 3D curvature measurements enable the derivation of curva-
ture in the direction of principal stress, κσ , underscoring the mechanical consequences
of notches aligned perpendicular to the loading direction. Finally, we also proposed
a model that integrates roughness and curvature data to compute a parameter, Kt˚ ,
that reflects the stress concentration induced by surface notches.
• Employing roughness, curvature, and Kt˚ metrics, we sought to predict crack initiation
during fatigue. These predictions paralleled the findings by [PER 20], which studied
the fatigue behavior of chemically etched E-PBF cylinders. Only a few defects re-
mained after etching, which made it possible to quantify their harmfulness using the
?
conventional area metric. Broadly speaking, our crack initiation site predictions
?
were comparable to the ones found using the area parameter. Notably, among
roughness, κσ , and Kt˚ , κσ demonstrated superior predictive ability. For instance, us-
ing κσ (rcurv “ 100 µm), the actual crack initiation site was among the 3 most severe
defects identified for 6 samples over 10. In a more general case, the proposed method-
ology offers new opportunities, being adaptable to surfaces with advanced topological
?
complexity, a challenge for the area metric. Furthermore, improved results could be
achieved by characterizing the surface at multiple scales, modifiable through parame-
ters like the cut-off wavelength λc and the convolution radius rcurv , used for roughness
and curvature measurements.

70 | Chapter 2 – A methodology for the 3D characterization of surfaces using X-ray computed

Materials characterization

Contents
3.1 Samples manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.1.1 Processing conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.1.2 Fatigue and tensile specimens . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2 Microstructure characterization . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.1 Microstructure of the bulk material . . . . . . . . . . . . . . . . . . . . . . . 76
3.2.2 Microstructure near the surface . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 Tensile properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4 Defects characterization using XCT . . . . . . . . . . . . . . . . . . . . . 84
3.4.1 Acquisition setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4.2 Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4.3 Surface defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

In Chapter 2, various materials were discussed through multiple case studies. However, the
next chapters of this manuscript will focus on a single material, namely Ti64 manufactured by
L-PBF. The upcoming sections will concentrate on some of its main characteristics, namely its
manufacturing process, microstructure, tensile properties, and defects.
It is worth noting that the manufacturing conditions for the samples match those used in
Quentin Gaillard’s work [GAI 23a]. Gaillard’s research mainly investigated microstructural and
static properties after diﬀerent heat treatments. To gain a comprehensive understanding of these
aspects, readers are therefore encouraged to refer to Gaillard’s work, which serves as a valuable
resource for this chapter.

3.1 Samples manufacturing

3.1.1 Processing conditions
The samples studied hereafter were manufactured by Laser Powder Bed Fusion (L-PBF)
using a FormUp 350 AddUp machine of the ﬁrst generation. The production was carried out
by the French company Dassault Aviation at Argonay (Savoie, France). The machine used was

| 71

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
equipped with two YAG 500 W lasers, operating at a wavelength of 1070 nm. Each laser was
responsible for one-half of the build plate.
The samples were made from Ti-6Al-4V Grade 5 powder produced by TEKNA through
plasma atomization. The results of particle size characterization by sieving are presented in
Tab.3.1. In addition, it is worth noting that 1% in mass of powder particles were found to have
a size larger than 25 µm.

D10 D50 D90

8 µm 14 µm 20 µm

Table 3.1 | Powder particle size characterization, performed by sieving.

The chemical composition of the powder can be found in Tab.3.2. It is worth highlighting
that only fresh powder was employed. This decision was made to minimize variations in oxygen
contamination, defect formation and mechanical properties, as these can arise from the recycling
of powder [MAH 22].

Ti Al V Fe O C N H
Base 6.22 3.87 0.18 0.14 0.009 0.02 0.004

Table 3.2 | Powder chemical composition (in mass %).

The powder bed was deposited using a roll. Although it enables obtaining a compact powder
bed, it requires some caution for the manufacturing of thin elements, as freshly printed thin layers
might be swept away by the roll. Thus, the printing of vertical specimens (presented thereafter)
required the use of "sarcophagus-like" support structures that consolidated the powder bed and
prevented specimens from bending towards the rolling direction. See Fig.3.4 and Fig.3.5 for
examples of such support structures.
Within the build chamber, a controlled atmosphere was maintained, with a laminar flow of
argon (Ar). The build plate did not undergo any pre-heating during the printing process.
Other processing parameters such as the laser power and scanning speed cannot be given for
confidentiality reasons. However, the scanning strategy is illustrated schematically in Fig.3.1,
as it is of primary importance to understand subsequent descriptions of microstructural features
and defects.
Each layer underwent scanning in two stages. First, the bulk material was melted using a
cross-hatching approach with a 90° angle between each layer. Subsequently, two contour scans
were implemented. The first contour scan is positioned on the outermost part of the sample.
Following the manufacturing process, the build plates were subjected to cleaning within an
ultrasonic bath containing water. This step is required to eliminate loose powder particles. The
build plates (plate + samples) were not sandblasted, primarily to prevent changes in surface
roughness and to avoid obscuring surface defects like valleys.
A stress-relief heat treatment of 2 h at 720 °C was then applied. XRD measurements
done by Quentin Gaillard [GAI 23a] showed that for the material considered, this heat treatment
achieves a complete relaxation of residual stresses. Here, the heat treatment was applied on the
entire build plate to prevent potential deformations that might arise if samples were detached
before undergoing stress relief.

72 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

90° rotation

Layer n

Caption

Cross-hatching bulk scanning strategy

First contour scan

Second coutour scan

Figure 3.1 | Scanning strategy adopted for sample manufacturing. Bulk material is
ﬁrst melted using a cross-hatching strategy, followed by two contour scans. A 90° rotation is
done between each layer.

Subsequently, specimens were detached from the build plate, and the support struc-
tures were manually removed. A ﬁnal ultrasonic bath was used to clean the samples, removing
residual powder particles trapped between the specimens and support structures. At this stage,
the samples were prepared for subsequent characterization.

3.1.2 Fatigue and tensile specimens

The samples used in the present PhD work were either ﬂat tensile or cylindrical fatigue
specimens. Their geometry is given in Fig.3.2 and Fig.3.3 respectively.

Figure 3.2 | Drawing of the ﬂat tensile specimens. The dimensions are given in mm. The
specimen width equals 3.00 ˘ 0.02mm.

Samples manufacturing | 73

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

9
R2

43
9

R20

Figure 3.3 | Geometry of the cylindrical fatigue specimens. Dimensions are given in mm.

Tensile specimens were printed in three orientations: vertical ((along the Z-axis), at a
45° angle with respect to the build direction, and horizontal (XZ). Pictures of the printed
samples with their support structures can be seen in Fig.3.4. The vertical specimens employed
"sarcophagus-like" support structures, while more standard supports were used for the 45° spec-
imens. Although not visible in Fig.3.4, the only points of contact between the supports and
specimens are located on the specimen heads. The middle portion of the support structures ap-
proaches the specimen without direct contact. Additionally, small pillars were used as supports
within the gauge length of the XZ specimens. However, after being manually removed, these
pillars left noticeable surface irregularities that are expected to adversely impact the mechanical
properties, please refer to Section 3.3 for further details.

For the fatigue specimens, two build orientations were investigated: vertical (Z) and
45°. Images of both types of specimens are shown in Fig.3.5. Similar support structures as
those used for the tensile specimens were employed. The primary distinction is that the 45°
specimens are not fully supported within the gauge length. Similar to the tensile specimens, the
only points of contact between these supports and the specimens are located on the specimen
heads.

74 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.4 | Pictures of printed tensile specimens and their support structures.
Specimens were printed in the (a) vertical (Z), (b) 45° and (c) horizontal (XZ) directions.
Pictures courtesy of Quentin Gaillard.

Figure 3.5 | Pictures of printed fatigue specimens and their support structures.
Specimens were printed in the (a) vertical (Z) and (b) 45° directions. Pictures courtesy of
Quentin Gaillard.

Samples manufacturing | 75

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
3.2 Microstructure characterization
All microstructure characterizations presented in this section are directly drawn from Quentin
Gaillard’s work [GAI 23a, GAI 23b]. They have been summarized to extract the most relevant
conclusions that contribute to the interpretation of the results in this manuscript. However, for
a more comprehensive understanding of this subject, readers are advised to refer to [GAI 23a,
GAI 23b], which provides a detailed and well-illustrated study. It is worth reiterating that
the manufacturing conditions employed in his study align with those used in the current work.
Notably, despite 15 mmˆ15 mmˆ15 mm cubes being used in Gaillard’s study for microstructure
characterization, the resulting microstructure mirrors that of the fatigue specimens employed in
this study.
Microstructure analysis involved the use of EBSD (Electron Back Scattered Diffraction)
maps, revealing the crystallographic orientation of grains. All these maps were generated using
the build direction (Z) as the reference.
The showcased maps can be classified into two types. The first type pertains to the α phase
orientation map, depicting the actual microstructure of the material following printing and the
subsequent stress-relief heat treatment. The second map represents the reconstructed prior β0
phase orientation map, illustrating the state of the β phase immediately after solidification and
before the subsequent martensite transformation − i.e. during the L-PBF process. Here, it
is reconstructed by using the α phase orientation map, following the procedure described in
[GAI 23a].

3.2.1 Microstructure of the bulk material

Fig.3.6a-f and Fig.3.7a-f illustrate EBSD maps depicting the bulk microstructure in a vertical
(Z) plane and a horizontal (XY) plane, respectively. In particular, Fig. 3.6a and Fig.3.7a display
α phase orientation maps, while Fig. 3.6d and Fig.3.7d showcase prior β0 phase orientation maps.
In Fig.3.6d, it is noticeable that immediately after solidification, the prior β0 phase shows
a columnar structure growing along the build direction, which is characteristic of addi-
tively manufactured materials. The XY map in Fig.3.7d exhibits a distinct chessboard pattern,
resulting from the cross-hatching scanning strategy employed for printing and the 90° rotation
between each layer.
The inverted pole figures, depicted in Fig.3.6e-f and Fig.3.7e-f, reveal that the prior β0 phase
is strongly textured. Specifically, the prior β0 grains were oriented with the (001) plane parallel
to Z, and both the (100) and (010) planes were aligned with the laser scanning directions.
Following rapid cooling, the prior β0 phase undergoes martensite transformation. Columnar
β0 grains transform into fine α1 (martensite) laths. Subsequent stress-relief (2 h at 720 °C)
initiates the martensite decomposition α1 Ñ α ` β. [GAI 23a] estimated, using Differential
Scanning Calorimetry (DSC), that this transformation can start at temperatures as low as
470 °C, well below the heat treatment temperature. The resulting microstructure consists of
a mixture of entangled α laths along with a minor fraction of β phase (2.2 ˘ 0.3%
according to [GAI 23b]). It is actually challenging to definitively ascertain whether the laths
consist exclusively of α phase or if some α1 remains. However, since this matter is not central
to the primary focus here, the subsequent analysis will proceed with the assumption that the
martensite decomposition is completed and only the α phase is present.

76 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.6 | EBSD analysis of the bulk material in a vertical plane (i.e. parallel to
the build direction Z), adapted from [GAI 23a]. Crystallographic orientation maps of (a)
α and (d) prior β0 phases. Pole ﬁgures of (b) α and (e) prior β0 phases. Inverse pole ﬁgures
of (c) α and (f) prior β0 phases. The build direction Z is taken as a reference. The maps are
1.25 mm ˆ 1.1 mm wide and are collected with a step size of 0.6 µm.

Microstructure characterization | 77

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.7 | EBSD analysis of the bulk material in a horizontal plane (i.e. perpen-
dicular to the build direction Z), adapted from [GAI 23a]. Crystallographic orientation
maps of (a) α and (d) prior β0 phases. Pole ﬁgures of (b) α and (e) prior β0 phases. Inverse
pole ﬁgures of (c) α and (f) prior β0 phases. The build direction Z is taken as a reference. The
maps are 1 mm ˆ 0.6 mm wide and are collected with a step size of 0.6 µm.

Fig.3.6a and Fig.3.7a provide insight into the morphology and crystallographic orientation
of α laths. Due to crystallographic constraints dictated by the Burger orientation relationships,
these laths can have only 12 orientations relative to a given prior β0 grain orientation. This
leads to a less pronounced texture compared to the prior β0 phase. EBSD data indicate
an average thickness of α-laths of 0.5 µm, correlating with a relatively high hardness of
394 ˘ 4 HV1 kg .

3.2.2 Microstructure near the surface

Noticeable diﬀerences exist between the microstructure within the bulk material and that
near the surface. Fig.3.8a-c provide a clearer representation of three distinct regions. The
bulk material is found at approximately 600 to 800 µm beneath the surface. The second re-
gion, referred to as the sample border, spans from about 150 µm below the surface. Lastly, the
third region termed the sample contour, encompasses a depth of around 150 µm, approximately
corresponding to the two laser contour scans conducted during printing.

78 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.8 | Optical microscope observation of the microstructure in a horizontal
plane (perpendicular to build direction), adapted from [GAI 23a]. (a) Large view
revealing the presence of three regions (bulk, border and contour). Zoomed view in (b) border
area and (c) bulk area.

Given that the microstructure of the bulk material has been previously discussed, the sub-
sequent focus will be solely on the distinctive features of the border and contour zones. In this
context, EBSD maps were generated near the surface within vertical (Z) and horizontal (XY)
planes − refer to Fig.3.9a-j and Fig.3.10a-j respectively. Similar to the bulk microstructure,
these maps characterize both α and prior β0 phases.
The border region (spanning from 150 µm to 600 to 800 µm beneath the surface) exhibits
a microstructure similar to that of the bulk material. Nevertheless, the prior β0 grains in
this zone appear to be twice the size of those in the bulk material − see Fig.3.8b-c and
Fig.3.10f. An explanation has been proposed by [GAI 23a] to rationalize this phenomenon. It
hypothesizes that at the beginning of a laser track, the preceding adjacent laser track might not
have had suﬃcient time to fully solidify near the surface. Consequently, in such instances, β0
grains might solidify from a melt pool twice the usual size, accounting for the observation of
larger β0 grains.
The contour area shows a notably distinct microstructure. In the immediate sub-
surface region (approximately within 0 to 50 µm), small prior β0 grains are apparent, elongated
in a direction close to Z. The slight deviation from the perfect alignment along the vertical axis
can be attributed to local variations in thermal gradients near the surface. Beyond 50 µm, prior
β0 grains exhibit elongation both towards Z and perpendicular to the surface.
Across all these regions and regardless of the prior β0 grain morphology and texture, the mi-
crostructure following printing and heat treatment consists of ﬁne α laths, with minor texturing.

Microstructure characterization | 79

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.9 | EBSD analysis of the microstructure near the surface in a vertical plane
(i.e. parallel to the build direction Z), adapted from [GAI 23a]. Crystallographic
orientation maps of (a) α and (f) prior β0 phases. Pole figures of (b) α and (g) prior β0 phases
in the border area. Inverse pole figures of (c) α and (h) prior β0 phases in the border area. Pole
figures of (d) α and (i) prior β0 phases in the contour area. Inverse pole figures of (e) α and (j)
prior β0 phases in the contour area. The build direction Z is taken as a reference. The maps are
1 mm ˆ 1 mm wide and are collected with a step size of 0.5 µm.

80 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.10 | EBSD analysis of the microstructure near the surface in a horizontal
plane (i.e. perpendicular to the build direction Z), adapted from [GAI 23a]. Crys-
tallographic orientation maps of (a) α and (f) prior β0 phases. Pole figures of (b) α and (g) prior
β0 phases in the border area. Inverse pole figures of (c) α and (h) prior β0 phases in the border
area. Pole figures of (d) α and (i) prior β0 phases in the contour area. Inverse pole figures of
(e) α and (j) prior β0 phases in the contour area. The build direction Z is taken as a reference.
The maps are 1.45 mm ˆ 0.5 mm wide and are collected with a step size of 0.4 µm.

Microstructure characterization | 81

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
3.3 Tensile properties
Tensile tests were conducted at room temperature using an MTS 250 machine with a 100 kN
load cell. The displacement rate was set to 0.525 mm/min. A GOM ARAMIS 6M Digital Image
Correlation (DIC) system was employed to measure strain.
The investigation involved two Z, three 45°, and three XZ specimens. For stress calculations,
the sections were assessed using XCT − refer to Section 3.4 for detailed XCT data acquisition
information. In this context, all specimens were scanned over a length of 2.3 mm, with a voxel
size of 6.5 µm. The mean of all XCT scan sections along the 2.3 mm distance was then computed
for each specimen and employed for stress computation.
Tensile curves are depicted in Fig.3.11. The mechanical properties were extracted from the
tensile response, including the Yield Stress (YS) measured at a plastic deformation of 0.02%,
the Ultimate Tensile Strength (UTS), and the elongation to failure (ϵf ).

1200
Engineering stress (MPa)

1000
YS (MPa) UTS (MPa) εf (%)
800 Z 1098 1147 7.8

45° 1087 ∓ 6 1145 ∓ 6 5.5 ∓ 0.8

600
XZ 1059 ∓ 4 1099 ∓ 8 0.8 ∓ 0.4

400
Vertical (Z)
45°
200
Horizontal (XZ)
0
0 2 4 6 8 10
Engineering deformation (%)

Figure 3.11 | Tensile curves measured for the three build directions. Values for YS,
UTS and ϵf are presented in the form (mean ˘ std). Standard deviation is not measured for Z
specimens, since two values are not enough to get a meaningful estimation.

The measured YS and UTS values exhibit strong reproducibility and consistency with re-
sults from the literature [NGU 22, GRE 17, KAS 15, BEN 18, VAN 12, CAI 15]. The slightly
lower strength of the horizontal (XZ) specimens can likely be attributed to the significant pro-
tuberances that remain after support structure removal. These defects substantially increase
the XCT-measured section of the specimen, though they are not expected to contribute to the
strength of the material.
In contrast to tensile strength, elongation to failure demonstrates substantial variation, de-
pending on the build orientation. Specifically, XZ specimens display very little defor-
mation before failure (ϵ¯f “ 0.8%). Once again, this could partially stem from support
structures within the gauge length, which leave pronounced surface defects. However, findings
from [GAI 22] suggest that an additional factor contributes to this lack of ductility. This study
reveals that even for tensile specimens with the bottom surface (affected by support structures)

82 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
machined, an unusually low elongation to failure is observed. Notably, the tensile specimens
used in [GAI 22] share the same manufacturing conditions as the samples in this study, except
for the machining step.
Microstructure might be responsible for this phenomenon. Indeed, Fig.3.12b reveals
a linear pattern on the fracture surface of XZ specimens, which is absent for the Z specimen in
Fig.3.12a. These lines are interpreted as indications of delamination at the interface between
prior β0 columnar grains. In the case of XZ specimens, these columnar grains are oriented
perpendicularly to the loading direction.

Figure 3.12 | SEM-SE observations of tensile fracture surfaces for (a) an XZ and (b)
a Z specimen. On the XZ specimen, delamination can be seen at the interface of prior β0 grain
boundaries, leaving linear patterns on the fracture surface (see orange arrows).

Tensile properties | 83

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
For more comprehensive insights into the tensile properties, readers are encouraged to refer
to the work of [GAI 23a], who characterized material manufactured under identical conditions as
in this study. Additionally, this work provides further details about the impact of microstructure
and its evolution under various heat treatments.

3.4 Defects characterization using XCT

3.4.1 Acquisition setup
XCT was extensively employed to scan a relatively large portion of the gauge length of both
tensile and fatigue specimens, primarily to assess their defects. The scanning procedure used a
cone beam phoenix | x-ray V | tome | x-laboratory tomograph, operating at a voltage of 140 kV
and a current of 80 µA, while acquiring 1200 projections. To counteract beam hardening artifacts,
a 0.1 mm Cu filter was introduced during the acquisition. Following the acquisition, radiographs
underwent a 3 ˆ 3 median denoising filter. Reconstruction was subsequently performed using
a standard filtered back projection algorithm facilitated by the proprietary software phoenix
datos|x. A numerical filter was additionally employed during reconstruction to further mitigate
the remaining beam hardening artifacts.
For each specimen, three scans were conducted along its longitudinal axis. The number of
scans was chosen primarily to ensure that for fatigue specimens, failure always occurs in an area
that was characterized by XCT. Details regarding the voxel size used, the resulting characterized
length (taking into account the 3 scans), and the number of specimens scanned are summarized
in Tab.3.3, for both tensile and fatigue specimens.

Tensile specimens Fatigue specimens

Number of scanned specimens Z: 2 - 45°: 3 - XZ: 3 Z: 29 - 45°: 27
Voxel size (µm) 6.5 3
Characterized length (mm) 23 11

Table 3.3 | Number of specimens scanned by XCT, voxel size and characterized length for
tensile and fatigue specimens.

Following reconstruction, the scans were converted into 8-bit unsigned volumes and merged
using an automated Python script developed in-house. To reduce noise, a 2 ˆ 2 ˆ 2 median filter
was subsequently applied to the resulting volume. This volume was then directly employed for
defect characterization, as detailed in the next sections.
As discussed in Chapter 1, defects can be classified into two primary groups: internal defects
(primarily pores) and surface defects. Due to their distinct shapes, origins, and methods to
characterize them, the two categories are examined separately in dedicated sub-sections. The
focus of this characterization, in both cases, is confined to the L-PBF Ti64 fatigue specimens.
The analysis was made using 29 Z specimens and 27 45° specimens.

84 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
3.4.2 Porosity
To perform a quantitative characterization of porosity, pores need first to be segmented from
XCT volumes. Here, like often, this step is done by thresholding. Since conventional automatic
threshold calculation methods were found to miss too many pores, a higher threshold value
has been used. Specifically, an intermediate value between conventional Otsu’s threshold and
the triangle threshold for bimodal histograms (as introduced in Chapter 2) was chosen. The
limit was set at 70%, meaning that the chosen threshold equals the sum of 70% of the triangle
threshold for bimodal histograms and 30% of Otsu’s threshold.
Despite applying a median filter to XCT scans before thresholding, residual noise remained,
leading to segmentation errors where some voxels were mistakenly identified as porosity. To
address this, all pores with an equivalent diameter (defined by Eq.3.1) smaller than 12 µm (=
4 voxels) were considered noise and subsequently removed.
Following this procedure, the resulting porosity is below 0.002% for all samples analyzed.
This remarkably low value confirms that the process parameters were efficiently optimized
to achieve a nearly fully dense material.
To further enhance the characterization, pores were evaluated in terms of equivalent diameter,
sphericity, and distance from the surface. The equivalent diameter Deq is defined as the diameter
of the sphere having the same volume as the pore (see Eq.3.1).
c
3 6¨V
Deq “ (3.1)
π
where V is the volume of the pore. It is calculated using the marching cubes algorithm [LOR 87].
Sphericity S offers insights into pore morphology, representing the ratio of the area of a
sphere with the same volume as the pore to the actual surface area of the pore (cf. Eq.3.2). A
value of 1 denotes a perfect sphere, while lower values indicate irregularly shaped pores. Fig.3.13
illustrates pores with varying sphericity values.
2
p6V q {3
1{3
π
S“ (3.2)
A
where V and A are respectively the volume and the surface area of the pore. The marching
cubes algorithm was used to measure both. This is an important computational detail because
using the number of voxels to measure volume, as is often done, can lead to noticeable errors
for small pores. In this study, for example, it resulted in many pores having a sphericity greater
than one, which is theoretically impossible.

Sphericity
0.52 0.61 0.7 0.8 0.9

Figure 3.13 | Scale showing the evolution of pore morphology with increasing sphericity. Pores
surfaces were extracted using the marching cubes algorithm.

Defects characterization using XCT | 85

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.3.14b-d display distributions for equivalent diameter, sphericity, and distance from the
surface. These distributions make easier quantitative comparisons between the two specimen
build orientations (Z and 45°), as well as variations between specimens of the same orientation.

a b
Z (29)
Total number of pores 35
Pore density (#pore / µm)

(#pore / µm / specimen)
in specimen: 100 45 (27)
30

Pore density
≈ Number of pores in the material
25
whose equivalent diameter Deq is
within a window of 1 µm around 20 µm
20

Number of pores below 20 µm 15

= green area
= 58 10

0
0 10 20 30 40 50 60 70
0 20 Deq (µm) Deq (µm)

c d
2000 Z (29) Z (29)
(#pore / µm / specimen)

1750 45° (27) 4 45 (27)

(#pore / specimen)

1500
Pore density
Pore density

3
1250

1000
2
750

500
1
250

0 0
0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 250 500 750 1000 1250 1500 1750
Sphericity Distance to surface (µm)

Figure 3.14 | Distributions of pores equivalent diameter Deq , sphericity and distance
to surface. (a) Schematic example of pore distribution. Distributions are here computed so
that their integral equals the number of pores in the specimen. (b) Deq , (c) Sphericity and
(d) Distance to surface distributions. For each graph, lines correspond to mean distributions
obtained for all samples while shaded areas represent an interval of one standard deviation
from mean lines. Mean distributions are computed separately for vertically built (Z) and 45°
specimens. The number of characterized specimens for each build orientation is speciﬁed in the
graph legend.

To accomplish this, normalized probability distributions were ﬁrst computed individually for
each specimen through kernel density estimation (KDE) using a Gaussian kernel. The bandwidth
was determined following Silverman’s rule of thumb [SIL 98, p.45]. To mitigate edge eﬀects, the
obtained distributions were cropped based on the minimum and maximum values found in the
data. Following cropping, they were normalized to ensure an integral of 1.

86 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The probability distributions for each specimen were then multiplied by the number of de-
tected pores in that specimen. That way, the resulting distributions can provide insights into
the number of pores rather than probabilities. For instance, as shown in Fig.3.14a, the area un-
der the distribution between 0 µm and 20 µm indicates the number of pores in the sample with
an equivalent diameter less than 20 µm. This approach enables diﬀerentiation between samples
exhibiting similar pore size probability distributions but varying pore densities.
As these distributions were calculated across numerous specimens, they were averaged to
yield a single mean distribution for each parameter (Deq , S, and distance from the surface).
To account for specimen variations, shaded areas around the distribution curves represent one
standard deviation from the mean curve.

Fig.3.14b demonstrates that for both build orientations, all pores exhibit equivalent
diameters smaller than 70 µm. The majority are notably smaller − 81% and 88% being less
than 30 µm for the Z and 45° specimens, respectively. This graph also indicates that more pores
are measured in 45° specimens, with an average increase of 29%.
In addition to their relatively compact size, pores tend to have high sphericity values,
as illustrated in Fig.3.14c. This tendency is more pronounced for Z specimens. Combining size
and sphericity data suggests that most pores are of the keyhole type, as opposed to lack of
fusion defects. Lack-of-fusion defects typically possess larger dimensions and irregular shapes,
resulting in lower sphericity values.
Fig.3.14d oﬀers insightful information about pore distribution within specimens. The ma-
jority is concentrated in a subsurface layer, with 62% and 71% being situated within
200 µm from the surface for Z and 45° specimens, respectively. The main peak in both Z and
45° distributions indicates that pores tend to form at a speciﬁc depth from the surface, ap-
proximately 140 µm. This aligns with the overlap region between contour laser tracks and bulk
material. One hypothesis to explain this could be that inappropriate laser contour scanning
parameters or overlap adjustment might be responsible for these pores. Excessive overlap could
for instance lead to locally high energy density, encouraging keyhole instabilities and pore for-
mation. It is worth noting that even considering this over-concentration of sub-surface porosity,
the specimens still show a high density (ą 99.9%).

An additional peak, close to 0 µm, is observed in the 45° distribution in Fig.3.14d. This
corresponds to pores detected in the down-skin region, just beneath the surface. XCT scan
resolution does not allow precise identification of whether these are actual internal pores. Most
are likely open pores or, more specifically, roots of thin surface valleys not accurately captured
during volume segmentation. Manual inspection of XCT scans supports this assumption for
the majority of these pores, although resolution is insufficient for some of them. Some may be
internal pores resulting from the closure of surface valleys. Notably, these pores tend to exhibit
an elongated morphology, which partly explains the higher number of pores with intermediate
sphericity values in the 45° specimens (see Fig.3.14c).

To summarize, the vast majority of pores found in 45° specimens are similar to those observed
in Z specimens in terms of size, shape, and location: most are small (ă 30 µm) and roughly
spherical pores located at the interface between contour and bulk laser scanning strategies
(« 140 µm deep).

Defects characterization using XCT | 87

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
3.4.3 Surface defects
The characterization of surface defects presents greater complexity compared to internal
defects, particularly for as-built surfaces. In many cases, defects are not isolated and well-
deﬁned objects. Instead, they often involve a combination of various irregularities on the surface,
forming for instance interconnected valleys. Nonetheless, certain typical surface features can still
be identiﬁed. To enable more quantitative comparisons, the surface roughness is then assessed
using standard roughness parameters.

3.4.3.1 Types of surface defects

Fig.3.15 and Fig.3.16 show the main surface defects identiﬁed on Z and 45° specimens. For
both types of specimens, it includes:
• Unmelted powder particles
• Spatters
• Large protrusions on vertical surfaces
• Dross in down-skin regions (see Section 1.2.2 for dross deﬁnition)
• Valleys (also referred to as notches)

Figure 3.15 | Examples of surface defects on a Z specimen. Enlarged views show SEM-
SE images of diﬀerent areas along the specimen.

88 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.16 | Examples of surface defects on a 45° specimen. Enlarged views show SEM-
SE images of diﬀerent areas along the specimen. Support structures used during manufacturing
are shown using a diﬀerent color.

Unmelted powder particles

Unmelted powder particles can be identified in various regions, with down-skin areas showing
the most prevalence, even after post-manufacturing cleaning steps.
While unmelted powders are not expected to affect the mechanical properties, they can
raise concerns about health and safety, as well as lead to other functional issues like fluid
contamination in hydraulic systems. They can also make the observation of more important
surface features difficult.
Efforts were made to further clean specimens after the removal of support structures, and
see if US bath cleaning is suited for this purpose. To do so, one Z and one 45° specimen were
subjected to XCT scanning at three stages:
1. Directly after support structures removal.
2. After 10 min ultra-sonic (US) cleaning in ethanol.
3. After an additional manual cleaning with a plastic brush and ethanol.

Defects characterization using XCT | 89

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The evolution of the surface condition is depicted in Fig.3.17a-c and Fig.3.18a-c. These
images represent the curvature (κmin ) in a 3D representation, derived from XCT scans at each
stage. The κmin representation helps in detecting powder particles, which exhibit a high positive
value (displayed in white). Images from each step were taken from the same area to facilitate
comparison, and Otsu’s threshold was used to enhance the visibility of powder particles during
segmentation.

Figure 3.17 | Surface of a Z specimen before and after several cleaning steps. The
same region is shown in three images, and roughness measurements were made using the surface
extracted from the whole XCT scan. Images are curvature 3D representations obtained from
XCT scans. The surface is segmented using Otsu’s threshold to better visualize powder particles.
Curvature and roughness were computed using the following parameters: λc “ 0.8 mm, rcurv “
20 µm and no S-ﬁlter.

The results indicate that US cleaning enables the removal of a large part of powder particles.
However, a signiﬁcant quantity of powder remains and can partially be removed by cleaning
with a plastic brush. This means that (for the experimental conditions investigated) even after
US cleaning, a substantial number of powder particles remain, although they are not ﬁrmly
attached to the surface. This can be a concern, as mentioned earlier.
From a research standpoint, this highlights the great importance of thorough surface clean-
ing. As depicted in Fig.3.18a-c, the valleys remain often concealed and hidden by particles
before cleaning. Given that the primary importance of comprehensive surface cleaning (includ-
ing brush-based methods) was realized only during the mid-phase of this PhD research, some of
the characterized specimens underwent US cleaning alone, while others also underwent manual
cleaning using a plastic brush. This can potentially contribute to dispersion in roughness mea-
surements, which show some variations depending on the cleaning step − see Fig.3.17a-c and
Fig.3.18a-c.

90 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.18 | Down-skin surface of a 45° specimen before and after several cleaning
steps. The exact same region is shown in three images. Images are curvature 3D representations
obtained from XCT scans. The surface is segmented using Otsu’s threshold to better visualize
powder particles. Curvature and roughness were computed using the following parameters:
λc “ 0.8 mm, rcurv “ 20 µm and no S-ﬁlter.

Spatters
Large particles can occasionally be observed (see e.g. Fig.3.16). Here, "large" refers to sizes
that signiﬁcantly exceed the expected maximum powder particle size of 25 µm. Although no
in-depth study has been carried out concerning their origin, they are considered to be most
likely spatters. For the sake of simplicity, they will therefore be referred to as such hereafter.
However, some may also originate from the initial powder itself.

Large protrusions
Large protrusions are frequently observed, particularly on vertical walls. In these areas, they
exhibit a distinctive morphology that is not commonly reported in the literature. Fig.3.19 illus-
trates some of these protrusions on a 3D representation of a Z specimen obtained via XCT. The
representation combines roughness and curvature data to make visualization easier. Notably,
protrusions are identiﬁed by their height, depicted in red in Fig.3.19.
These protrusions are generally quite large (around 300 to 400 µm wide) and have a curved
water-drop shape. Furthermore, their spatial distribution appears non-random, although it can
vary signiﬁcantly between specimens. For instance, Fig.3.19 shows that several protrusions are
vertically aligned. They also tend to be denser on one side of the specimen, while the other
side presents a more conventional plate-pile appearance. The most plausible explanation is that
these protrusions correspond to the initiation or termination points of the laser tracks, where
variations in energy density or melt-pool dynamics could contribute to the formation of such
defects.

Defects characterization using XCT | 91

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.19 | 3D representation of a Z specimen. Height and curvature are superim-
posed for better visualization, the curvature being shown with some transparency. Height and
curvature are measured using λc “ 0.8 mm, rcurv “ 12 µm “ 4 ˆ voxel size and λS “ 9 µm “
3 ˆ voxel size. Surface segmentation from the XCT scan was done using Otsu’s threshold.

Dross
Dross is widely visible in down-skin regions − refer to Section 1.2.2 for dross deﬁnition.
This leads to the presence of signiﬁcant protrusions and valleys in between, see Fig.3.20. The
protrusions vary in shape and size, but are frequently observed to be a few hundred µm wide.

92 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.20 | Down-skin surface of a 45° specimen. Height and curvature measurements
are superimposed for better visualization, the curvature being shown with some transparency.
Height and curvature were measured using λc “ 0.8 mm and rcurv “ 20 µm. No S-ﬁlter was
used. The TTBH was used to better reveal dross and the resulting valleys.

Valleys

Valleys can be seen everywhere on the surface but with diﬀerent morphologies depending on
their location.

Defects characterization using XCT | 93

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
In up-skin and vertically built areas, valleys primarily stem from the plate-pile effect. In this
case, they tend to be relatively evenly spaced and aligned perpendicular to the build direction.
However, the spacing between valleys does not precisely correspond to the layer height as one
might anticipate. For instance, their spacing is found equal to 75 µm in the green window in
Fig.3.15, which is almost twice the layer height (40 µm). Additionally, these valleys are often
elongated yet shallow.
Conversely, valleys in down-skin regions are related to dross formation. They are often deeper
than plate-pile-like valleys in vertical areas and more irregular. Since dross formation occurs
parallel to the build direction, resulting valleys tend to be oriented at 45° with respect to the
surface (see Fig.3.21). Furthermore, large quantities of unmelted powders attached to the surface
and filling valleys often obscure them. This complexity can make their characterization using
conventional surface assessment tools like tactile probes or interferometers more difficult. Even
XCT scans may struggle to accurately detect valleys due to the presence of powder particles
that obstruct their visibility. These issues are illustrated in Fig.3.21.

a b c

XCT (3D) Conventional surface

Real surface characterization (2.5D)
Dross

Notch root
smoothed due to
limited resolution

Notch missed All notches partly

because of missed in directional
unmelted powder 2.5D characterization

Figure 3.21 | Schematic representation of the down-skin region of a 45° specimen,

showing dross formation and resulting valleys oriented at 45°. (a) Real surface, (b)
Surface as it could be seen by XCT, and (c) Surface as seen by a conventional (2D or 2.5D
directional) surface characterization tool such as interferometry. Since valleys are inclined, a
2.5D characterization will miss a large part of them. XCT provides better results, although
unmelted powder particles can also obscure notches.

94 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To prevent unmelted powder particles from hiding notches, the use of the Trian-
gle Threshold for Bimodal Histograms (TTBH, defined in Chapter 2) to segment
specimens’ surface was found to be very effective.
In Fig.3.22a, the conventional use of Otsu’s threshold results in powder particles masking sig-
nificant portions of the surface. Conversely, Fig.3.22b demonstrates that the triangle threshold
for bimodal histograms substantially reduces this effect, uncovering other surface features like
dross and resulting valleys. This thresholding approach yields outcomes similar to those of spec-
imen cleaning using both ultrasonic bath treatment and brushing, as depicted in Fig.3.22c. The
triangle threshold for bimodal histograms even surpasses physical cleaning in its efficiency in re-
moving powder particles. Consequently, this threshold will be extensively employed throughout
this work, including for surface roughness quantification.

Defects characterization using XCT | 95

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 3.22 | Comparison of the eﬃciency of sample cleaning and the use of the
TTBH to eliminate powder particles from XCT scans. (a) Surface obtained before
any cleaning (step t0 in 3.18) using Otsu’s threshold, (b) Surface obtained before any cleaning
using TTBH and (c) Surface obtained after both US and brush cleaning, obtained using Otsu’s
threshold. Roughness and curvature were computed using λc “ 0.8 mm, rcurv “ 20 µm and no
S-ﬁlter. Roughness and κmin are superimposed in central pictures by displaying κmin with some
transparency. Pictures on the side represent κmin only.

96 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
3.4.3.2 Quantitative surface roughness
To provide a more quantitative description of the surface states of the specimens, several
roughness parameters were computed. The roughness data employed for this analysis was ob-
tained from XCT scans. The calculated parameters encompass the arithmetical mean height
(Sa ), root mean square height (Sq ), maximum height (Sz ), maximum valley depth (Sv ), height
distribution skewness (Ssk ) and kurtosis (Sku ). The mathematical definitions and practical
implementations used to compute these parameters are detailed in Chapter 2.
Similar to the porosity characterization, roughness parameters are measured for 29 Z spec-
imens and 27 45° specimens. Since the surface states significantly differ between the up-skin
and down-skin regions of the 45° specimens, the parameters are computed separately for these
areas. The definitions for up-skin and down-skin regions used in the calculations are depicted in
Fig.3.23. In the case of Z specimens, the whole surface was used. The calculation region has a
length of 9 mm along the specimen axis, accounting for a portion of the specimen’s fillet radius.

Down Up
skin skin

45°
Specimen
axis

Figure 3.23 | Deﬁnition of up-skin and down-skin regions used for roughness parameters
calculation on 45° specimens.

Results were summarized in the form of violin plots, which are enhanced versions of box
plots. The way those plots can be read is detailed in Fig.3.24.

25
Probability density
Unlikely values
distribution
Q3
20
Adjacent values
Sa (µm)

Q1
Values within 1.5 IQR
of the nearest quartile
15 Deﬁnes limits above which points
are considered to be outliers Inter-quartile range (IQR)
Space between 1st and 3rd quartiles

10 Most frequent value

Median

5
Vertical (Z) 45° / Down-skin 45° / Up-skin
Figure 3.24 | Presentation of violin plots for roughness parameters.

Defects characterization using XCT | 97

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Given the size of the observed surface features, a cut-off wavelength of 0.8 mm was estimated
to be the best compromise for roughness computation. However, it is worth noting that other
values may be relevant and lead to significant differences in roughness parameter values. As an
example, Fig.3.25a-b show violin plots for Sa parameter using two cut-off wavelengths: 0.25 mm
and 0.8 mm.
λc = 0.25 mm λc = 0.8 mm
30 30
a b
25 25

20 20

Sa (µm)
Sa (µm)

15 15

10 10

5 5

0 0
Vertical (Z) 45° / Down-skin 45° / Up-skin Vertical (Z) 45° / Down-skin 45° / Up-skin

Figure 3.25 | Violin plots for Sa parameter computed for (a) λc “ 0.25 mm and (b)
0.8 mm. In both cases, an S-ﬁlter was applied using λS “ 0.03 mm “ 10 voxel size.

Although tendencies between build orientations and down-skin/up-skin areas are the same
for both cut-off wavelengths, absolute values are significantly different. For instance, the median
Sa value in down-skin area of 45° specimens is around 12.2 µm for λc “ 0.25 mm whereas it is
only about 18.8 µm for λc “ 0.8 mm. One should therefore always be careful when comparing
roughness parameter values computed with different workflows, in particular, different cut-
off wavelengths. In the present case, all characterizations done hereafter use the same cut-
off wavelength to overcome this issue. In all cases, an S-filter is also applied after roughness
computation, using a cut-off λS “ 0.03 mm.
Violin plots for Sa , Sq , Sz , Sv , Ssk and Sku parameters are shown in Fig.3.26a-f. Mean values
for these parameters are summarized in Tab.3.4.

Vertical 45° Down-skin 45° Up-skin

Sa(µm) 7.3 ˘ 0.5 19.3 ˘ 2.6 6.4 ˘ 0.7
Sq(µm) 10.3 ˘ 0.7 26.1 ˘ 3.7 8.6 ˘ 1.0
Sz(µm) 150 ˘ 25 269 ˘ 51 120 ˘ 21
Sv(µm) 64 ˘ 15 117 ˘ 30 59 ˘ 17
Ssk 1.0 ˘ 0.3 0.8 ˘ 0.2 0.2 ˘ 0.5
Sku 4.2 ˘ 1.6 2.5 ˘ 1.0 3.6 ˘ 3.0

Table 3.4 | Roughness parameters, computed for Z specimens (29 sp.) as well as
down-skin and up-skin regions of 45° specimens (27 sp.). Roughness was computed
from XCT data using the 3D characterization workflow introduced in Chapter 2. The cut-off
values for filters were set to λc “ 0.8 mm and λS “ 0.03 mm. Values presented in the format
(mean ˘ std).

98 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

30
20
25

Sq (µm)
Sa (µm)

15 20

15
10
10
5
Vertical (Z) 45° / Down-skin 45° / Up-skin Vertical (Z) 45° / Down-skin 45° / Up-skin

400
c Sz 180 d
Sv
350 160

300 140

120
Sv (µm)
Sz (µm)

250
100
200
80
150 60

100 40

Vertical (Z) 45° / Down-skin 45° / Up-skin Vertical (Z) 45° / Down-skin 45° / Up-skin

2.0
e Ssk 16 f
Sku
1.5 14

12
1.0
10
Sku
Ssk

0.5 8

6
0.0
4
0.5 2

0
Vertical (Z) 45° / Down-skin 45° / Up-skin Vertical (Z) 45° / Down-skin 45° / Up-skin

Figure 3.26 | Violin plots for roughness parameters (a) Sa , (b) Sq , (c) Sz , (d) Sv (e) Ssk
and (f) Sku . Roughness is computed using λc “ 0.8 mm and λS “ 0.03 mm. Violin plots were
computed using 29 Z and 27 45° specimens.

Defects characterization using XCT | 99

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.3.25a and Fig.3.26a show that, on average, down-skin surfaces of 45° specimens are
signiﬁcantly rougher compared to up-skin and vertical surfaces. Vertical specimens are also
slightly rougher than up-skin areas of 45° specimens. All these observations are consistent with
qualitative observations made on SEM images in Fig.3.15 and Fig.3.16.
Similar conclusions can be drawn from Sz and Sv parameters in Fig.3.26b-c, even though
discrepancies between diﬀerent conditions are smaller. This could be explained by the fact that
even though vertical surfaces and up-skin regions are much smoother on average, they may still
contain some relatively large defects. This could also explain why Sku can reach higher values
for those two conditions − see Fig.3.26f. However, these observations may also come from the
fact that the area characterized is eight times smaller for up-skin and down-skin regions of 45°
specimens than for Z specimens.
Regarding the skewness measurements presented in Fig.3.26e, most surfaces tend to be peak-
dominated (Ssk ą 0). However, some of the 45° specimens display negative skewness in up-skin
regions. This positive skewness is attributed to the occurrence of protrusions in vertical regions
and to dross in down-skin areas. In contrast, up-skin regions usually lack such protrusions but
often feature plate-like valleys. This might account for the instances of negative skewness. Yet,
spatters appear sometimes in these areas, possibly explaining the exceptionally high skewness
(and kurtosis) observed in a few specimens.

100 | Chapter 3 – Materials characterization

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary
Ti64 samples were manufactured by L-PBF and submitted to a stress-relief at 720 °C for
2 h to completely alleviate residual stresses. The surfaces were maintained in their as-built
state to study surface defects. To investigate the inﬂuence of build orientation, fatigue
specimens were produced in two orientations (Z and 45°), while tensile specimens were
prepared in three orientations (vertical (Z), 45°, and horizontal (XZ)).

The material exhibited a microstructure and tensile properties consistent with the existing
literature. The microstructure consisted of fine α laths (averaging around 0.5 µm thickness)
along with a minor fraction of β phase (2.2 ˘ 0.3%). Prior β0 grains were reconstructed
from EBSD data and showed a typical columnar morphology with the growth direction
parallel to the build direction. They showed a significantly different morphology near the
surface, which can probably be explained by differences in local energy density and thermal
gradients during printing. Everywhere, though, the final microstructure is primarily
composed of α laths and exhibits only little texturing. For more detailed information
regarding the material microstructure, readers are encouraged to refer to [GAI 23a] and
[GAI 23b].

Compared to conventional wrought Ti64 [NGU 22], the tensile strength was relatively
high (Y¯S “ 1098 MPa for Z specimens), but the elongation to failure was rather limited
(ϵ¯f “ 7.8% for Z specimens). This can be attributed partly to the aforementioned ﬁne
microstructure. XZ specimens displayed notably low ductility (ϵ¯f “ 0.8%), possibly due to
the presence of signiﬁcant defects left after support removal. Additionally, prior β0 grain
boundaries, aligned perpendicular to the loading direction for these specimens, seemed to
serve as weak points contributing to premature failure.

XCT-based characterization revealed very low porosity (< 0.002%) with predominantly
near-spherical and small-sized pores (over 80% smaller than 30 µm). However, the majority
of these pores were located beneath the surface (« 140 µm deep), likely originating from
contour scans.

Surface roughness was also evaluated via XCT, utilizing the 3D characterization approach
introduced in Chapter 2. The surface roughness was found to be relatively low in com-
parison to literature values, particularly for up-skin and vertical regions. For instance, the
mean Sa value across all Z specimens was 7.3 µm. In contrast, down-skin regions exhibited
higher roughness with a mean Sa of 19.3 µm. Among the observed surface defects with the
potential to inﬂuence fatigue properties, valleys were prominent. These valleys originated
from the plate-pile eﬀect in up-skin and vertical surfaces, and from dross formation in down-
skin regions. These valleys could exhibit intricate morphologies with re-entrant features,
frequently concealed by unmelted powder particles. Conventional 2.5D surface characteri-
zation tools may encounter limitations in properly assessing such surfaces, making XCT a
suitable solution for this purpose.

Defects characterization using XCT | 101

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Fatigue properties and crack

initiation mechanisms of as-built and
post-treated specimens

Contents
4.1 Fatigue properties and crack initiation mechanisms for as-built specimens104
4.1.1 Protocol for fatigue tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.2 Results of fatigue tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.3 Killer defects identiﬁcation: the need for a 3D characterization . . . . . . . 111
4.2 Alpha-case formation and surface embrittlement induced by heat treat-
ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.1 Alpha-case formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.2 Impact on tensile properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.3 Impact on fatigue properties . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.3 Improvement of fatigue properties by Plasma electrolytic Polishing (PeP)131
4.3.1 Principle of PeP surface treatment . . . . . . . . . . . . . . . . . . . . . . . 131
4.3.2 Conditions of PeP treatments applied to fatigue specimens . . . . . . . . . 133
4.3.3 Roughness of specimens polished by PeP . . . . . . . . . . . . . . . . . . . 138
4.3.4 Fatigue resistance improvement and change in crack initiation mechanism . 146
Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

| 103

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
4.1 Fatigue properties and crack initiation mechanisms for as-built
specimens
4.1.1 Protocol for fatigue tests
Fatigue tests were performed under laboratory conditions (room temperature and pressure)
using servo-hydraulic machines equipped with 100 kN load cells. These tests were carried out
at constant stress amplitude using a sinusoidal waveform and at a frequency f of 10 Hz. The
stress ratio was set at R “ 0.1. Tests that exhibited no failure after 107 cycles were stopped and
categorized as run-outs.

Specimens were first tested with their as-built surfaces. A stress-relief heat treatment of
2 h at 720 °C was applied to these specimens, ensuring they were free of residual stresses. In
this condition, 22 Z and 18 45° specimens underwent testing. Among these, 10 Z and 10 45°
specimens were XCT scanned both before and after failure for in-depth analysis.
2
The stress levels were calculated based on a theoretical section of S “ π p3 mmq
4 « 7.07 mm2 .
Yet, minor geometrical deviations from the CAD definition meant that the effective specimen
gauge diameter was slightly reduced. Thus, the mean section of the gauge area for all XCT
scanned specimens was determined, and the average and standard deviation were calculated
for each build orientation. The obtained values are 6.91 ˘ 0.07 mm2 for Z specimens and
6.99 ˘ 0.32 mm2 for 45° specimens. They were used for correcting stress levels in Wöhler
curves. However, for the sake of simplicity, the same rounded theoretical stress levels − i.e.
those obtained using the theoretical section − will be used in the discussions for both Z and 45°
specimens.

In addition to as-built specimens, four machined and manually polished (M&P) vertical
specimens were tested as a reference. These specimens were initially machined from rectangular
bars through turning and then manually polished with SiC P1200 paper. Thus, they were free of
both the as-built surface roughness and the high concentration of subsurface pores highlighted
in Section 3.4.2. Following their preparation, a supplementary stress-relief heat treatment of
2 h at 660 °C was performed. This ensured the elimination of any residual stresses potentially
arising from machining and polishing while keeping the microstructure almost unchanged.

4.1.2 Results of fatigue tests

The Wöhler curves obtained for the diﬀerent conditions are shown in Fig.4.1. These curves
will be analyzed in light of fracture surface observations and previous defect characterizations.
Wöhler curves indicate that fatigue strengths of as-built specimens are consistent
with those found in the literature. See Tab.1.1 presented in the state of the art for a
comparison. For both Z and 45° specimens, the fatigue strength at 107 cycles is estimated to be
around 250 MPa.

104 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1100
R = 0.1
1000

900

800
σmax (MPa)

Vertical | as-built
700 45° | as-built
600 Vertical | M&P
Number of
2 4 3 superimposed points
500

400 32 4

300 2
2
200
104 105 106 107
Nf
Figure 4.1 | Wöhler curves of as-built specimens. M&P specimens are also represented
as a reference. They were submitted to a stress-relief treatment (2 h at 660 °C) after machining
and polishing. Fatigue tests are performed at R “ 0.1.

Fig.4.2a-b show fracture surfaces for as-built Z and 45° specimens. The relatively flat region
in the upper parts of Fig.4.2a-b show unambiguously that in both cases, cracks initiated from
the surface. More generally, it has been observed that crack initiation always occurred at
the surface for as-built specimens. It is worth noting that no single defect can be identified
in Fig.4.2a-b, and that crack initiation seems to have occurred in vast zones. This does not
necessarily mean that multiple cracks initiated at different locations since this usually leads to
the formation of more ridges than what can be seen in Fig.4.2a-b. The mechanism behind this
particular fracture surface morphology is discussed in Section 4.1.3.2.
Fig.4.1 shows that M&P specimens exhibit significantly higher fatigue strength
than as-built specimens. Despite the low number of tested specimens, fatigue strength at
106 cycles of M&P specimens can be estimated at about 850 MPa. By looking at M&P fracture
surfaces in Fig.4.3a-b, this huge difference can be explained by a change in the crack initiation
mechanism. Indeed, all crack initiations occurred on an internal pore in the case of M&P
specimens. The pore was deep inside the material (ă 290 µm below the surface) in 3 cases
over 4, whereas it was directly beneath the surface in one case. Fig.4.2a-b show two surface
crack initiations on as-built specimens and Fig.4.3a-b show two crack initiations on bulk pores
of different sizes in polished ones.

Fatigue properties and crack initiation mechanisms for as-built specimens | 105

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.2 | SEM-BSE topographical mode observations showing crack initiation at
as-built specimens surface. (a) Fracture surface of a Z specimen tested at 300 MPa and that
failed after 73 547 cycles cycles. (b) Fracture surface of a 45° specimen tested at 300 MPa and
that failed after 193 381 cycles.

These results conﬁrm that in our case and as it was reported several times in the literature,
surface defects of as-built samples are much more critical than internal pores. Only
when surface defects are removed by machining/polishing, do pores become potential crack
initiation sites. Note that for as-built specimens, the high density of sub-surface pores described
in Section 3.4.2 seem to have negligible impact on the fatigue resistance compared to surface
defects.
Concerning the M&P specimens, it is also interesting to note that their fatigue strength
seems to be in the upper range of what can be found in the literature for L-PBF Ti64 (see again
Tab.1.1 in the state of the art for details). This can be primarily attributed to the good bulk
material integrity, which, as shown in Section 3.4.2, contains only a few pores of limited size
and rather high sphericity. This can be conﬁrmed by the fracture surfaces in Fig.4.3a-b of both
polished specimens that were tested at 950 MPa. In those two cases, crack initiation was caused
by small rounded pores. This is especially true in Fig.4.3b, where the killer defect is a perfectly
spherical pore of only 11 µm.

106 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.3 | SEM observations of two polished specimens fracture surfaces showing
crack initiation at small rounded pores. Large views are obtained in BSE topographical
mode, whereas higher resolution images are obtained in SE mode.

The good properties of polished specimens still need to be qualified because of the reduced
number of specimens and the large scatter observed. The two specimens tested at 950 MPa reach
for instance very different fatigue lives: one breaks at 21 441 cycles while the other lasts almost
100 times longer (1 714 472 cycles). This can be attributed to significant differences between
the defect populations of both specimens. Thus, the equivalent diameter can be estimated to
be 52 µm in Fig.4.3a, much larger than the 11 µm found in Fig.4.3b. Furthermore, the pore in
Fig.4.3a is much less spherical than the one in Fig.4.3b. A Hot Isostatic Pressing treatment
(HIP) could in this case be relevant to remove porosity and therefore achieve more reproducible
fatigue properties.
However, a HIP treatment would significantly affect the microstructure. This could have ad-
verse effects, since the fine microstructure obtained after low-temperature heat treatment (2 h at
720 °C) could also contribute to the good observed properties. However, it seems difficult to draw
a conclusion on this question since contradictory studies exist to explain which microstructure
should be the most efficient regarding fatigue properties [POL 14]. The lamellar microstructure
obtained with L-PBF could be rather advantageous regarding crack propagation, as cracks could
tend to deviate from their trajectory during propagation to follow α-laths interfaces [GAL 17].
All these bifurcations imply that more energy is needed to propagate cracks. This hypothesis
is coherent with the high density of micro-cracks found in the propagation zone of the fracture
surface and which most likely follow α-laths interfaces, see e.g. Fig.4.4. Some studies even state
that finer α-laths will result in better fatigue strength [LIU 19, LIU 17]. As a reminder, the
α-laths width in the present study is particularly small (« 0.5 µm, see [GAI 23b]). However,
other studies suggest conversely that better fatigue strengths can be reached with coarser laths
[GAL 17, WAN 23] and further work would thus be needed to clarify the matter.

Fatigue properties and crack initiation mechanisms for as-built specimens | 107

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.4 | Crack bifurcations visible on the propagation zone of a fracture sur-
face. The specimen is a vertical one, tested at 350 MPa and which failed after 241 406 cycles.
Observations are made using an SEM in SE mode.

Regarding 45° specimens, it is interesting to note that all crack initiations occurred
in the down-skin region, as has already been found in the literature [PEG 18]. Again,
this is coherent with roughness measurements in Section 3.4.3.2 showing that the down-skin
region is much rougher than the up-skin one. This probably explains why, as can be noticed
in Fig.4.1, the fatigue life of the 45° specimens is on average slightly lower than that of the Z
specimens, especially at lower stresses. For instance, at 350 MPa, median cycles to failure for Z
and 45° specimens are 255 152 cycles and 97 800 cycles, respectively. Nevertheless, the distinction
between Z and 45° orientations is minimal compared to the larger diﬀerence between as-built
and M&P specimens.

108 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Looking in more detail into the position of initiation sites in 45° specimens, it can be noticed
that failure occurred most frequently in the lower section, closer to the build plate during
fabrication (see Fig.4.5a). This observation is not reﬂected in Z specimens. Fig.4.5b highlights
this trend, plotting the initiation sites position along the specimen axis in 10 Z and 10 45°
specimens. A position of 0 means that the initiation site is located in the middle of the gauge
length.

-4 -2 0 2 4 Position along
0 specimen axis (mm)

1
Build plate

b
Z (10)
0.4 45° (10)
Probability density

0.3
Gauge
length
0.2

0.1

0.0
4 3 2 1 0 1 2 3 4
Build plate
Position along specimen axis (mm)
Figure 4.5 | Initiation sites positions along specimen axis for Z and 45° specimens.
Each point corresponds to a specimen. 10 Z and 10 45° were used for this study. For each
specimen, the initiation site has been identiﬁed on the corresponding XCT scan following the
methodology described in Section 4.1.3. The position of the initiation site was then estimated.
A position of 0 means that the initiation site is located in the middle of the gauge length.
Probability density curves were computed by kernel density estimation.

Fatigue properties and crack initiation mechanisms for as-built specimens | 109

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Calculations were done to determine if the preferential crack initiation in lower sections of
45° can be explained by roughness inhomogeneity. More precisely, roughness parameter proﬁles
were computed along the specimens’ axis. To do so, for a given specimen and for 50 evenly
spaced positions along its axis, points within a 1 mm high transverse segment were selected
(see Fig.4.6a). Then, Sa and Sv roughness parameters were calculated using the local height
measurements performed at these points from XCT data using a cutoﬀ λc = 0.8 mm.
Roughness measurement area
= 1 mm high transverse segment
centered on the measurement point

-4 -2 0 2 4 Position along
a 0 specimen axis (mm)

b c
Sa (µm) Sv (µm) Sa (µm) Sv (µm)
140 140
30 Z Sa 30 45° | downskin
Sv 120 Sa 120
Initiation site position
25 25 Sv
Initiation site position
100 100
20 20
80 80

15 15
60 60

10 10
40 40

4 2 0 2 4 4 2 0 2 4
Position along specimen axis (mm) Position along specimen axis (mm)

Figure 4.6 | Roughness profiles along specimens’ axis. (a) Scheme showing the area used
to measure roughness parameters at each position along a specimen axis. (b) Sa and Sv profiles
for Z specimens. (b) Sa and Sv profiles for down-skin regions of 45° specimens. In (b) and (c),
gray points indicate the position of initiation sites. Roughness parameters were measured using
only specimens where the initiation site has been identified (10 Z and 10 45° specimens). Curves
represent the average roughness profile (among specimens of the same build orientation) whereas
shaded areas correspond to a distance of one standard deviation from the average curves.

Refer to Fig.4.6b for roughness parameters proﬁle visuals for Z specimens, and Fig.4.6c for
those of 45° specimens. Due to the consistent failure in the down-skin region of 45° specimens,
only points from this region were used in roughness computations. For a clearer deﬁnition of
this down-skin region, see Fig.3.23.
As can be seen in Fig.4.6b, Sa and Sv roughness parameters show only little variations along
the specimen axis of Z specimens. This is consistent with the uniform distribution of initiation
sites marked by gray spots.

110 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Conversely, Fig.4.6c reveals an upward trend in both Sa and Sv values along the axis in the
down-skin region of 45° specimens. This is probably due to the fact that because of the presence
of the fillet radius, the upper part of 45° specimens is nearly horizontal during manufacturing.
Conversely, lower parts are near vertical, which leads to smoother surfaces.
Surprisingly, initiation sites predominantly cluster in lower parts of 45° specimens,
i.e. where Sa and Sv values are lowest. This tends to indicate that surface roughness might
not be the sole determinant of initiation site position. Another influencing factor might for
instance be the presence of support structures under the upper sections of 45° specimens during
manufacturing (see Fig.3.5 for an image of support structures). Lower parts of the specimens
where failure occurs did not require the use of support structures and were therefore built
directly on the powder bed. Such factors might lead to a slightly different microstructure (due
to differences in thermal history) or minor geometric shifts (due to residual stresses), potentially
influencing fatigue crack initiation.
In any case, these results indicate that roughness parameters (or at least the sim-
ple Sa and Sv parameters) are not necessarily reliable indicators for the fatigue
performance of L-PBF as-built specimens.

4.1.3 Killer defects identiﬁcation: the need for a 3D characterization

Fracture surfaces like the one in Fig.4.2a-b demonstrate that crack initiates from the surface
on as-built specimens. However, no defect could be identiﬁed and therefore the question remains:
which precise defects are responsible for failure?

4.1.3.1 The only defects identiﬁable using SEM: spatters

Actually, SEM observations enabled the identification of the killer defect in only two spec-
imens in total, one for each build orientation. In both cases, the crack initiated from a large
particle that is considered to be a spatter, see Fig.4.7a-d.
Both fracture surfaces are still relatively different and it is interesting to note that in Fig.4.7b,
the crack did not seem to initiate from the spatter only. River lines that emanate from the surface
suggest indeed that the crack could have initiated from a more extended defect, similar to what
can be seen in Fig.4.2a-b. Hence, SEM provides limited clarity on the exact defect causing the
failure.
For all other specimens, SEM observations are even less informative, see e.g. Fig.4.2a-b.
The crack seems in this case to have started from an extended circular segment rather than
from one specific point. With no distinct defect identifiable, a new strategy is essential for
identification. Incidentally, this also means that it probably is not possible here to quantify
?
defects’ severity using the conventional area parameter. It seems for instance impossible in
Fig.4.2a and hazardous in Fig.4.2b. A different approach will therefore be tempted in Chapter
5.

Fatigue properties and crack initiation mechanisms for as-built specimens | 111

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.7 | SEM fracture surface observations showing crack initiation from a
spatter. (a) Observation of a Z specimen in BSE topological mode. The specimen was tested
at 300 MPa and failed after 7 154 298 cycles. (b) Observation of a 45° specimen in SE mode. The
specimen was tested at 300 MPa and failed after 125 534 cycles.

4.1.3.2 A 3D workﬂow for a more extensive initiation site identiﬁcation

To go further in the initiation site identification, XCT was used to get a broader 3D point of
view. To this end, 10 Z and 10 45° fatigue specimens were scanned before and after failure. The
workflow for initiation site identification is described in Fig.4.8a-d. The specimen taken as an
example is the one that is shown in Fig.4.7b. This allows to check whether river lines emerging
from the surface away from the spatter could indicate crack initiation from another defect.

112 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.8 | Workﬂow for crack initiation site identiﬁcation using both SEM and
XCT scans (before and after failure). The specimen used as an example was built at 45°
and failed after 125 534 cycles at 300 MPa. (a) SEM-SE fracture surface observation. (b) Surface
mesh of the fracture surface obtained from XCT scan. (c) Crack initiation site observed on the
fracture surface from the side of the specimen. Gray levels correspond to κmin , estimated using
rcurv “ 20 µm. (d) Same area as in (c), trace backed on initial XCT scan before failure.

Fatigue properties and crack initiation mechanisms for as-built specimens | 113

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The fracture surface is first observed by SEM due to its superior resolution, see Fig.4.8a.
Even using SEM, initiation site identification can be complex in some cases since initiation was
not caused by a clearly delineated defect. To help the identification in such cases, SEM-BSE
topographical mode (instead of the conventional SEM-SE mode) was sometimes found very
useful. It can help better visualize river marks and trace back the initiation sites at their origin
more easily. An example of a BSE topographical image can be seen in Fig.4.7a. In the particular
case of Fig.4.8a, only the SEM-SE observation is shown since it is enough to identify both the
spatter and the river lines that emerge from the surface a little further.

Once the initiation site is identiﬁed on the SEM image, the same fracture surface is observed
in 3D using XCT. For this purpose, the surface is extracted from the XCT scan containing the
fracture surface. The mesh extracted from the XCT scan using the conventional marching cubes
algorithm is for instance shown in Fig.4.8b. By comparing Fig.4.8a and Fig.4.8b, the initiation
site is identiﬁed on the XCT fracture surface.

Finally, the initial XCT scan before the fatigue test is compared to the XCT fracture sur-
face to identify what surface feature caused crack initiation. This could be done using mesh
representations of surfaces before and after failure. But while mesh representations give a good
perception of volume, they might not always eﬀectively render the most relevant surface fea-
tures. On the opposite, it has been shown in Section 2.2.3 that the minimum curvature greatly
highlights surface valleys, which, as will be shown, are the surface features of primary interest.
Therefore, for the killer defect identiﬁcation surfaces were rather visualized by using a point
cloud of the surface colored using a minimum curvature measurement, see Fig.4.8c-d.

The spatter that was previously seen on the SEM observation can be seen in Fig.4.8b-d
obtained from XCT data as well. The crack path seems to follow a surface valley over a
long distance, despite the fact that this valley is not aligned on a horizontal plane.
This is highlighted by the magnifying windows of Fig.4.8c-d. It is precisely the same area that
was identified previously as a potential initiation site. Although such a valley can be identified
in the 3D representation (especially when they are underlined by the 3D rendering of curvature),
it cannot be identified using SEM only, see Fig.4.8a. This confirms the relevance of such a
3D characterization method for crack initiation site identification on L-PBF as-built
surfaces.

The precise alignment of the crack with valleys indicates their contribution to fatigue failure.
Parallel river lines in SEM images, rather than emerging solely from the spatter, suggest crack
propagation from the entire area, including the spatter and valleys. Two scenarios could be
considered:

1. Valleys and the spatter both served as crack initiation points, leading to simultaneous crack
initiation over a vast perimeter. Although the valleys may appear to be less pronounced
defects than the spatter, their extent could allow many cracks to initiate along the surface
before coalescing, causing rapid damage and short fatigue life, as reported by [PEG 18] in
L-PBF Ti64.

2. The crack primarily began at the spatter, but valleys accelerated subsequent crack prop-
agation due to the generated stress concentration.

114 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To confirm these theories, advanced techniques, like in-situ crack propagation in synchrotron
XCT, would have been necessary. As of now, no definitive answer can be provided.
Nonetheless, the 3D initiation site characterization in Fig.4.8a-d offers valuable insights.
Fig.4.9a magnifies an area where a small detail emerges: a profound, narrow pit-like defect
likely present before fatigue cycling. This defect seemingly led to a crack that merged with the
primary one, implying it might have contributed to crack initiation.
Conventional pre-failure surface characterization (e.g. using a profilometer or an interfer-
ometer) could easily overlook such defects. The XCT characterization, however, accounts for
them. As can be seen in Fig.4.9e, the valley located behind the spherical spatter is indeed clearly
visible. Furthermore, Fig.4.9d-e show the pit-like defect is visible as well. To see it, one has to
look to the surface from the inside, i.e. as if one would stand in the core of the material. This
point of view is made possible by the XCT characterization and surface extraction. This show-
cases the capability of 3D XCT in delving deeper into crack initiation mechanisms, especially in
potentially uncovering obscured features commonly found in as-built surfaces.
Using the presented workflow, initiation sites were identified for 10 Z and 10 45° specimens.
Apart from the two crack initiations caused by spatters − i.e. in most cases, failure was
caused by surface valleys.
While both build orientations exhibited valleys, their morphologies varied. The morphology
of valleys in the 45° specimens is showcased in Fig.4.8d. The focus here, and in subsequent
discussions, is on valleys in the down-skin area, as failures in all specimens occurred in this
region. Compared to Z specimens, these valleys appear deeper and more irregular.
The Z specimens’ valleys, illustrated earlier in Fig.3.19, are generally shallower, horizontal,
and regular unless interrupted by prominent protrusions. Fig.4.10a-d depicts the identification
of a plate-pile valley as the crack initiation site on a Z specimen. It can be noted that for
the specimen chosen as an example, two initiation sites can be identified. This is a frequently
encountered case for both Z and 45° specimens, especially at high stresses.

Fatigue properties and crack initiation mechanisms for as-built specimens | 115

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.9 | Revealing of deep notches hidden by a spatter using the 3D XCT
characterization. The sample characterized is the same 45° specimen as in Fig.4.8. (a) SEM-
SE observation of the crack initiation zone, where a spatter can be identiﬁed. This spatter hides
a deep valley and an even deeper and sharp pit-like defect. (b-e) 3D height maps showing (b) a
global view, (c) a magnifying window centered on the initiation site, (d) the crack initiation area
observed from the inside, revealing the presence of a deep pit-like defect and (e) a longitudinal
cut showing the presence of a valley behind the spatter and the deep pit shown in (d).

116 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.10 | Workflow for crack initiation site identification using both SEM and
XCT scans (before and after failure). The specimen used as an example is a vertically built
one which failed after 129 339 cycles at 350 MPa. (a) SEM-BSE (topographical mode) fracture
surface observation. (b) Fracture surface obtained from XCT scan. Colors correspond to κmin
measurements performed using rcurv “ 20 µm and λS “ 10 voxel size. (c) Crack initiation site
observed on the fracture surface from the side of the specimen. (d) Same area as in sub-figure
c, identified on the initial XCT scan before fatigue failure. A valley is identified to be the cause
of crack initiation by comparison with (c).

Fatigue properties and crack initiation mechanisms for as-built specimens | 117

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Further analysis revealed more crack initiations in the specimen, not visible in Fig.4.10a-d.
These cracks, termed secondary cracks, emerged in the specimen’s lower section but did not
result in failure. With the provided 3D characterization workflow, one can also identify these
secondary cracks if their opening is large enough to be detected using XCT, as displayed in
Fig.4.11a-c.
Fig.4.11a shows a view of the crack from the inside (in black), called internal view. Fig.4.11b
shows the same area but from the outside. This is the "natural" point of view, which would for
instance be possible using SEM. Secondary cracks are highlighted by white arrows. A comparison
with the surface of the specimen before fatigue testing in Fig.4.11c highlights that these cracks
generally follow the surface valleys, reinforcing the observations from the 45° specimen in
Fig.4.8.
All these results obtained from as-built specimens show that surface defects, and primar-
ily surface valleys, control fatigue properties. The impact of surface valleys can probably
be explained by the so-called notch effect, which results in stress concentrations at valley roots.
The extent of this notch effect is known to depend on mechanical properties such as ultimate
tensile strength and/or ductility [SCH 09, Ch.7].
Titanium alloys are very sensitive to oxygen content, which has a strong impact on these
properties [OGD 55]. Surface oxygen enrichment might happen during heat treatments such as
the one used to relieve residual stresses after the L-PBF process. Thus, it may also affect the
extent of the notch effect and its consequences on the fatigue properties. This issue is addressed
in the following section.

118 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.11 | Identification of a secondary crack − i.e. a crack that didn’t propagate
until failure − in a lower section of the fracture specimen. The fatigue specimen in
question is the same as in Fig.4.10. Grey levels in all sub-figures correspond to a minimum
curvature measurement. (a) Fractured specimen observed from the inside to reveal the partially
propagated crack in black. This internal view is possible thanks to the fact that the surface was
extracted from XCT data. (b) Fractured specimen observed from the outside, in the same area
as in sub-figure a. The crack path is highlighted by white arrows. (c) Same area as in sub-figure
b, but before fatigue testing. It can be seen that the secondary crack highlighted in (a) and (b)
follows valleys that were initially present.

Fatigue properties and crack initiation mechanisms for as-built specimens | 119

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
4.2 Alpha-case formation and surface embrittlement induced by
heat treatments
This section focuses on the effects of the oxygen-enriched surface layer formed dur-
ing heat treatment, commonly called α-case, on fatigue properties1 . First, the discussion
will present the main findings obtained by Quentin Gaillard about α-case formation and its in-
fluence on tensile properties. For a more in-depth exploration, readers are directed to references
[GAI 23b, GAI 23a]. Subsequently, findings from the joint research about its implications on
fatigue properties will be presented. As for Section 3.2, the results obtained by Quentin Gaillard
can be extended to the material studied in the present PhD work, since he used samples manu-
factured with the same L-PBF machine, process parameters and Ti64 powders. Industrial heat
treatments were also performed by the same company (THERMI-GARONNE), and laboratory
heat treatments were performed using the same furnace. In both cases, the same thermal cycles
were used.

4.2.1 Alpha-case formation

Two types of heat treatments using different furnaces (an industrial and a laboratory furnace)
have been studied by Quentin Gaillard during this PhD [GAI 23b, GAI 23a]. The industrial
furnace could reach a nominal vacuum better than 10−4 mbar. Yet, instances of α-case formation
were observed during the heat treatments in this furnace. Conversely, with the same thermal
cycle, the laboratory furnace did not lead to any α-case formation. Such contrasts facilitated a
comparative analysis to distinguish between the influences of microstructure changes and α-case
emergence, both of which are tied to the selected heat treatment temperature.
Temperatures employed for heat treatments were: 720 °C, 800 °C, 860 °C, 920 °C, and 980 °C,
each sustained for 2 h. Additionally, a reference condition was evaluated. It corresponds to
samples that did not undergo any heat treatment after printing.
To gauge the depth of the α-case, polished sections of samples were chemically attacked
using a Weck reagent, composed of ammonium bifluoride dissolved in ethanol and water. This
procedure renders the α-case layer white on polished sections observed with optical microscopy,
as illustrated for diverse heat treatments in Fig.4.12a.
No α-case can be seen before any heat treatment, but it can be seen after all heat treatments
in the industrial furnace. Its depth ranges from 11 µm for the 720 °C heat treatment to 185 µm
for the 980 °C heat treatment. Interestingly, even at the highest temperature of 980 °C, the
laboratory furnace treatment did not yield any α-case formation.

4.2.2 Impact on tensile properties

Fig.4.13a-b show the evolution of the yield stress σy , the ultimate tensile strength σu and
the elongation to failure ϵf after the diﬀerent heat treatments.
The main observation that will be discussed here is the evolution of the elongation to failure
in Fig.4.13b. It can be seen that it increases steadily as the heat treatment temperature increases
when the heat treatment is performed in the laboratory furnace, without the formation of α-case.
1
Collaborative study conducted with Quentin Gaillard, PhD student, and Amélie Larguier, ﬁnal year student
at INSA Lyon

120 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.12 | Optical microscope micrographs revealing α-case for diﬀerent heat
treatment conditions , from [GAI 23b, Fig.19]. Micrographs were taken at the edge of cross-
sections in the horizontal XY plane of (a) cubes before and after various heat treatments in
the industrial furnace, and (b) half tensile specimens before and after heat treatments in the
laboratory furnace. The etched material is obtained with a Weck reagent.

When performed in the industrial furnace, the heat treatment leads to comparable results
for temperatures lower or equal to 800 °C. Yet, at 860 °C and beyond, there is a reduction
in elongation to failure, contrasting with the previous increase. For instance, at 980 °C, the
elongation to failure after the industrial treatment is measured at 4.0 ˘ 0.8%, while a value of
18.7 ˘ 0.7% is obtained after the laboratory treatment. This significant drop in elongation
to failure stems from the growth of a pronounced α-case layer, resulting in surface
brittleness.
Evidence of this surface fragility is the profusion of cracks observable on the surfaces of tensile
specimens post-failure, as illustrated in Fig.4.14a.
These cracks remain confined to the brittle α-case layer. They likely serve to release the
tensile stresses, given the α-case layer’s inability to accommodate deformation due to its limited
ductility. The propensity of α-case to initiate cracks prompts questions about its
potential influence on fatigue damage. Consequently, this led to further research into the
effects of α-case on fatigue properties.

Alpha-case formation and surface embrittlement induced by heat treatments | 121

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.13 | Evolution of tensile properties of the vertically built samples in the
as-built condition and subjected to various heat treatments in both the industrial
and laboratory furnaces : (a) Yield stress σy and ultimate tensile strength σu , (b) Elongation
to failure ϵf . Drawn from [GAI 23b, Fig.19]

Figure 4.14 | Surface cracks formed in α-case layer during tensile testing. (a) Visual-
ization of cracks on the surface of a broken ﬂat tensile specimen subjected to the industrial heat
treatment at 920 °C for 2 h. Visualization of cracks on a cross-section of the same specimen (b)
before and (c) after a chemical attack with Weck reagent. (d) Visualization of blunt cracks at
the α-case boundary. Drawn from [GAI 23b, Fig.13]

122 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
4.2.3 Impact on fatigue properties
4.2.3.1 Heat treatment conditions and resulting α-case and microstructure
Only Z specimens were used for this study, with the same geometry (3 mm diameter cylin-
drical gauge area) and L-PBF manufacturing conditions as the ones detailed in Section 3.1.
Three heat treatment conditions are compared in this study. They consist of a stress relief
possibly followed by a higher-temperature heat treatment:
• The stress relieved (SR) condition, which consists of a 720 °C|2 h treatment performed
in the industrial furnace used in Quentin Gaillard’s study. Note that it corresponds to
the state called as-built elsewhere in this manuscript (same thermal cycle and furnace).
It is a convenient denomination in the rest of the manuscript which focuses on the study
of surface topography, since it refers to the fact that the surface topography is kept in its
as-built state. However, this denomination would not be appropriate for this study which
focuses on heat treatments and their impact on α-case formation. Furthermore, the choice
has been made to call this condition SR in the present section to prevent confusion with
the work of Quentin Gaillard, where the as-built state refers to the condition without any
heat treatment.
• The HT condition refers to samples that received a 860 °C|4 h heat treatment in the
laboratory furnace after the stress-relief treatment in the industrial furnace.
• The HT|α-case condition pertains to samples treated at 860 °C for 4 h in the industrial
furnace following stress-relief.

The α-case layers observed for each condition are presented in Fig.4.15a-c.

Figure 4.15 | α-case characterization in SR, HT and HT|α-case conditions. The optical
micrographs were obtained from longitudinal cross-sections of fatigue specimens subjected to a
chemical attack with Weck reagent. The Weck reagent reveals the α-case layer which appears
in white. (Courtesy of Quentin Gaillard)

Alpha-case formation and surface embrittlement induced by heat treatments | 123

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Previously, the SR condition revealed an α-case layer approximately 10 µm thick. In the
HT|α-case scenario, this layer expands to 70 µm. The HT condition, despite having initially
undergone a stress-relief treatment known to induce a 10 µm α-case layer, exhibited no consistent
α-case layer. The most likely explanation is that the oxygen from the α-case layer formed during
stress-relief treatment further diﬀused below the surface during the treatment at 860 °C, resulting
in its dispersion in the core material.

Section 4.2.2 showed that the thin α-case layer formed in the SR condition did not have any
significant impact on tensile properties, including elongation to failure. As a first approach, the
influence of α-case in the SR condition is therefore considered to be negligible, even for fatigue
properties. The only condition considered to induce a significant α-case layer is the HT|α-case
condition. This enables the separation of the influence of microstructure and α-case formation.

Therefore, a comparison between HT and HT|α-case states will provide insights

into the impact of the 70 µm α-case layer in the latter. In contrast, when comparing
SR and HT heat treatments, one can expect only the inﬂuence of the microstructure. The
microstructure after stress relief and after the 860 °C|4 h heat treatment are shown in Fig.4.16a-
b. These ﬁgures reveal a thickening of α laths as well as an increase in β phase fraction induced
by the 860 °C|4 h heat treatment.

Figure 4.16 | SEM-BSE images showing the microstructure of the bulk material
(a) of a SR sample and (b) of an HT|α-case sample. The β phase appears brighter than the α
phase. (Courtesy of Quentin Gaillard)

To quantify the α-laths thickening, two EBSD maps were obtained on two fatigue samples
in the SR and HT|α-case states (HT|α-case and HT conditions are equivalent in terms of bulk
microstructure). The respective EBSD maps were 600 ˆ 445 µm2 and 600 ˆ 500 µm2 in size,
captured with a 0.25 µm step size. The determined α-laths thickness shifted from 0.61 µm in the
SR state to 1.06 µm for the 860 °C|4 h heat treatment.

124 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
4.2.3.2 Inﬂuence of α-case on fatigue properties
Fatigue tests on HT and HT|α-case samples were performed under the same conditions as
those used for the SR specimens (R “ 0.1 and f “ 10 Hz). However, due to time constraints,
tests were stopped here before 107 cycles when no failure occurred.
Fig.4.17 shows the Wöhler curves for the three conditions. Results for M&P specimens
presented in Section 4.1.1 are also shown as a reference.

1200

1000
σmax (MPa)

800
SR
HT
HT|α-case
600
M&P
2 4 Number of
3 superimposed points
400 3 4

200
104 105 106 107
Nf
Figure 4.17 | Wöhler curves for SR, HT and HT|α-case conditions. M&P specimens
are also represented as a reference. They did undergo a stress-relief treatment after polishing (2 h
at 660 °C), ensuring that no residual stress remains from machining. Fatigue tests are performed
at R “ 0.1. All specimens were manufactured vertically.

There seems to be no noticeable diﬀerence in fatigue resistance between the SR and HT

conditions. This suggests that the microstructural changes from the 860 °C|4 h heat
treatment have a limited influence on the fatigue resistance. The small reduction
in fatigue life measured for the HT|α-case condition vanishes at the lowest stress level
tested (300 MPa).
It is worth noting that the differences observed among the SR, HT, and HT|α-case conditions
are much smaller than those seen with M&P specimens. Still, it would be interesting to study
HT and HT|α-case heat treatments on M&P specimens. This may for instance reveal that
in the absence of the most detrimental surface defects, microstructure or α-case have a more
pronounced influence on the fatigue resistance.

Alpha-case formation and surface embrittlement induced by heat treatments | 125

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To better understand the influence of α-case on fatigue properties, the fracture surfaces of
the specimens were examined using SEM. Some HT|α-case samples were also studied with XCT
both before and after failure.
SEM observations revealed that all the cracks that lead to failure started from the surface,
regardless of the heat treatment applied. Additionally, XCT characterization showed that most
cracks began at surface valleys. Intriguingly, one HT|α-case sample displayed a crack initiation
site with no discernible defect on the initially scanned XCT surface, a feature not seen in SR
samples. This might indicate a change in crack initiation mechanisms due to the α-case layer,
but it may also simply be caused by a lack of resolution which does not enable to capture the
killer defect.
Detailed SEM observations revealed traces of the α-case layer in the final failure area of the
broken samples, i.e. the part of the fracture surface formed after the fracture toughness KIC is
reached. For the three heat treatment conditions, dimples are visible in this region (Fig.4.18a-
c), indicating a ductile fracture mechanism. However, near the surface, a transition to a brittle
fracture mode can be seen. The width of this zone is of the order of magnitude of the one of the
α-case layer. Fig.4.18a-c show the surface in the final failure area for the three heat treatment
conditions. One can note from these figures that the width of the fragile fracture layer is not
completely homogeneous. This can for instance be seen in Fig.4.18b, where the fragile layer is
visible on the left, whereas only cupules are visible on the right.

Figure 4.18 | SEM-SE observation of the ﬁnal failure area of fatigue samples, reveal-
ing the brittle behavior of α-case layer (a) for an SR specimen that failed after 262 790 cycles
at 350 MPa, (b) an HT specimen that failed after 22 477 cycles at 550 MPa and (c) an HT|α-case
specimen that failed after 25 852 cycles at 450 MPa.

The observations indicate that the α-case layer can make the surface more brittle. This effect
is especially pronounced for the HT|α-case condition, which has the thickest α-case layer. Thin
brittle layers in the SR and HT conditions might also play a role. While it was previously stated
in Section 4.2.2 that thin α-case layers do not affect tensile properties, their influence on fatigue
could differ. They might, for example, facilitate the early stage of the crack initiation process.
It is interesting here to draw a parallel with the study of [VI 23]. The latter discussed the
influence of a 10 µm brittle oxide layer on the fatigue behavior (R “ 0.1) of 7075-T6 aluminum
alloy. By comparing specimens with and without this oxide layer, it has been shown that
the fatigue lifetime was reduced by a factor of 3 in the presence of the oxide layer. SEM
fractography and acoustic emission analysis during fatigue cycling were used to examine further

126 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
the mechanisms behind this observation. The ﬁndings suggested two potential mechanisms. The
ﬁrst involves the oxide layer cracking, promoting crack initiation in the underlying material.
The second proposes that continuous crack formation in the oxide layer aids in the main crack’s
propagation, a phenomenon referred to as the continuous initiation process.
A parallel can be made here, with the oxide layer corresponding in the present case to the
brittle α-case oxygen-enriched metallic layer. In particular, the hypothesis of a continuous
initiation process seems coherent when looking at our fracture surfaces, which show comparable
appearances as the ones obtained by [VI 23, Fig.13] for specimens with the oxide layer. [VI 23]
noticed that when no oxide layer was present, cracks initiated from point defects. However,
when the brittle oxide layer was present, cracks seemed to initiate from a large circular segment,
possibly due to the occurrence of a continuous initiation process. Such large crack initiations
are typically observed in the present work, as can be seen in Fig.4.19a-c.

Figure 4.19 | SEM-BSE topological mode observation of fracture surfaces for (a)
a SR specimen that failed after 73 547 cycles at 300 MPa, (b) an HT specimen that failed af-
ter 40 901 cycles at 450 MPa and (c) an HT|α-case specimen that failed after 198 645 cycles at
300 MPa. In every case, crack initiation seems to occur from a large portion of the surface.

Yet, the observed fracture surface patterns are consistent across all conditions (SR, HT,
and HT|α-case) and are not speciﬁc to the HT|α-case. One possibility is that α-case always
induces a continuous initiation since some α-case layer exists in all treatments, even though it
seems discontinuous in the HT condition. A second explanation could be that the particular
fracture surface morphology observed comes from the presence of elongated surface valleys rather
than α-case. Indeed, it has been shown in 4.1.3.2 that cracks tend to follow surface valleys
inherited from the L-PBF process. This suggests that surface valleys could either promote
crack initiation or facilitate crack propagation by a continuous initiation-like mechanism. More
advanced characterizations would be required to better understand these mechanisms. Acoustic
emission analysis, as performed by [VI 23], would for instance have brought valuable information.
To verify if the presence of α-case could promote crack initiation, the crack density within
specimens in SR, HT and HT|α-case conditions was studied. For this purpose, the aim was to
count the number of cracks that did initiate, even though they did not propagate until failure.
In other words, cracks were searched below (or above) the fracture surface.

Alpha-case formation and surface embrittlement induced by heat treatments | 127

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
It is worth mentioning here that two types of cracks can be seen on specimens after fracture.
The ﬁrst one will be referred to as ﬁnal failure cracks and is observed primarily on HT|α-case
specimens. Examples of such cracks are given in Fig.4.20a-b.

Figure 4.20 | Observation of ﬁnal fracture cracks in HT|α-case specimens. (a) SEM-
SE image (b) Optical microscope image of a longitudinal cross-section after a chemical attack
with Weck reagent to reveal α-case in white.

Fig.4.20b reveals these cracks to be notably wide but shallow. Their propagation seems to
stop at the end of the α-case layer. They are very similar to cracks observed by Quentin Gaillard
[GAI 23b, Fig.13] on tensile specimens after failure, as shown in Fig.4.14z=a-d. Interestingly,
these cracks reside near the final fracture zone, where the material can be expected to bear very
high loads at the end of crack propagation. Towards the end of fatigue life, in the ligament
ahead of the propagating main crack, the material is submitted to high levels of stress This
likely explains the similarity between the fracture surfaces observed in the final fracture region
and the ones found on tensile specimens.
A closer look reveals the densest clustering of these cracks at the fringe of the final failure
region, as illustrated in Fig.4.20 and schematically represented in Fig.4.21.
Such cracks underscore the potential of the α-case layer to diminish fracture toughness,
accelerating the terminal failure. This behavior mirrors Quentin Gaillard’s findings during
tensile tests, illustrated in Fig.4.13b. Nevertheless, the presence of these cracks is only related
to the final failure phase, excluding the crack initiation or growth phases.
Therefore, in what follows, the focus has been made on a second type of crack. These initiated
and propagated during fatigue cycling, even though they did not necessarily propagate through
the whole specimen section. They will be referred to as fatigue cracks in what follows. Such
cracks are much slimmer and straighter, perpendicular to the loading direction, in contrast to
the final failure cracks. Moreover, their propagation often extends beyond the α-case layer, as
can be seen on the longitudinal cross-section in Fig.4.22.

128 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Crack initiation

Highest density
of ﬁnal
failure cracks
Crack propagation
area

Final rupture
area

Final failure cracks

Figure 4.21 | Schematic representation of an HT|α-case fracture surface, showing
the preferential locations of ﬁnal failure cracks. Adapted from [LAR 23].

Figure 4.22 | Fatigue cracks observed by optical microscopy on an HT|α-case spec-

imen longitudinal cross-section. The cross section was polished and subjected to Weck
reagent, which best revealed thin cracks.

Alpha-case formation and surface embrittlement induced by heat treatments | 129

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
An extensive study was conducted using 11 SR, 9 HT and 11 HT|α-case specimens tested at
different stress levels. After failure, a longitudinal cross-section was obtained from one-half of
each specimen. This cross-section was polished, and the Weck reagent was applied. This proved
effective in highlighting thin cracks, as illustrated in Fig.4.22. Within a specified distance h
from the fracture surface (see Fig.4.23a), fatigue cracks were counted. The resulting count was
normalized by the length of the inspected surface, yielding a crack linear density in mm−1 , as
presented in Fig.4.23b.

b 5
a
Fatigue crack linear density (mm-1)
SR
HT
4 HT| -case
Run-out
h 3

2
e
1.1⤫e 1

0
250 300 350 400 450 500 550 600 650 700
σmax (MPa)

Figure 4.23 | Fatigue crack linear density measurements on SR, HT and HT|α-case
specimens at diﬀerent stress levels. (a) Scheme specifying the height h over which cracks
are counted for each specimen. (b) Cracks densities measured for all specimens, as a function
of the maximum stress during fatigue test.

The results shown in Fig.4.23b indicate that the crack density increases with the value of
maximum cyclic stress for all SR, HT, and HT|α-case specimens. Such a trend is anticipated,
aligning with ﬁndings that a greater number of nucleated cracks appear at elevated stress levels
[SCH 09, Fig.2.20].
Further analysis of Fig.4.23b reveals that, excluding the readings at 300 MPa, the HT|α-case
specimens exhibit considerably higher fatigue crack densities. This supports the theory that the
70 µm α-case layer in HT|α-case specimens promotes crack initiation during cycling,
potentially reducing fatigue lifespan. Densities for SR and HT specimens are similar, with a
marginally elevated count for SR. This could be due to the wider and more regular α-case layer
present in SR specimens, although it is only 10 µm wide.
A deeper exploration could shed light on how the α-case aﬀects fatigue properties. For
instance, testing specimens with thicker α-case layers or M&P specimens with an equivalent
α-case layer would provide complementary information. Alternative characterization techniques
could have also proven useful. For instance, using acoustic emission, as explored by [VI 23], could
provide insights into the timing of crack initiation during cyclic loading and its relationship with
the α-case presence.

130 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
This research establishes that the 70 µm α-case layer developed during heat treatment slightly
diminishes the fatigue resistance of L-PBF Ti64 specimens. Combined with the surface rough-
ness considerations discussed in Section 4.1, it underscores the significance of L-PBF Ti64’s
surface integrity on fatigue properties. M&P specimens’ superior properties hint that strategic
surface polishing might significantly enhance fatigue attributes by both smoothing the surface
and eliminating the α-case layer. Although the impact of surface roughness seems much more
pronounced than the one of the α-case layer, the use of surface polishing could still be interesting
regarding the α-case issue. Indeed, one of the advantages of using material-removal polishing is
that it would enable heat treatments to be carried out with fewer constraints on the vacuum to
be achieved in the furnace, as the α-case layer (if of a reasonable depth) would be eliminated by
polishing.
However, polishing additively manufactured components poses challenges. Such components
typically have rougher surfaces than their machined counterparts. Moreover, additive manufac-
turing techniques like L-PBF are prized for making parts with intricate designs, making most
polishing methods unsuitable due to inaccessibility. Despite these issues, various technologies
aim to address these challenges [MAL 21, KHA 21, NUT 19, BOB 21]. In this PhD work, a
specific technique, Plasma Electrolytic Polishing (PeP), was explored to enhance fatigue prop-
erties.

4.3 Improvement of fatigue properties by Plasma electrolytic

Polishing (PeP)
4.3.1 Principle of PeP surface treatment
Plasma electrolytic Polishing (PeP) is a surface finishing method for metals and alloys. It
was first introduced in 1979 [DUR 79] and has since aroused increasing interest. Several studies
and reviews mention its application to AM parts [CUT 22, SEO 21, NAV 22, ZEI 22, HEN 18,
BAS 22]. This can be partly explained by its ability to treat surfaces of complex geometries.
As an example, Fig.4.24a-b show results from a study that characterized titanium additive
manufactured parts before and after PeP [NAV 22]. Significant improvements can be seen, both
in terms of roughness and fatigue properties.
Fig.4.25 shows schematically the PeP process principle. PeP setup is at first sight similar
to the one that could be used in conventional Electrolytic Polishing (EP). The metallic part to
polish is placed in a temperature-controlled electrolytic bath and connected to a cathode. A
voltage difference is then applied between the workpiece (acting as an anode) and the cathode.
However, PeP differs from EP in many aspects.
Firstly, polishing mechanisms are different. They are multiple and complex in the case
of PeP, and still a topic of discussion nowadays [BAS 22, HUA 21, BEL 20]. Without going into
too much detail, the reactions occurring at the anode produce gas (O2 ) which rapidly forms an
insulating gas layer around the workpiece − see Fig.4.25. Due to the high potential difference
between the anode and the electrolyte, this so-called vapour-gaseous envelope (VGE) is ionized
which leads to the formation of a plasma. The polishing action comes both from the chemical
reaction with the electrolyte and the action of this plasma layer.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 131

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.24 | Roughness and fatigue characterization of titanium additive manu-
factured parts (a) before and (b) after PeP [NAV 22]. Parts present eight tensile stress
concentration spots that are highlighted in (a). They were tested with a constant amplitude at
R “ 0.1 and a peak load of 11.793 kN. Fatigue data is estimated from [NAV 22, Fig.6].

Figure 4.25 | Principle scheme of PeP process drawn from [HUA 21].

132 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Because polishing mechanisms are different from the ones active in EP, PeP does not re-
quire the use of a cathode whose geometry is similar to the one of the workpiece.
Simple geometries such as rings or plates can therefore be used [ZEI 22], which is a huge advan-
tage for the treatment of parts of complex geometry. Some constraints still exist, like the need
for the cathode to be of significantly larger size than the workpiece [ZEI 22, NAV 22, HUA 21]
which could be limiting for large industrial parts. Furthermore, material removal rate di-
rectly depends on the immersion depth [ALE 05, BEL 20]. This issue comes from the
effect of hydrostatic pressure on the gas layer. Some inhomogeneity can therefore be expected,
especially for large parts. However, some solutions could exist in those cases, such as the use of
electrolyte jet PeP [HUA 21] where the electrolyte is used in the form of a jet propelled towards
the workpiece rather than a bath.
One of the characteristics of PeP is the high applied voltage and current density which
are needed for the gas layer to form and the polishing to occur. Typical voltage values range
from 200 to 350 V and current densities from 0.2 to 0.5 A/cm2 [HUA 21]. This makes PeP much
more demanding than conventional EP and can require the use of costly and energy-consuming
generators for large parts.
On a more positive side, electrolytes used for PeP (e.g. made of low-concentration salt
solutions) are generally considered to be less harmful to the environment than the ones
used in conventional EP [BAS 22, ZEI 22, NAV 22, HUA 21]. For titanium alloys, the most
commonly used electrolyte is a (low concentrated) fluorine salt-based aqueous solution [HUA 21].
On the whole, PeP presents interesting perspectives for the surface treatment of metallic
additive manufactured parts of complex geometries. However, as it is more recent than other
methods such as (electro-)chemical polishing or shot-peening, more studies are needed to esti-
mate its practical relevance.

4.3.2 Conditions of PeP treatments applied to fatigue specimens

To investigate the effect of PeP treatment on fatigue properties, specimens with as-built
surfaces were utilized. These specimens only underwent a 720 °C|2 h stress-relief heat treatment
after production. Given the shallow depth of the α-case layer in this state (10 µm), it is believed
that PeP’s effects on fatigue properties are not due to the α-case layer’s removal.
PeP treatments were performed by the French company Ineosurf on several Z and 45° spec-
imens. Specific PeP operational details remain confidential. To assess the evolution of surface
roughness, specimens underwent XCT scanning both before and after PeP. The XCT acquisition
and data processing conditions are the same as those of the as-built specimens, as outlined in
3.4.1.
Different surface treatment conditions were applied depending on the build orientation. Two
treatment durations were assigned for Z specimens: 40 min and 1 h 20 min. All 45°
specimens received a 40 min polish. However, some of the 45° specimens were oriented
at 45° in the electrolyte bath. The positioning of the specimens in the electrolyte during
treatment is illustrated in Fig.4.26. Treatment conditions in subsequent discussions will be
labeled as Z|Vertical Vertical Vertical 45°
40min , Z|1h20 , 45°|40min , and 45°|40min .
Fig.4.28 shows a polished specimen and the number of specimens available for each condition.
It also indicates by a red cross specimens for which PeP treatment failed and led to a clear
degradation of surface state on part or the whole specimen − see Fig.4.27 for an example. The
reason behind these dysfunctions remains unknown.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 133

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.26 | Position of the specimens in the electrolyte bath during PeP treatment.

Figure 4.27 | 45°|Vertical

40min specimen surface after a failed PeP treatment.

134 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.28 | PeP conditions for Z and 45° fatigue specimens. "Position" refers to the
specimen orientation of the specimen in the electrolytic bath. Red crosses indicate the number
of specimens for which treatment failed, see Fig.4.27 for an example. Note that the hole used
to hold the specimens in the bath is visible on the 45° specimen support structures. The hole
was mistakenly made for the 45° specimen in the vertical condition, but another hole was made
on the opposite head to allow the treatment to be performed in the vertical position.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 135

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
4.3.2.1 Analysis of material removal
Tab.4.1 indicates the thickness of removed material for all the treatment conditions. For a
given specimen, it corresponds to the diﬀerence between the minimum radius of the as-built
specimen and the minimum radius of the polished specimen, both measured from XCT data.

Treatment condition Z|Vertical

40min Z|Vertical
1h20 45°|Vertical
40min 45°|45°
40min
Matter removal (µm) 182 ˘ 22 345 ˘ 3 171 ˘ 7 223 ˘ 15

Table 4.1 | Matter removals measured for the diﬀerent PeP treatment conditions.
It corresponds to the diﬀerence between the minimum radius of the as-built specimens and the
minimum radius of the polished specimen, both measured from XCT data. The values shown
are the average and standard deviation for all specimens treated with the same condition.

The thickness of removed material is approximately 180 µm for specimens treated for 40 min
(i.e. Z|Vertical Vertical 45° Vertical condition, material removal
40min , 45°|40min , and 45°|40min conditions). For the Z|1h20
is nearly double, at approximately 340 µm.
Polishing at 45° resulted in a slightly higher matter removal than polishing in the vertical
position for 45° specimens, although the diﬀerence is small and could be considered within the
measurement uncertainty. Although no clear explanation was found, it could be attributed to
the specimen’s position in the bath. The position might alter the gas layer behavior around the
samples during PeP, inﬂuencing the material removal rate.
It should be noted that the polishing is not homogeneous across the specimen surfaces.
Fig.4.29 demonstrates that material removal thickness evolves along the specimen axis. Note
that each curve in Fig.4.29 represents a single specimen, not an average as in Fig.4.28.

450

400
Matter removal thickness (µm)

350

300 Slopes from

linear regression

250 - 8.25 µm/mm

- 12.15 µm/mm
200
+ 5.02 µm/mm
150 - 8.80 µm/mm
Gauge
100
length
50

0
4 2 0 2 4
Position along specimen axis (mm)

Figure 4.29 | Matter removal thicknesses measured along the specimen axis of 4
specimens which underwent PeP with diﬀerent treatment conditions. By convention,
the position along the specimen axis decreases in the direction of the built plate (for instance in
the direction of the residual support structures for 45° specimens.)

136 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Material removal appears to have a linear trend along the specimen axis. Maybe the most
striking point here is that the slope of the linear regression is negative for the Z|Vertical Vertical
40min , Z|1h20
45° Vertical
and 45°|40min conditions, whereas it is positive for the 45°|40min one. This can be correlated
to the position of the specimens during the PeP treatment. For the Z|Vertical Vertical and
40min , Z|1h20
45°
45°|40min conditions, the specimens are pointing downwards. Hence, the matter removal rate is
at its maximum for the smallest immersion depth. For the 45°|Vertical40min condition, specimens are
pointing upwards and therefore the positive slope shows that here again, the matter removal
rate is at its maximum for the smallest immersion depth.
However, these results seem difficult to explain regarding the existing literature. Indeed, the
effects of immersion depth on the material removal rate were found to be explained by variations
in the hydrostatic pressure. Hydrostatic pressure is an important parameter since it influences
the shape and width of the gas layer, which itself conditions the treatment efficiency. Such an
effect makes the material removal rate higher for higher immersion depths [ALE 05, BEL 20],
which is the exact opposite of what is observed in the present work.
Furthermore, the effect of hydrostatic pressure seem much less pronounced in the study of
[ALE 05] on a 12Cr-18Ni-10Ti steel than what is seen here. For instance, assuming a treatment of
40 min, the evolution of the thickness of removed material with immersion depth was according
to their measurements of 0.15 µm/mm, which is about 50 times less than what is observed
in the present study. One can also add that the material removal thickness evolution seems
very similar between the Z|Vertical 45°
40min and the 45°|40min conditions, although for this specimen the
immersion depth variation is smaller than for the 45°|45° 40min condition where the specimen is
oriented at 45°. Inversely, the slope is smaller of the 45°|Vertical
40min condition. All these observations
suggest that the material removal rate variations along the specimens’ axis are likely induced by
another factor than hydrostatic pressure variation. Since no other explanation could be found,
the observed material removal inhomogeneity will be simply taken into account as
they are in the subsequent analysis.
Chapter 3 highlighted that a unique microstructure and pore distribution are observed near
the surface. It is thus relevant to analyze the relationship between material removal thickness and
these distributions under different polishing conditions. Key findings are compiled in Fig.4.30.
Only the average matter removal presented in Tab.4.1 is shown here, but it is worth keeping in
mind that matter removal can vary among specimens and within a specimen itself.
Fig.4.30 superimposes:

• the distributions of the pores distance to the surface for both Z and 45° specimens
• a crystallographic orientation map (EBSD) of the reconstructed β0 grains close to the
surface
• vertical lines showing the mean value of the thickness of the removed material for each
PeP treatment condition

Comparing average material removals with pore distance distributions, the Z|Vertical
1h20 treat-
ment seems to entirely eliminate sub-surface porosity. For other conditions, the removal depth
aligns with the tail-end of sub-surface porosity. This could imply that residual sub-surface
pores could get exposed during polishing, forming new surface ﬂaws that would negatively
impact fatigue resistance.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 137

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.30 | Relation between matter removal during PeP, pores distance to surface
and near-surface microstructure. Pore distance to surface distributions is a zoomed view of
Fig.3.14. Similarly, the near-surface microstructure in the background is an EBSD orientation
map of prior β0 grains that has been shown in Fig.3.10.

Similarly, for the Z|Vertical

1h20 condition, polishing exceeds the microstructural contour region
discussed in 3.2. The surface microstructure after PeP resembles the border chessboard mi-
crostructure referenced in 3.2. For other conditions, the end of polishing can be expected to
lie astride the boundary between contour and border regions. EBSD analyses were conducted
on Z|Vertical Vertical specimens to verify these observations and ensure PeP did not alter
40min and Z|1h20
surface microstructure. As expected, the surface microstructure after PeP corresponds to the
border region that has been characterized on as-built samples. Detailed results are available in
Appendix C.
Lastly, residual stress measurements were also performed on one of the specimens’ heads and
led to the conclusion that PeP did not generate any residual stress.

4.3.3 Roughness of specimens polished by PeP

Due to roughness inhomogeneity in as-built specimens (Z vs 45°; up-skin vs down-skin)
and in the subsequent PeP treatment, different surface states are obtained after PeP. They
will be presented for each treatment condition, and finally summarized in terms of roughness
parameters.
Fig.4.31 displays the surface of a Z specimen before and after a 40 min PeP treatment
(Z|Vertical
40min condition). Notably, the original surface roughness disappears. The deepest valley,
however, remains at 38 µm. At greater magnification, tiny sharp pits become visible, contribut-
ing to this measured depth. These pits are due to sub-surface pores exposed during
polishing, as highlighted in Section 4.3.2.1. Such defects could compromise fatigue performance
despite the otherwise smooth surface.

138 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.31 | Surface of a Z specimen before and after a 40 min PeP treatment
(Z|Vertical
40min condition). The color used to represent the surface combines the roughness (using
λc “ 0.8 mm) as well as the minimum curvature (using rcurv “ 20 µm) to exacerbate valleys
and details in general. Although PeP completely smoothed the original surface roughness, a few
small pits can be identiﬁed on the surface. They are attributed to pores that were brought to
the surface by polishing.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 139

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.4.32 shows that after a longer treatment (1 h 20 min), surface pits can no longer be
seen. The maximum Sv value measured in this condition is accordingly low (6 µm). This com-
forts the hypothesis that those defects originate from sub-surface pores in Z|Vertical
40min specimens,
Vertical
since they are all eliminated for Z|1h20 specimens. Due to the higher matter removal, the
specimen section is however greatly reduced, which may cause other dimensional issues.

Figure 4.32 | Surface of a Z specimen before and after a 1 h 20 min PeP treatment
(Z|Vertical
1h20 condition). The color used to represent the surface combines the roughness (using
λc “ 0.8 mm) as well as the minimum curvature (using rcurv “ 20 µm) to exacerbate valleys and
details in general. This 1 h 20 min treatment enables to erase all the initial roughness, and no
surface holes caused by pores can be seen as in Fig.4.31 for Z|Vertical
40min condition.

140 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The inﬂuence of PeP on the 45°|Vertical
40min condition is demonstrated in Fig.4.33. Polishing
yields a uniformly smooth surface, but upon closer inspection, the down-skin region reveals
some pits. They bear resemblance in depth and form to those seen in the Z|Vertical
40min condition.
Vertical
Since the matter removal thickness for the 45°|40min condition (171 ˘ 7µm) is much larger than
the maximum valley depth of the as-built specimen (88 µm), it is considered here that these
defects originate as well from sub-surface pores that were brought to the surface by PeP.
Looking at the two magnifying windows, it is possible to notice that most of the surface pits
are found in the bottom region of the sample. This observation is consistent with the results
shown in 4.3.2.1. Indeed, Fig.4.30 showed that the polishing was done until the near-end of the
sub-surface pores. Fig.4.29 showed on its part that the matter removal thickness was lower at
the bottom of the specimen. Thus, the higher density of defect at the bottom in Fig.4.33 may
be explained by the lower matter removal in this region, which did not enable the complete
removal of the surface layer where most pores are located.

Figure 4.33 | Surface of a 45° specimen before and after a 40 min PeP treatment
performed vertically (45°|Vertical
40min condition). The color used to represent the surface com-
bines the roughness (using λc “ 0.8 mm) as well as the minimum curvature (using rcurv “ 20 µm)
to exacerbate valleys and details in general. A few small pits can still be identiﬁed on the surface
after PeP, especially at the bottom of the specimen.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 141

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.4.34 shows that for 45°|45°
40min specimens, PeP leaves significant defects, particularly at the
top of the down-skin region. Their presence here, and not in 45°|Vertical
40min specimens, likely stems
from different material removal profiles between the conditions. As evidenced in Fig.4.29, the
45°|45°
40min condition has minimal material removal at the top of the down-skin area. Addition-
ally, this region is nearly horizontal during fabrication, resulting in maximal roughness. The
combination of this initial roughness and minimal material removal explains the presence of
residual valleys after PeP.

Figure 4.34 | Surface of a 45° specimen before and after a 40 min PeP treatment at
45° (45°|45°
40min condition). The color used to represent the surface combines the roughness
(using λc “ 0.8 mm) as well as the minimum curvature (using rcurv “ 20 µm) to exacerbate
valleys and details in general. Large and deep valleys can still be identiﬁed at the top of
the specimen. They are remains of the down-skin surface roughness. Most of the valleys in
question are not taken into account by roughness measurements since they are out of the zone
characterized by XCT. An optical microscopy image shows that defects grow even larger outside
this zone.

142 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
However, since XCT scans only enable the characterization of a 9 mm high portion of the
specimens, these large defects are mostly not accounted for. The optical microscope observation
in Fig.4.34 shows that even larger defects can be seen in higher parts of the specimens that
are not characterized by XCT. Yet, as they are far away from the specimens’ gauge area, their
impact on fatigue tests might be minimal.
Finally, 45°|45°
40min specimens are the only ones for which large defects remain after PeP. These
defects are located in the upper part of the down-skin region. As evidenced in Fig.4.29, the
45°|45°
40min condition has minimal material removal at the top of the down-skin area. Additionally,
this region is closest to horizontal during creation, resulting in maximal roughness (see Fig.4.6c).
The combination of this initial roughness and minimal removal can explain the residual valleys
after PeP and their location in the upper part of the down-skin region.
It is important to notice here that the XCT characterization is limited to a height of 9 mm,
which is barely enough to reach the beginning of those defects. The optical microscope image
in Fig.4.34 shows indeed that many defects can be seen further in the radius ﬁllet. Although
these defects are very large, the fact they are not situated in the specimens’ gauge length may
lead in practice to a limited impact on fatigue tests.
To understand PeP’s impact on roughness evolution more broadly, the evolution of Sa and
Sv roughness parameters were analyzed for each specimen. Vertical specimens’ roughness evo-
lution (Z|Vertical Vertical conditions) is showcased in Fig.4.35. Each specimen’s roughness
40min and Z|1h20
parameters, before and after PeP, are illustrated. The arithmetical mean height Sa on the x-axis
reﬂects the overall average surface condition, while the maximum valley depth Sv on the y-axis
sheds light on the deepest remaining surface defects.

Before PeP (as-built)

80
After PeP

60
Sv (µm)

Z
0
1 2 3 4 5 6 7 8
Sa (µm)
Figure 4.35 | Evolution of Sa and Sv roughness parameters induced by the PeP
treatment of Z specimens. Roughness was computed using λc “ 0.8 mm and λS “ 0.03 mm.
Dashed lines are represented only to enable the identiﬁcation of points that were obtained before
and after the PeP treatment of a given specimen. They do not mean that a linear relationship
exists between Sa and Sv parameters.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 143

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
It can be seen on Fig.4.35 that Sa and Sv values are very reproducible for both Z|Vertical
40min and
Z|Vertical
1h20 conditions. Similar Sa values are reached in both cases. This is in good agreement with
a preceding discussion concerning Fig.4.31 and Fig.4.32 where it could be seen that on average,
very smooth surface states are reached for both conditions.
However, using the Sv parameter which is more sensitive to the deepest defects, a clear
distinction appears between the two conditions. S¯v equals to 5 µm for the Z|Vertical
1h20 condition,
which is less than two voxels. In other words, no actual notch-like defect could be detected. On
the other side, S¯v equals to 37 µm for the Z|Vertical
40min condition, revealing the presence of pores
brought to the surface after polishing as identiﬁed in Fig.4.31.
Considering 45° specimens, Fig.4.36 reveals that after polishing, down-skin areas remain
comparatively rougher. For instance, in the 45°|45° ¯
40min condition, Sa equals to 2.36 µm in down-
skin and to 0.88 µm in up-skin.

Before PeP (as-built)

175
After PeP
150

125
Sv (µm)

100

25
45° down-skin
5 10 15 20 25
Sa (µm)
Figure 4.36 | Evolution of Sa and Sv roughness parameters induced by the PeP
treatment of 45° specimens in the down-skin region. Roughness was computed using
λc “ 0.8 mm and λS “ 0.03 mm. Dashed lines are represented only to enable the identiﬁcation
of points that were obtained before and after the PeP treatment of a given specimen. They do
not mean that a linear relationship exists between Sa and Sv parameters.

Results are also more scattered, especially for Sv . For example, the 45°|45°40min specimen
illustrated in Fig.4.34 shows an abnormally high Sv value of 97 µm. Fig.4.34 showed the existence
of large remaining valleys in the upper part of the down-skin region, responsible for this high
measured Sv . Furthermore, the optical microscope observation has shown that even larger and
deeper defects are present further away from the gauge length. The same is true for other
45°|45°
40min specimens, which means that their relatively low value is due to the fact that the Z
span of the XCT scan is not suﬃcient to capture these defects. Sv values for the 45°|45° 40min
condition can thus be considered to be underestimated. However, one may also argue
that the large defects in question should not be taken into account, since they are too far from
the gauge length. This makes crack initiation unlikely in this area.

144 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
On the contrary, 45°|45°
40min specimens are smoother in the up-skin region than 45°|40min
Vertical

specimens, see Fig.4.37. They even show results comparable to Z|1h20 Vertical specimens with an
average S¯v of 5 µm. These results may indicate that surfaces oriented downwards during PeP
are better polished than surfaces oriented upwards, since for the 45°|45°
40min condition, the up-skin
region is oriented downwards.

Before PeP (as-built)

60 After PeP

40
Sv (µm)

10
45° up-skin
1 2 3 4 5 6
Sa (µm)
Figure 4.37 | Evolution of Sa and Sv roughness parameters induced by the PeP
treatment of 45° specimens in the up-skin region. Roughness was computed using λc “
0.8 mm and λS “ 0.03 mm. Dashed lines are represented only to enable the identiﬁcation of
points that were obtained before and after the PeP treatment of a given specimen. They do not
mean that a linear relationship exists between Sa and Sv parameters.

Table 4.2 presents a variety of roughness parameters calculated after PeP across the four
tested conditions. For a more detailed description of roughness parameter computations, refer
to Chapter 2 and Section 3.4.3.2.
While many key insights from these results are highlighted in previous figures, a notable
observation is the predominance of negative skewness Ssk in most cases. This contrasts with
the mainly positive skewness seen in as-built specimens. This difference arises because, post-
PeP, the surface mostly consists of smooth areas interspersed with occasional notches that have
negative local heights. Such a pattern is characteristic of surfaces dominated by valleys, which
result in negative skewness.
Regarding kurtosis, elevated values in some conditions can be correlated to the existence
of occasional, yet notably deep, defects. The observed variability suggests that the occurrence
and depth of these defects can differ significantly between specimens. As a result, variations
in fatigue behavior might also be anticipated. Conversely, the low kurtosis value for
conditions like Z|Vertical
1h20 signifies the lack of exceptionally deep surface irregularities. It indicates
that PeP effectively produces a consistently smooth and even surface. Such a surface is
anticipated to exhibit superior fatigue resistance.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 145

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
45°|Vertical 45°|45°
Z|Vertical
40min Z|Vertical
1h20
40min 40min
Down-skin Up-skin Down-skin Up-skin
Sa (µm) 1.11 ˘ 0.02 0.82 1.83 ˘ 0.35 1.09 ˘ 0.13 2.36 ˘ 0.31 0.88 ˘ 0.07
Sq (µm) 1.43 ˘ 0.02 1.03 2.30 ˘ 0.50 1.41 ˘ 0.18 3.22 ˘ 0.69 1.08 ˘ 0.10
Sz (µm) 43 ˘ 2 9 34 ˘ 14 26 ˘ 7 59 ˘ 42 8˘2
Sv (µm) 37 ˘ 2 5 26 ˘ 12 21 ˘ 7 42 ˘ 33 5˘2
Ssk ´0.12 ˘ 0.11 ´0.03 ´0.38 ˘ 0.06 ´0.44 ˘ 0.48 ´0.54 ˘ 0.81 0.02 ˘ 0.06
Sku 4.78 ˘ 1.01 ´0.10 1.01 ˘ 0.82 5.65 ˘ 4.57 8.41 ˘ 12.52 ´0.35 ˘ 0.05

Table 4.2 | Roughness parameters measured for specimens polished by PeP. Rough-
ness was computed from XCT data using the 3D characterization workflow introduced in Chapter
2. The cut-off values for filters were set to λc “ 0.8 mm and λS “ 0.03 mm. Roughness param-
eter values are presented in the format (mean ˘ std) for conditions where at least 3 specimens
were available. Otherwise, only the mean value is given.

4.3.4 Fatigue resistance improvement and change in crack initiation mecha-

nism

Having characterized the surface state post-PeP, the next step is to assess its impact on
fatigue properties. Uniaxial fatigue tests were conducted under conditions identical to those
of as-built specimens, with details available in Section 4.1.1. Unlike as-built specimens, the
PeP specimens’ fatigue tests were stopped at 2 ˆ 106 cycles when no failure occurred. Non-
failed specimens underwent further testing at elevated stresses for 2 ˆ 106 cycles, increasing the
maximum stress value until failure and starting again the cycle count from zero at each stress
level. On Wöhler curves, 2 ˆ 106 cycles run-outs get marked by an arrow, and points from the
same specimen tested under varying stresses are connected with a dashed line.

Due to the substantial material removal by PeP, XCT was used to measure the specimen’s
section to set stress levels correctly. The average section across the gauge length was used.
Results from M&P and as-built specimens are shown to allow for comparison and better gauge
the enhancements resulting from PeP.

The S-N diagram in Fig.4.38 displays results for 45° specimens. It can be seen from Fig.4.38
that PeP signiﬁcantly enhances fatigue life. A comparable enhancement was observed for both
45°|Vertical 45°
40min and 45°|40min conditions, though this would require further specimen testing for vali-
dation.

Considering available data, the fatigue strength for 45° specimens post-PeP polishing is esti-
mated between 450 MPa and 550 MPa. This translates to an 80%-120% improvement over
as-built specimens. Although signiﬁcant, this enhancement is still far from the one provided
by machining and manual polishing. The limited improvement can be primarily attributed to
pores either exposed on the surface or slightly beneath it by PeP. Fig.4.39a-b illustrate
two examples of crack initiation sites on 45°|Vertical
40min specimen fractures where such defects can
be identiﬁed.

146 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1200
45° R = 0.1
1000 As-built
PeP
σmax (MPa)

PeP
800 M&P
2 Number of
3 superimposed
points
600 2
2
400 32

2
200
104 105 106 107
Nf
Figure 4.38 | Wöhler for 45° specimens before and after PeP in both 45°|Vertical
40min and
45°
45°|40min conditions. Results for 45° as-built specimens and M&P specimens are also shown
for comparison.

Fig.4.39a corresponds to the failure on a pore brought to the surface by PeP. It is the most
frequent initiation site observed. It can be seen that the inner surface of the defect is somehow
degraded, probably under the effects of PeP treatment (see Fig.4.39b for comparison with a
subsurface pore with a smooth wall/inner surface).
Fig.4.39b shows another case, where the failure occurred from a sub-surface pore that was
brought just beneath the surface by PeP. These two examples confirm the detrimental effect of
the few remaining defects on the fatigue properties of specimens polished after PeP, especially
if the polishing depth falls close to the sub-surface layer where most pores are located.
For one 45°|45°
40min specimen, the main crack initiated from a residual surface valley which was
not entirely suppressed after PeP. The specimen is the one shown in Fig.4.34, which presented
the largest defect among all 45°|45°
40min specimens. The initiation site can be seen in the SEM-SE
observation in Fig.4.40a.
One crack initiation site found on another 45°|45°
40min specimen, shown in Fig.4.40b, could also
correspond to a residual surface valley, although its origin could not be precisely identified. It
may as well be a pore brought to the surface as the ones in Fig.4.39a-b. Nonetheless, this shows
that insufficient polishing, especially in down-skin areas, residual valleys can facilitate fatigue
failure. However, similar fatigue lives for both 45°|Vertical 45°
40min and 45°|40min suggests this is not a
major influence on fatigue resistance in our case. This can be explained by the fact that residual
valleys are mostly located in the upper part of the specimen, far away from the gauge area.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 147

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.39 | Crack initiation sites of 45°|Vertical
40min specimens, observed by SEM-SE.
(a) Pore brought to the surface by PeP. The specimen failed after 54 092 cycles at 550 MPa. (b)
Pore brought in sub-surface by PeP. Only a thin ligament separates the pore from the surface.
The specimen failed after 1 173 759 cycles at 550 MPa.

Figure 4.40 | Crack initiation sites of 45°|45° 40min specimens, observed by SEM-SE.
(a) Residual surface valley in the upper part of the down-skin region. This observation has
been observed on the same specimen that was used for Fig.4.34. The specimen failed after
2 ˆ 106 cycles at 450 MPa followed by 91 990 cycles at 550 MPa. (b) Crack initiation site whose
origin could not be found with certainty. It corresponds most probably to a residual surface
valley or a pore brought to the surface by PeP. The specimen failed after 2 ˆ 106 cycles at
550 MPa followed by 55 086 cycles at 650 MPa.

148 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
1200
Z R = 0.1
1000 2
As-built
2
PeP
σmax (MPa)

800 PeP
M&P
Number of
600 4 3 superimposed
points

400 4
4
3
200
104 105 106 107
Nf
Figure 4.41 | Wöhler for Z specimens before and after PeP in both Z|Vertical 40min and
Vertical
Z|1h20 conditions. Results for Z as-built specimens and M&P specimens are also shown for
comparison.

Fig.4.41 presents the S-N diagram for Z specimens. The data indicates that the Z|Vertical
40min
polishing yields lesser improvements than observed for 45°|Vertical
40min specimens. The enhancement
in terms of fatigue strength at 2 ˆ 106 cycles can be estimated to be around 150 MPa. In terms
of initiation site, all failures of Z|Vertical
40min specimens occurred from pores brought to the surface
by PeP, similarly to what has been shown in Fig.4.39a.
However, the Z|Vertical
1h20 condition displayed impressive fatigue resistance, as evi-
denced by one specimen enduring successive cycles at increasing stresses and finally failing at
1100 MPa. As a reminder, tensile tests in Section 3.3 enabled to estimate the yield strength of
Z specimens to be about 1098 MPa. This exceptional resistance might result from the gauge
section being particularly small in the Z|Vertical
1h20 condition, reducing the probability of having a
large defect.
Thus, Fig.4.35 showed for instance that no significant surface defects could be found on
Z|Vertical
1h20 using the XCT scan with a 3 µm voxel size. As a result, the crack leading to failure
initiated from internal defects in this case.
In terms of porosity, only a few pores could be identified on XCT scans in both Z|Vertical
1h20
specimens. For instance, only 5 pores could be identified within the specimen that showed the
highest fatigue resistance. Apart from one having an equivalent diameter of 32 µm, all have
equivalent diameters inferior to 13 µm. The sub-surface pore that caused crack initiation could
not even be identified on the XCT scan due to resolution limitation. It is shown on the SEM-SE
fractography in Fig.4.42a.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 149

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.42 | Crack initiation sites of Z|Vertical
1h20 specimens, observed by SEM-SE. (a)
9 µm spherical sub-surface pore. The specimen underwent 2 ˆ 106 cycles successively at 550 MPa,
650 MPa, 750 MPa, 850 MPa, 950 MPa, 1000 MPa, 1050 MPa and failed after 924 055 cycles at
1100 MPa. (b) Unidentiﬁed ﬂat defect. The specimen underwent 2 ˆ 106 cycles successively at
950 MPa and 1000 MPa, and failed after 155 739 cycles at 1050 MPa.

Even fewer pores could be identiﬁed on the second Z|Vertical

1h20 specimen, with 3 pores in total
and a maximum equivalent diameter of 17 µm. However, crack initiation still occurred
from a larger sub-surface defect, as can be seen in Fig.4.42b. This defect could however not be
identified on the XCT scan either, probably because of its very flat shape. The nature of this
defect is very puzzling due to its very particular rough aspect. No convincing hypothesis could
be found to explain its origin.
Fig.4.43 offers a summarized view of the fatigue results for PeP specimens. While as-built
specimens mostly show crack initiation on surface valleys, PeP-treated specimens for 45°|Vertical
40min ,
45°|45°
40min and Z| Vertical conditions exhibit initiation primarily on surface or near-surface pores.
40min
This change in the crack initiation site is associated with a limited gain in terms of fatigue
resistance. When material removal goes beyond the sub-surface layer that gathers most pores,
as is the case for Z|Vertical
1h20 specimens, a drastic increase of fatigue resistance is observed. Only
a few defects remain in this condition, often of reduced size. However, a high scatter can be
expected in this condition, since a single larger defect would certainly significantly reduce the
fatigue resistance.
To give a better idea of predominant initiation sites for each PeP condition, Tab.4.3 counts
the number of surface and sub-surface crack initiations for each PeP condition. Results were
obtained by analyzing both SEM fractography and XCT scans before/after failure.

150 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 4.43 | Evolution of fatigue resistance and crack initiation mechanisms with
increases PeP matter removal. Contour pores refer to the pores located at about 140 µm
from the surface, between the contour scans and the bulk material (see Section 3.4.2).

Z|Vertical
40min Z|Vertical
1h20 45°|Vertical
40min 45°|45°
40min
Pore brought to the surface 5 2 2
Pore brought near the surface 1 1
Residual valley 1
Unidentiﬁed nature 1 1

Table 4.3 | Number of crack initiation on surface defects (mostly pores brought to
the surface by PeP) and sub-surface pores for each PeP condition. The number of
defects that could not be identiﬁed on the surface extracted from the XCT scan is also speciﬁed.
This issue comes either from a lack of resolution or simply from the fact that the sub-surface
defect had no connection on the surface and was therefore not present in the extracted surface.

Improvement of fatigue properties by Plasma electrolytic Polishing (PeP) | 151

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary
This chapter focused on assessing the uni-axial (R “ 0.1) fatigue properties of L-PBF Ti64
with diﬀerent post-treatment conditions. Initial tests targeted specimens with as-built
surfaces, which had only been subjected to a stress-relief heat treatment for 2 h at 720 °C.
Specimens were manufactured both vertically and at 45°, to gauge the inﬂuence of build
orientation.

For as-built specimens, all failures originated from the surface. Detailed insights from 3D
XCT characterization before and after failure reveal that killer defects are primarily surface
valleys. These drastically reduced the fatigue resistance of the material, evident when
comparing against the fatigue resistance of vertically built, machined and manually polished
(M&P) specimens. Fatigue strength at 2 ˆ 106 cycles for as-built specimens, regardless of
build orientation, is found close to 250 MPa. In contrast, M&P specimens reached a fatigue
strength of roughly 850 MPa. There is a small yet notable diﬀerence in the fatigue life
between vertically built and 45° specimens. The latter display shorter fatigue lives, likely
due to the increased roughness in their down-skin area, which consistently was the site of
failure.

Another part of the study concerned specimens subjected to a more intense heat treatment
(4 h at 860 °C). This focused on the eﬀects of the α-case, a brittle oxygen-contaminated
surface layer forming during heat treatments. In our case, the treatment yielded a 70 µm
layer. This layer further compromised the fatigue resistance of as-built specimens, though
possibly not as seriously as surface roughness. The diminished fatigue life could be due to
the α-case’s tendency to promote crack initiation.

Plasma Electrolytic Polishing (PeP), a recent surface polishing method, was tested as a
solution to mitigate both roughness and α-case in as-built surfaces. While some specimens
experienced uneven material removal or unexplained polishing failures, the process effec-
tively minimized initial surface roughness. Because material removal during a 40 min PeP
treatment was insufficient (about 180 µm), sub-surface pores were brought to the surface,
leading to the formation of surface pits. Consequently, cracks emerged from these exposed
pores, limiting the improvement in fatigue resistance. The fatigue resistance at 2 ˆ 106 cycles
after such treatment improved from 250 MPa to approximately 400 to 550 MPa. A more
prolonged PeP treatment (1 h 20 min) polished beyond this porosity overconcentration layer,
significantly enhancing fatigue resistance. Thus, the fatigue strength at 2 ˆ 106 cycles for
these specimens was larger than 1000 MPa, slightly surpassing M&P samples.

While PeP oﬀers enhanced fatigue properties, it may not be feasible for many practical
scenarios. Factors such as complex part shapes or prohibitive post-processing costs and time
considerations might hinder its adoption. Hence, understanding, modeling, and predicting
the inﬂuence of as-built surface defects on fatigue properties remain crucial. Traditional
?
models, like the Murakami parameter area, fall short when addressing the complexities
of surface valleys leading to failure. Thus, the search for alternative methods is explored in
Chapter 5.

152 | Chapter 4 – Fatigue properties and crack initiation mechanisms of as-built and
post-treated specimens
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Chapter 5

Quantitative analysis of surface

defects harmfulness

Contents
5.1 Surface defects segmentation . . . . . . . . . . . . . . . . . . . . . . . . . 154

5.1.1 Detailed methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

5.1.2 Segmentation of valleys and pit-like defects . . . . . . . . . . . . . . . . . . 158

5.1.3 Segmentation of protrusions/peaks . . . . . . . . . . . . . . . . . . . . . . . 160

5.2 Relationships between fatigue resistance and parameters derived from

the surface characterization using XCT . . . . . . . . . . . . . . . . . . . 160

5.2.1 Parameters selected to reﬂect surface defects harmfulness . . . . . . . . . . 161

5.2.2 Correlation between the selected parameters and the fatigue lifetime . . . . 163

5.3 Killer defect prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

5.3.1 Application to surface-treated specimens (PeP) . . . . . . . . . . . . . . . . 169

5.3.2 Application to specimens with as-built surfaces . . . . . . . . . . . . . . . . 178

Intermediate summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Chapter 4 primarily focused on the description of fatigue properties and defects at the origin
of failure. This chapter aims at rationalizing the impact of surface defects on fatigue, based
on the analysis of XCT data. Specimens in their as-built state and those polished by PeP will
be analyzed. Specimens subjected to a heat treatment at 860 °C have been excluded from our
analysis to alleviate the effect of α-case formation or changes in microstructure.
The first section introduces the methodology employed to segment surface defects, i.e. to
identify them. In other words, this step consists of going from a description of the surface as a
scalar field (= values such as Kt˚ at each surface voxel) to a defect-based description (= a single
Kt˚ value associated with each segmented defect). This defect segmentation will be used in the
subsequent sections.

| 153

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The second section is dedicated to the study of the relationship between parameters measured
from XCT data (e.g. Kt˚ ) and the experimentally measured fatigue resistance of samples with
diﬀerent surface conditions (as-built Z and 45°, PeP...). The aim is to evaluate whether the
parameters measured from XCT such as Kt˚ can give quantitative and reliable information
about the fatigue lifetime at a given stress level, or about the maximum stress level that can be
reached for a targeted fatigue lifetime.
Finally, the third section evaluates the ability of the methodology to predict the location
of killer defects, similarly to what has been shown for chemically etched E-PBF specimens in
Chapter 2, in Section 2.4.

5.1 Surface defects segmentation

The main objective of this section is to provide a tool to segment surface defects that
are likely to initiate cracks during fatigue loading. It concerns mainly valleys in as-built
specimens and pit-like defects (= pores brought to the surface) in specimens polished by PeP.
The developed methodology has also been used to segment protrusions/peaks, i.e. the highest
surface features. Although they are unlikely to initiate a crack, their segmentation is required
for some calculations in the subsequent sections. This explains why this topic is addressed here.
Fundamentally, the same method has been used to segment all kinds of surface defects in
specimens with diﬀerent surface states (vertical or 45° specimens, before or after PeP). However,
the method requires to adjust several parameters to take into account the particularities of each
case.
Identifying pit-like defects in PeP-polished specimens and protrusions in as-built specimens
turned out to be relatively straightforward. These defects are usually well-identiﬁed. How-
ever, isolating valleys in as-built specimens is more challenging due to several reasons discussed
hereafter:

• The very elongated (and sometimes tortuous) shape of valleys.

• The fact that valleys are not well separated from each other. They often form a sort of
valley network that can make it diﬃcult to deﬁne separated defects, especially in down-skin
regions in 45° specimens.
• The resolution of XCT scans is often too low to accurately capture the valleys’ depth
and morphology. This likely leads to underestimating their depth and curvature, further
complicating their separation from the rest of the surface. Moreover, the underestimation
of valley depth and curvature is not necessarily homogeneous along valleys. As a result,
a single valley can appear from the XCT scan as several fragments of valleys that are
separated because the resolution was not high enough to properly capture the entire valley.
Finally, the limited resolution makes results (especially curvature) noisy, which further
hinders the segmentation of such defects.

The segmentation method was chosen to best overcome these difficulties. Its different steps
are detailed hereafter for an as-built (45°) specimen. Examples are then shown for the segmen-
tation of potential killer defects (i.e. valleys or pit-like defects) for all surface states, and finally,
other examples are given for the segmentation of protrusions.

154 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
5.1.1 Detailed methodology
The defect segmentation is achieved in 3 steps:
1. Thresholding of the defects points (i.e. surface points that are considered to belong to a
defect) based on their curvature or height value.
2. Clustering of these points to deﬁne defects.
3. Cleaning the identiﬁed defects by removing the smallest ones.

As shown in Section 4.1.3.2, valleys are most easily identified based on their low κmin or κσ
values (the sign being taken into account). Thus, defect points are first separated from the rest
of the surface by thresholding, based on the local curvature.
While κmin successfully reveals valleys, it does not seem suited for their segmentation since
it does not allow to differentiate those that are parallel to the specimens’ axis from those that
are perpendicular to it. Therefore, using κmin would make it very difficult to segment individual
valleys, as many of the valleys perpendicular to the specimen axis would be connected by valleys
that are parallel to the specimen axis. This would lead to the segmentation of valley networks,
especially in down-skin regions of 45° specimens, see Fig.5.2c.
To avoid this problem, the curvature oriented in the direction of the specimens’ axis, κσ , is
used instead of κmin . A drawback of this method is that when computed at a fine scale, κσ
is more noisy than κmin . For this reason, the curvature is computed here at a relatively large
scale, i.e. using rcurv “ 50 µm.
The value for the threshold is determined automatically based on the histogram of the κσ
values. The method is similar to that of the triangle threshold for bimodal histograms
(TTBH) introduced in Section 2.2.1. The definition of the threshold is schematically illustrated
in Fig.5.1.
As for the TTBH, the value of the threshold is determined by maximizing the distance d
between the hypotenuse of the triangle (orange dashed lines in Fig.5.1) and the normalized
histogram. This enables the separation of the main histogram peak and the left tail, which
contains points with the lowest κσ , i.e. defect points. However, the obtained threshold value is
not always the most suitable. In some cases, slightly increasing or decreasing the threshold value
allows to achieve better results. To allow for such flexibility, an offset parameter is included and
can be set between -1 and 1. This modifies the final threshold value as schematized in Fig.5.1.
Fig.5.2a-f show the surface before and after thresholding of a 45° specimens (offset “ `0.4), in
both down-skin and up-skin regions. The surface is represented in grayscale using the minimum
curvature computed at a rather low scale (rcurv “ 20 µm) to provide a detailed visualization
of the surface. Fig.5.2c shows the complex network of interconnected valleys that would be
identified if κmin was used instead of κσ . Contrariwise, the use of κσ breaks this network
and results in separated defects, see Fig.5.2d,f.
A clustering algorithm called DBSCAN is then used to gather defect points together, and
thus obtain a set of isolated and well-identified defects. Details on how this algorithm works
can be found in [EST 96]. Two parameters are required in this model: Eps and MinPts. The
meaning of the MinPts will not be discussed extensively here, since its value has always been set
to 1 in our case. With MinPts“ 1, the DBSCAN algorithm simply clusters together all points
that are separated by a distance shorter than Eps.

Surface defects segmentation | 155

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

dmax
+oﬀset
-1 0 0.4 1

Normalized
Threshold curvature
-1 Decreasing κ 0 Increasing κ 1
Figure 5.1 | Deﬁnition of the method for the automatic determination of the thresh-
old for valleys thresholding, based on κσ values. The threshold is computed using the
normalized histogram computed for κσ prcurv “ 50 µmq by KDE.

A small adjustment is introduced here to further improve the ability of the model to cluster
elongated valleys that are perpendicular to the specimen axis, without connecting two parallel
transverse valleys that are close to one another. To do this, all points are dilated along the
specimen axis by a factor aspect_ratio, e.g. equal to 5 for as-built 45° specimens. In other
words, the component of each point along the specimen axis (i.e. the Z component for vertical
specimens) is multiplied by aspect_ratio, which artificially increases the distance between points
along the specimen axis. This prevents to some extent the algorithm from merging two
parallel valleys that are very close to each other at some point.
After clustering, many of the identified clusters are small in size. This leads to high numbers
of clusters that are not necessarily significant. For instance, 2949 clusters were found for the
specimen shown in Fig.5.2. To reduce this number and keep only the most significant clusters,
only the ones having a size greater than min_cluster_size were kept. The minimum cluster size
was set at 500 for as-built 45° specimens, which resulted in a final number of clusters of 692.

156 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.2 | Identiﬁcation of surface defects (valleys) by thresholding the local cur-
vature. Example based on the surface of an as-built 45° specimen. (a) Downskin
area represented by κmin prcurv “ 20 µm|λS “ 0.03 mmq. (b) Color code for all other subﬁgures.
(c) Thresholded points (in orange) using as threshold κmin prcurv “ 50 µmq “ ´0.0071 µm´1 .
(d) Thresholded points (in orange) in down-skin using as threshold κσ prcurv “ 50 µmq “
´0.0071 µm´1 . (e) Upskin area represented by κmin prcurv “ 20 µm|λS “ 0.03 mmq. (f) Thresh-
olded points (in orange) in up-skin using as threshold κσ prcurv “ 50 µmq “ ´0.0071 µm´1 .

Surface defects segmentation | 157

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
5.1.2 Segmentation of valleys and pit-like defects
The parameters used for each surface state for the segmentation of valleys or pit-like defects
are given in Tab.5.1. They were all chosen manually.

as-built Z as-built 45° PeP

oﬀset +0.5 +0.4 -0.1
Eps 20 10 2
aspect_ratio 5 5 1
min_cluster_size 500 500 20

Table 5.1 | Parameters used for the main valleys segmentation in each surface state condition:
as-built Z or 45°, as well as specimens polished by PeP (all conditions included).

The result of the clustering for the as-built 45° specimen shown in Fig.5.2 is presented in
Fig.5.3a-b. Likewise, the results of the clustering for an as-built Z specimen and for a Z specimen
polished by PeP are shown in Fig.5.4a-b and Fig.5.5a-b respectively.

Figure 5.3 | Final valleys segmentation in (a) down-skin and (b) up-skin regions of
an as-built 45° specimens. A random color is attributed to each cluster.

Overall, the quality of the segmentation seems reasonable given the resolution of the 3D
images. However, we can see in Fig.5.3a-b and Fig.5.4a-b that a signiﬁcant proportion of
the valleys are not or only partially captured, a problem that could be circumvented by
using more advanced segmentation techniques and/or higher resolution XCT scans.

158 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.4 | Valleys segmentation on a Z specimen. (a) Surface before segmentation. (b)
Segmentation result. A random color is attributed to each cluster.

Figure 5.5 | Pit-like defects segmentation on a Z specimen polished by PeP (Z|Vertical

40min
condition). (a) Surface before segmentation. (b) Surface after segmentation. A random color
is attributed to each cluster.

Surface defects segmentation | 159

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Since protrusions have much larger width and height, and exhibit a more rounded morphology,
their segmentation is easier than that of valleys. The κσ parameter is not necessarily the most
suited in such a case. A more simple roughness measurement is contrariwise very well adapted.
Thus, the methodology presented for valley segmentation can simply be applied, using the
roughness measurement instead of κσ . Clusters will thus be formed from surface points that
have the highest height. To do that, the thresholding method illustrated in Fig.5.1 is adapted
to isolate the right tail of the height distribution (highest values) instead of the left tail.
Eps and aspect_ratio parameters were taken equal to 2 and 1 respectively, which roughly
means that only points in direct contact will be clustered together, without any preferential
direction. oﬀset and min_cluster_size parameters are taken equal to -0.25 and 1000 respectively,
for both as-built Z and 45° specimens. The segmentation of protrusions has not been performed
on specimens polished by PeP, since none is left after polishing.
As an example, the result of protrusions segmentation in the same area of the Z specimen
shown in Fig.5.4 is presented in Fig.5.6a-b. All the main protrusions are identiﬁed.

Figure 5.6 | Protrusions segmentation on a Z specimen. (a) Surface before segmentation.

(b) Surface after segmentation. A random color is attributed to each cluster.

5.2 Relationships between fatigue resistance and parameters de-

rived from the surface characterization using XCT

The aim of this section is to test the ability of diﬀerent parameters (e.g. the roughness
parameter Sa ), calculated for a given specimen, to inform about the impact of surface defects
on the fatigue resistance.

160 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
5.2.1 Parameters selected to reflect surface defects harmfulness
Several parameters possibly reflecting the surface defects’ harmfulness were evaluated, all
computed using roughness and curvature measurements based on XCT data. Roughness and
curvature were measured using λc “ 0.8 mm, rcurv “ 20 µm and λS “ 0.03 mm. All parameters
were computed over a height of 9 mm centered on the specimens’ gauge area. Here, all the
surface is used, i.e. measurements are not restricted to the down-skin area for 45°
specimens.
First, conventional Sa and Sv roughness parameters were used. An additional parameter
was evaluated, in an attempt to combine several roughness parameters in a more comprehensive
way than considering each roughness parameter alone. This parameter is based on a model
proposed by [ARO 99] that has been introduced in Section 1.4.2.2. The model derives an effective
stress concentration factor from roughness measurements, a value that should quantify the notch
effect generated by the surface roughness. This parameter is conventionally used to predict the
fatigue limit of specimens based on roughness data. However, here, it was used as an index
reflecting the impact of surface roughness on the fatigue resistance.
Since this model was defined for 2D roughness profile measurements, some adjustments were
made here to extend it in 3D. The obtained formula for the effective stress concentration
factor K̄tAR is given by the following equation:

Sz˚
ˆ ˙ˆ ˙
Sa
K̄tAR “1`2 ˚ (5.1)
ρ̄ S10z
Sa is the arithmetic mean roughness. Sz˚ and S10z ˚ deﬁnitions are close to the ones of standard

ISO parameters Sz and S10z [Int 21]. Still, it is slightly diﬀerent because the method used in
our case for valley segmentation (presented in the previous section) is diﬀerent from the one
advised in the ISO standard [Int 21]. Thus, in the present case, Sz˚ is the sum of the highest
peak height and the deepest valley depth, whereas S10z ˚ is the sum of the average depth of the

five deepest valleys and the average height of the five highest peaks. Finally, ρ̄ is the average
radius of curvature measured on the five deepest valleys. The radius of curvature is calculated
here as the inverse of the curvature along the loading direction. Valleys and peaks are segmented
using the method presented in the previous section. When less than 5 valleys or 5 peaks are
detected, the average is computed from the few detected ones.
The last parameter that is evaluated is the maximum value of the estimated stress
concentration factor Kt˚ measured on a whole specimen, denoted Kt,max ˚ . A slight
˚
modification was made compared to the definition of Kt given in Section 2.2.4. This modification
aims to take into account the variation of the section along its axis. Since the specimen
section is larger within a fillet radius, the stress experienced by the material is reduced in this
area. Kt˚ is thus corrected at each point of the surface by the ratio of the minimum section
along the whole specimen (i.e. within the gauge length) Smin and the section S at this point.
The resulting formula for Kt˚ is given in Eq.5.2:

Smin ´ a ¯
Kt˚ “ ¨ 1 ` 2 height ¨ κσ (5.2)
S
Smin
Fig.5.7a-d show the Kt˚ values before and after correction, as well as the S factor, measured
on a 45° specimen.

Relationships between fatigue resistance and parameters derived from the surface
characterization using XCT | 161
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.7 | Correction of the Kt˚ parameter to take into account the variation of
the section along the specimens’ axis. (a) Kt˚ map before correction, (b) Smin S correction
factor, (c) Kt˚ map after correction, and (d) Internal view of the Kt˚ map after correction.

162 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Note that theoretically, the stress concentration at the intersection of the gauge area and the
ﬁllet radii (which is diﬀerent from the simple impact of section variation mentioned previously)
could also be taken into account. However, FEM calculations were conducted from the spec-
imens’ CAD geometry and the measured stress concentration reached a maximum of 1.03. It
was considered negligible here and was thus not taken into account.

5.2.2 Correlation between the selected parameters and the fatigue lifetime
To evaluate the correlation between the four parameters deﬁned previously (Sa , Sv , K̄tAR
˚
and Kt,max ) and the fatigue resistance, S-N diagrams are created for each parameter, with data
points color-coded according to the parameter value. Figure 5.8a provides an ideal example
of such a diagram, assuming a strong correlation between the investigated parameter and the
fatigue resistance.

b
a

+
Nf ≈ cste

+ σmax
Parameter (harmfulness)

Parameter +
c
σmax

σmax = cste
+

Nf +
Nf

Parameter +
Figure 5.8 | Schematic representation of an ideal correlation between fatigue re-
sistance and a given parameter representative of surface defects harmfulness. (a)
Hypothetical S-N diagram that would be obtained if the fatigue lifetime of specimens was di-
rectly correlated to the harmfulness parameter. (b) Plot of σmax as a function of the harmfulness
parameter, for specimens that failed after a given number of cycles. (c) Plot of Nf as a function
of the harmfulness parameter, at a given maximum stress level.

In Fig.5.8a, specimens with defects that are not severe (low harmfulness, reﬂected by the
selected parameter) tend to have longer lifespans or endure higher maximum stresses. This
uniﬁed graph representation allows for a comprehensive overview.

Relationships between fatigue resistance and parameters derived from the surface
characterization using XCT | 163
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To assess possible correlations, the detailed plots in Fig.5.8b-c are helpful. These graphs
plot the maximum stress σmax or the number of cycles to failure Nf against the harmfulness
parameter. A strong correlation in these plots indicates the parameter’s reliability in predicting
fatigue resistance. However, caution is needed when generalizing these results, as they require
plotting data for a single σmax or Nf value only.
Plots are first shown for Sa in Fig.5.9a-c. All surface states (e.g. as-built and PeP) are
represented on the same plot. This enables to evaluate whether Sa is relevant to reflect defects
harmfulness for very different surfaces.

700 b
a
Sa (µm) 600 Nf ∈ [5e4, 7e4]

σmax (MPa)
1200
500
10
1000 400

8 300
σmax (MPa)

800 2 6 10
Sa (µm)
6 3·105 c
600
σmax = 350 MPa
2·105
4
Nf

400
2 105

200
104 105 Nf 106 107 6·104
7 9 11
Sa (µm)

Figure 5.9 | Correlation between Sa and the fatigue resistance of specimens with
diﬀerent surface states. (a) S-N curve with points coloring according to the specimens’
Sa . (b) σmax vs. Sa plot for point with Nf P r5 ˆ 105 , 7 ˆ 104 s. (c) Nf vs. Sa plot for
σmax “ 350 MPa.

The S-N diagram in Fig.5.9a shows that some correlation exists between Sa and the fatigue
resistance. Specifically, specimens with a lower Sa , such as those polished by PeP, tend to have
longer fatigue lives and withstand higher stresses.
Additionally, the plot of fatigue life (Nf ) against Sa in Fig.5.9c shows a fairly good correlation
at 350 MPa. Sa effectively captures the fact that, on average, as-built Z specimens
reach higher fatigue lives than as-built 45° specimens which exhibit higher Sa values.
However, there is a noticeable scatter: for example, some Z specimens have similar fatigue
lives in comparison with 45° specimens, despite having significantly different Sa values. The
quality of the correlation also depends on the stress levels. For instance, at 300 MPa, an as-built

164 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
45° specimen failed after 193 381 cycles with Sa “ 6.4 µm, while an as-built Z specimen failed
after only 125 534 cycles while Sa “ 11.0 µm. Therefore, while a correlation between Nf and
Sa is observed, especially for as-built surfaces, it should be interpreted cautiously due to the
observed scatter and limited number of specimens.
The σmax vs. Sa graph in Fig.5.9b shows a stronger correlation of Sa with σmax . This plot
is based on specimens with fatigue lives between 5 ˆ 104 and 7 ˆ 104 cycles. However, all PeP-
polished specimens show similar Sa values, even though they sustained diﬀerent stress levels.
This trend is also noticeable in Fig.5.9a, where PeP-polished specimens are distinctly separated
in dark blue (indicating low Sa ) from as-built specimens in green/yellow (higher Sa ). Hence, Sa
might not be the best parameter for polished surfaces which are smooth but may contain a few
harmful defects.
In such cases, Sv , which better represents the most critical defect across the entire surface,
might be more appropriate. Corresponding plots for Sv are presented in Fig.5.10a-c.

700 b
a Sv (µm)
1200 600 Nf ∈ [5e4, 7e4]

σmax (MPa)
160
500
1000
400
120
σmax (MPa)

800 300
20 40 60 80
Sv (µm)
80
600 3·105 c
2·105
σmax = 350 MPa
400 40
Nf

105
200 5 7
104 10 Nf 10 6
10
6·104
60 100 140
Sv (µm)

Figure 5.10 | Correlation between Sv and the fatigue resistance of specimens with
diﬀerent surface states. (a) S-N curve with points coloring according to the specimens’ Sv . (b)
σmax vs. Sv plot for point with Nf P r5 ˆ 105 , 7 ˆ 104 s. (c) Nf vs. Sv plot for σmax “ 350 MPa.

In Fig.5.10b, the use of Sv avoids the saturation eﬀect seen with Sa for PeP-polished spec-
imens. However, the correlation of Sv with fatigue life (Nf ) is not as good as that
of Sa . Notably, for as-built 45° specimens at 350 MPa, the highest fatigue lives paradoxically
correspond to the higher Sv values. Thus, while Sv generally shows that as-built Z specimens
have higher fatigue lives and lower Sv values on average than as-built 45° specimens, it fails to
adequately explain the scatter in fatigue lives among specimens with similar surface states.

Relationships between fatigue resistance and parameters derived from the surface
characterization using XCT | 165
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
700 b
a 600
Nf ∈ [5e4, 7e4]

σmax (MPa)
1200
3.00 500

1000 2.75
400
2.50
300
σmax (MPa)

800 1.2 1.8 2.4

2.25

2.00
600
3·105 c
1.75 σmax = 350 MPa
2·105
1.50
400

Nf
1.25 105
200
104 105 Nf 106 107
6·104
1.8 2.4 3.0

Figure 5.11 | Correlation between K̄tAR and the fatigue resistance of specimens with
diﬀerent surface states. (a) S-N curve with points coloring according to the specimens’
K̄tAR . (b) σmax vs. K̄tAR plot for point with Nf P r5 ˆ 105 , 7 ˆ 104 s. (c) Nf vs. K̄tAR plot for
σmax “ 350 MPa.

Fig.5.11a-c presents results for K̄tAR which lie between those of Sa and Sv , leaning more
towards Sa . This was somewhat expected, since K̄tAR is derived from Sa and Sz˚ (Sz˚ is the
average of the deepest valley depth and the highest peak height and is thus close to Sv ). The
contribution of curvature (at least with a convolution radius of 20 µm) appears to
be limited in this case.
˚
Fig.5.12a-c displays plots for Kt,max , showing trends similar to those for Sv . This similarity
˚
is likely due to the fact that Kt,max , like Sv , is a parameter that mainly characterizes the most
˚
critical defect. Thus, the correlation of Kt,max with Nf at 350 MPa (Fig.5.12c) is not very good,
despite the clear distinction between the average behaviors of as-built Z and 45° specimens.
In contrast, the correlation between σmax and Kt,max ˚ in Fig.5.12b is much more
pronounced and effectively differentiates between varying surface states, i.e. for both
˚
as-built and PeP-polished specimens. It could be expected that a parameter such as Kt,max
˚
yields a better correlation with σmax than Nf , since Kt,max quantifies a stress concentration
˚
effect. Fig.5.12a also shows that Kt,max allows to better account for the transition between
as-built and PeP specimens, without the distinct separation between the two as was noticed
with Sa .

166 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

σmax (MPa)
1200
7 500

1000 6 400

300
σmax (MPa)

800 5 3 4 5 6

4 3·105 c
600
2·105
σmax = 350 MPa
3
400

Nf
2 105

200
104 105 Nf 106 107 6·104
4.5 5.5 6.5

Figure 5.12 | Correlation between Kt,max ˚ and the fatigue resistance of specimens
with diﬀerent surface states. (a) S-N curve with points coloring according to the specimens’
˚
Kt,max ˚
. (b) σmax vs. Kt,max plot for point with Nf P r5 ˆ 105 , 7 ˆ 104 s. (c) Nf vs. Kt,max
˚ plot
for σmax “ 350 MPa.

A possible solution to improve these results, especially the correlation between parameters
and Nf , would be to consider the shape of surface defects, particularly for as-built specimens. For
˚
example, Kt,max being a local measurement, it does not account for the lengths of surface valleys.
Yet, as discussed in Section 4.1.3.2, the elongated nature of valleys might be a critical factor in
fatigue failure. To estimate whether taking this factor into account could improve predictions,
the length of the crack initiation zone (ℓCIZ ) was manually measured on the fracture surfaces of
as-built specimens using SEM. ℓCIZ represents the length of the surface perimeter from which
river lines appear to originate. Assessing whether including valley length improves the predictive
˚
accuracy of Kt,max could be an interesting pathway in the future.
Fig.5.13a demonstrates that at higher maximum stress levels, the values of ℓCIZ (length
of the crack initiation zone) tend to be higher. This trend is likely because increased stress
levels promote the initiation of multiple cracks [SCH 09, Fig.2.20], leading to larger ℓCIZ values.
Additionally, there is a noticeable trend because specimens with smaller crack initiation zones
achieve longer fatigue lives at a given stress level, as shown in Fig.5.13b. This observation
suggests that considering the length of valleys might be essential for a more accurate
prediction of their impact on fatigue resistance.

Relationships between fatigue resistance and parameters derived from the surface
characterization using XCT | 167
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
a (mm)
600 >4
Z|as-built b
550 3.5 3·10 5
45°|as-built
500 3 2·105 σmax = 350 MPa
σmax (MPa)

Nf
450 2.5
105
400 2
6·104
350 1.5
0.5 1.5 2.5
300 1 (mm)

Z|as-built 45°|as-built
250 < 0.5
5 7
104 10 Nf 10 6
10

Figure 5.13 | Correlation between ℓCIZ and the fatigue resistance of as-built spec-
imens. (a) S-N curve with points coloring according to ℓCIZ . (b) Nf vs. ℓCIZ plot for
σmax “ 350 MPa.

5.3 Killer defect prediction

Roughness, curvature and Kt˚ measurements are now used for the prediction of killer defects.
As in the previous section, the analysis has been conducted solely on as-built specimens that
were subjected to a stress relief (no high-temperature heat treatment) as well as specimens
polished by PeP.
The methodology followed here is similar to the one used in the study conducted on chemically
polished E-PBF specimens, see Section 2.4. The main difference is that the defects segmentation
presented in Section 5.1.1 has been applied to achieve more meaningful results for as-built
specimens.
Each specimen was scanned by XCT before and after failure. The most harmful defects
(valleys for as-built specimens and pit-like defects for PeP specimens) were automatically seg-
mented from the initial scan before fatigue testing using the methodology described in Section
5.1.1. Among them, the killer(s) defect(s) were manually identified from the (SEM and XCT)
fracture surface observation as detailed in Section 4.1.3.2. Note that if several defects were
thought to play a significant role in fatigue failure (i.e. in the case of multiple crack initiations),
those defects were merged and considered as one for the subsequent analysis. For example, the
maximum Kt˚ value of the killer defect will be considered to be the maximum value among all
defects that lead to the initiation of a crack of significant size. For some specimens, the killer
defect could not be identified due to the insufficient resolution of XCT scans.
Roughness, κσ , and Kt˚ were calculated using λc “ 0.8 mm, rcurv P t10, 20, 50, 100, 150uµm
and λS “ 0.03 mm. Note that Kt˚ has been calculated using the section correction presented
in Section 5.2. For each defect segmented from the initial XCT scan, the most critical value is
computed for each parameter. This value corresponds to the highest one for the Kt˚ parameter
and to the lowest one for roughness and κσ .

168 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
At this step of the analysis, we have, for each specimen, a set of defects automatically
segmented from the 3D image. Each defect is characterized by a single value of depth, as well
as a value of κσ and Kt˚ for five different values of rcurv (hence 1 ` p2 ¨ 5q “ 11 values in total
for a given defect). Among these defects, we also know which ones were responsible for failure
in experimental fatigue tests based on the fracture surface observation using SEM.
Two scores were then calculated to inform about the ability of the model to predict the killer
defect. The first one is the rank, which corresponds to the number of defects initially segmented
from the XCT scan that is expected to have a more detrimental effect on the fatigue lifetime
compared to the killer defect(s). A rank of 1 means that the killer defect is well-predicted with
our approach. A rank of 10 means that 9 defects were predicted to be more harmful than the
actual killer defect(s). The second score is a percentage that corresponds to the ratio of the
maximum Kt˚ value measured at the killer defect(s) and the maximum Kt˚ value measured all
over the characterized surface. This percentage score has been calculated for the Kt˚ parameter
only, since it is the one showing the most physical meaning.
The results will be first shown for specimens polished by PeP. It is the most advantageous
situation since defects can be segmented more easily than valleys in as-built specimens. This
case is similar to that of chemically polished E-PBF specimens presented in Section 2.4. In a
second part, results will then be shown for as-built specimens.

5.3.1 Application to surface-treated specimens (PeP)

5.3.1.1 Ranks and percentage
Fig.5.14 shows the inﬂuence of rcurv on the median rank for κσ and Kt˚ . The median rank
is calculated from the ranks obtained for all specimens when the methodology could be applied
(i.e. when the killer defect could eﬀectively be captured by XCT).

6 6
Sv values of
5 PeP specimens 5
Median rank - κσ

Median rank - Kt*

4 4

3 3

2 2

1 1

0 0
0 25 50 75 100 125 150
rcurv (µm)
Figure 5.14 | Median ranks obtained for κσ and Kt˚ for the diﬀerent values tested
for rcurv . The range of Sv values obtained for PeP specimens (except the Z|Vertical
1h20 condition) is
given in green.

Killer defect prediction | 169

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
For κσ , the median rank decreases when rcurv increases and reaches a plateau between 20 µm
and 50 µm. This is consistent with the observation made in Section 2.4, where the optimum
rcurv value was found to be close to the maximum valley depth Sv . Indeed, Sv values lie between
20 µm and 50 µm for specimens polished by PeP, except Z|Vertical
1h20 ones which could anyway not
be part of this analysis since the killer defect could not be found in the XCT scan before failure.
Given these results, ranks of PeP specimens considering the κσ metric will be studied using
rcurv “ 50 µm.
For specimens polished by PeP, rcurv does not have a strong inﬂuence on the ranks obtained
based on the Kt˚ parameter. Slightly better results are yet obtained for rcurv “ 20 µm, which is
why ranks and percentage scores for Kt˚ will be presented using this value.
Fig.5.15 provides a visual representation to help understand rank and percentage scores. It
shows in blue the Probability Density Function (PDF « continuous histogram) of Kt˚ values
obtained for all surface points where this parameter is deﬁned (i.e. where height ă 0 and
κσ ă 0). Most surface points exhibit Kt˚ values below 1.5. The distribution extends into a very
long tail, for a large part representing points within pit-like defects. Thus, the orange points,
corresponding to the maximum Kt˚ value found among all points of each defect, are found in
this tail section of the distribution.

Probability density
(surface points Kt*) PeP

3 PDF (≈ histogram)
of Kt* values of surface points

Max Kt* values

1
of segmented defect
1 point = 1 defect

0 0
Probability density
PDF of max Kt values*
0.5 (defects max Kt*)
of segmented defects

1 2 3 4

Figure 5.15 | Distributions of surface points and defects Kt˚ values for 45°|Vertical
40min -2
˚
specimen. The blue line shows the PDF for Kt values of surface points, considering only
points with negative heights and κσ . The orange points and curve represent the maximum Kt˚
value for each segmented defect and its PDF. The killer defect is marked in purple. PDFs are
calculated using the KDE method with Silverman’s rule of thumb for bandwidth selection.

170 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Additionally, an orange PDF illustrates the maximum Kt˚ values distribution for the seg-
mented defects, and the killer defect is distinctly marked in purple, with its rank/percentage
scores. In the given example, two defects exhibit higher maximum Kt˚ values than the killer
defect, placing its rank at 3. For a detailed description of the results, speciﬁc values for each
specimen are summarized in Tab.5.2.

PeP Rank Percentage

Height κσ Kt˚ Kt˚
λc “ 800 µm λc “ 800 µm
Specimen ndef ects λc “ 800 µm rcurv “ 50 µm
rcurv “ 20 µm rcurv “ 20 µm
Z|Vertical
40min - 1 29 7 9 8 75
Z|Vertical
40min - 2 41 1 1 1 100
Z|Vertical
40min - 3 12 2 2 1 100
Z|Vertical
40min - 4 10 5 4 5 74
Z|Vertical
40min - 5 21 1 1 1 100
Z|Vertical
1h20 -1 ˆ Killer defect not captured by XCT
Vertical
Z|1h20 - 2 ˆ Killer defect not captured by XCT
45°|Vertical
40min - 1 43 11 12 8 85
45°|Vertical
40min - 2 30 2 2 3 86
45°|Vertical
40min - 3 ˆ Killer defect not captured by XCT
45°|45°
40min - 1 11 1 1 1 100
45°|45°
40min - 2 ˆ Killer defect not captured by XCT
45°|45°
40min - 3 ˆ Killer defect not captured by XCT
45°|45°
40min - 4 ˆ Killer defect not captured by XCT
Median 25 2 2 1 93

Table 5.2 | Killer defect predictions scores (rank & percentage) for specimens pol-
ished by PeP using roughness, κσ and Kt˚ parameters. The Specimen column refers to
the name of individual specimens and the ndef ects column gives the total number of defects that
were automatically segmented. Colors are given for guidance, the color tending to green for a
successful killer defect prediction while it tends to red for a less accurate prediction.

First, we can notice that for 5 specimens out of 14, the killer defect could not
be captured by XCT due to a lack of resolution. Therefore, the analysis could not
be conducted for those specimens. For the remaining specimens, the ranks obtained using
roughness, κσ or Kt˚ are all very similar. This is likely due to the small and point-like shape of
the defects in PeP specimens. Notably, their point-like shape reduces the beneﬁt of using the
curvature oriented in the direction of the specimen axis.
Among the 8 specimens that could be analyzed, the killer defect could be suc-
cessfully predicted in 4 cases using the Kt˚ parameter. These results are comparable to or
even slightly better than the ones obtained on chemically polished E-PBF samples in Section
2.4. This may be because the surface obtained for L-PBF polished by PeP is smoother and the
few remaining defects show a more regular/globular morphology.

Killer defect prediction | 171

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
For specimens that showed higher ranks (e.g. Z|Vertical Vertical
40min - 1 and 45°|40min - 1), the bad scores
could be at least partly explained by other factors such as microstructure are not taken into
account. For example, crack formation has been reported to be more frequent along basal and
prismatic planes in α phase of Ti64 alloy [BRI 08], which means that the local crystallographic
orientation of surface grains can be expected to play a role. The percentage scores for Z|Vertical
40min -
Vertical
1 and 45°|40min - 1 specimens being relatively high (75% and 85% respectively), crystallographic
orientation inhomogeneity may have caused crack initiation to occur at a defect that does not
generate the highest stress concentration.
Nonetheless, it is also very likely that a lack of resolution also plays a major role as already
emphasized. This is evidenced by the fact that the killer defect could not even be found (even by
visual inspection) on the surface extracted from the XCT scan for 5 specimens out of 14. Since
the resolution of XCT scans seems to be a major issue regarding critical defects’ characterization,
a more detailed analysis has been conducted on a few specimens.

5.3.1.2 The impact of XCT scan resolution on the prediction of killer defect
The objective here is to compare the killer defect as characterized by XCT with its charac-
terization with SEM. Since SEM provides a better resolution than XCT, it is considered as a
ground truth. This is possible for specimens polished by PeP since defects can be unambiguously
delineated on SEM observations, which is not possible for valleys in as-built specimens.
Four cases will be successively discussed:
• a case where the killer defect is eﬀectively captured by XCT, leading to a successful pre-
diction (example 1);
• a case showing that the method can also be used to account for near-surface internal pores
(example 2);
• a case where the limited XCT scan resolution enables to only partially capture the killer
defect, leading to a much less accurate prediction (example 3);
• a case where the killer defect could not even be found in the surface extracted from XCT
scans (example 4).

Example 1: successful prediction

Fig.5.16a-e show a first example where the killer defect could be successfully identified on
the surface extracted from the XCT scan. The defect is, in this case, a pore that was brought
to the surface by PeP, leading to the formation of a pit-like defect.
Fig.5.16c shows the killer defect on the 3D height map of the initial surface, but the external
point of view barely enables one to see it. The defect can be observed much more accurately when
looking from inside the sample, see Fig.5.16d-e. In Fig.5.16e, one can recognize the particular
shape of the defect that can be seen with more detail in the SEM image shown in Fig.5.16a.
Moreover, the dimensions of the defect are in good agreement with the ones found using SEM. As
a result of this relatively good defect segmentation, the killer defect is successfully
identified (rank “ 1) regardless of the parameter used (height, κσ , or Kt˚ ). However, even
for such a large defect, it can be seen that due to the limited XCT scan resolution, the 3D
morphology appears smoother/rounder than the one visible on the fracture surface observed by
SEM.

172 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.16 | Crack initiation site identiﬁcation on the 45°|45° 40min -1 specimen. The
specimen failed after 50 780 cycles at 550 MPa. The killer defect is a sub-surface pore that was
brought to the surface during PeP. (a) SEM-SE observation of the initiation site on the fracture
surface. (b) Global 3D height map of the specimen before fatigue testing. (c) Zoom on the killer
defect. (d) Same area as in (c), observed from the inside, revealing the morphology of the killer
defect that cannot be seen from the outside. (e) Killer defect observed from above, showing a
similar geometry as in the SEM observation.

Killer defect prediction | 173

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
It is worth noting here that the killer defect could be successfully segmented from
the XCT scan thanks to the use of the Triangle Threshold for Bimodal Histograms
(TTBH). As shown in the transverse slices of the XCT scan containing the defect in Fig.5.17a-
c, a conventional threshold obtained from Otsu’s method fails to capture this defect due to its
limited width.

a b c
Killer defect Otsu (124/255) TTBH (195/255)

30 µm Specimen axis

Figure 5.17 | Comparison of the abilities of Otsu’s threshold and TTBH to capture
the killer defect shown in Fig.5.16. (a) Transverse XCT slice (voxel size = 3 µm). (b)
Application of Otsu’s threshold, which does not capture the defect. (c) Successful segmentation
of the killer defect using the TTBH.

Example 2: prediction for a sub-surface pore

Another example of the beneﬁts of using the TTBH is illustrated by the second example
shown in Fig.5.18a-e. It shows the analysis of the crack initiation site on another 45°|45° 40min
specimen.
The killer defect is a sub-surface pore that has no direct connection with the surface, although
the ligament that separates the two is very thin. Fig.5.18c shows that Otsu’s threshold does
partially segment the pore, but that the resulting defect is not connected to the surface. Since
only surface defects are taken into account in the methodology used here, such a defect will be
ignored in the analysis, and the killer defect would not be possible to identify. Conversely, the
use of a higher threshold (in terms of gray level) erodes the surface and enables to erase
the ligament that separates the pore from the surface, see Fig.5.18d. It is advantageous
here, as a pore so close to the surface can reasonably be considered equivalent from a mechanical
point of view to the same defect located at the surface.
In this regard, the surface segmentation could be adapted to take into account all subsurface
pores and consider them as equivalent surface defects. A simple way would be to artiﬁcially
remove the ligaments that separate sub-surface pores from the surface, and then characterize the
obtained surface defect. Fig.5.18e shows an example, where the ligament is erased in the case
of Otsu’s threshold. Here, this operation was performed manually, but it could be automatized
fairly easily.
Finally, it can be seen that even if the defect can be segmented from XCT data using the
TTBH, its exact morphology cannot be extracted. This highlights, once again, that the use of
laboratory XCT with a voxel size of 3 µm reaches its limits, especially considering the small size
of the observed defects.

174 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

30 µm 30 µm
SEM-SE Specimen axis

c d
Otsu (126/255) TTBH (198/255)

Defect
not connected
to surface

Manual removal
of the ligament

Figure 5.18 | Comparison of the abilities of Otsu’s threshold and TTBH to capture
the killer defect on the 45°|Vertical
40min -2 specimen. The specimen failed after 1 173 759 cycles
at 550 MPa, and the killer defect is a sub-surface pore that was brought just beneath the surface
by the PeP treatment. (a) SEM-SE observation of the initiation site on the fracture surface. (b)
Transverse XCT slice (voxel size = 3 µm). (c) Application of Otsu’s threshold, which enables
capturing some part of the defect only. Since the defect is not connected to the surface, it would
not be considered a surface defect. (d) Successful segmentation of the defect using the TTBH.
The small ligament that separated the pore from the surface was removed by the TTBH. (e)
An alternative solution to take into account sub-surface pores in the present workﬂow dedicated
to surface defects analysis. The ligament that separates the pore from the surface is artiﬁcially
removed, revealing the underlying defect even when using Otsu’s threshold.

Killer defect prediction | 175

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Example 3: defect only partially captured due to an insufficient resolution
This is even clearer in the third example shown in Fig.5.19a-d, where the killer defect is
significantly blurred due to this lack of resolution. As a result, its depth is significantly un-
derestimated when it is measured from the XCT scan, even using the TTBH for the surface
segmentation, see Fig.5.19a,d. Its shape is not well described since it is only a few voxels large.
This will necessarily lead to erroneous estimations of the depth and curvature measurements.
This is probably the major reason accounting for the unsuccessful killer defect prediction in this
peculiar case (rank “ 8 ´ 12).

a Initiation site b
XCT

10 µm 30 µm
Specimen axis

c d
Otsu (124/255) TTBH (198/255)

Even using the TTBH, the resolution is insuﬀicient

to properly capture the defect
Figure 5.19 | Comparison of the killer defect observed by SEM and the segmentation
from the XCT scan on the 45°|Vertical40min -1 specimen. The specimen failed after 54 092 cycles
at 550 MPa, and the killer defect is a sub-surface pore that was brought to the surface by the PeP
treatment. (a) SEM-SE observation of the initiation site on the fracture surface. (b) Transverse
XCT slice (voxel size = 3 µm). (c) Surface segmentation using Otsu’s threshold. (d) Surface
segmentation using the TTBH.

176 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Example 4: defect not captured at all due to an insuﬃcient resolution
Finally, Fig.5.20a-d show the last case where the killer defect could not be captured at all
from XCT data at all because of insuﬃcient resolution.

a b
Initiation site XCT

SEM-SE

Specimen axis

10 µm 30 µm

c d
Otsu (124/255) TTBH (199/255)

The defect is Only a small portion

entirely missed of the defect is captured

Figure 5.20 | Unsuccessful segmentation of the killer defect on the 45°|45° 40min -3 spec-
6
imen due to a lack of resolution. The specimen failed after 2 ˆ 10 cycles at 550 MPa
followed by 52 369 cycles at 650 MPa. The killer defect is a subsurface pore that was brought to
the surface by PeP, although its connection to the surface is very small. (a) SEM-SE observa-
tion of the initiation site on the fracture surface. (b) Transverse XCT slice (voxel size = 3 µm).
(c) Application of Otsu’s threshold, which entirely misses the killer defect. (d) Application of
the TTBH which enables capturing some part of the defect only, and fails at establishing its
connection with the surface.

The SEM observation in Fig.5.20a shows that the killer defect is a pore that is barely con-
nected to the surface. Moreover, the portion of the defect close to the surface seems very thin,
making its characterization by XCT very diﬃcult. As a result, Fig.5.20b shows that only a
limited portion of the defect can be detected in the XCT scan. It corresponds to the bottom of
the defect, which is larger. This portion of the defect is successfully segmented using the TTBH
threshold, see Fig.5.20d. However, most of the defect is not captured, and in particular its con-
nection to the surface is not accounted for. Thus, the methodology for killer defect prediction
could not be applied.

Killer defect prediction | 177

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
In light of the four examples discussed above, the quality of the killer defect prediction
seems well correlated with the ability of XCT to properly capture defects. In other words, the
resolution of XCT scans plays a major role, and even a 3 µm voxel size is often insuﬃcient given
the small size of the defects. This resolution issue can also be expected to play a major role for
as-built specimens and the ability of XCT to properly capture the roots of surface valleys.

5.3.2 Application to specimens with as-built surfaces

For the sake of simplicity, the measurements of roughness, κσ , and Kt˚ for the as-built
specimens were conducted using the same parameters (λc , rcurv , and λS ) as those used for PeP-
polished specimens. Despite testing other values, no signiﬁcant improvements or changes in
interpretation were observed.
As-built specimens have markedly diﬀerent surface topologies compared to PeP specimens,
leading to distinct Kt˚ PDFs and a greater number of segmented defects. This is illustrated
in Fig.5.21, which shows for Z-7 specimen the surface points’ Kt˚ distribution in blue and the
maximum Kt˚ distribution for segmented defects in orange.

Probability density
(surface points Kt*) Z|as-built

1.0

0.5

0 0
Probability density
(defects max Kt*)

0.5

1 2 3 4

Figure 5.21 | Distributions of surface points and defects Kt˚ values for as-built Z-7
specimen. The blue line shows the PDF for Kt˚ values of surface points, considering only
points with negative heights and κσ . The orange points and curve represent the maximum Kt˚
value for each segmented defect and its PDF. The killer defect is marked in purple. PDFs are
calculated using the KDE method with Silverman’s rule of thumb for bandwidth selection.

178 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The surface points’ Kt˚ distribution has a less extended tail than that observed in PeP
specimens. While the killer defect appears in the tail of the surface points’ Kt˚ distribution, its
rank is quite high (176), which illustrates the more complex killer defect prediction for as-built
specimens. The rank and percentage scores for the killer defect predictions across all as-built Z
specimens are detailed in Tab.5.3.

Z Rank Percentage
Height κσ Kt˚ Kt˚
λc “ 800 µm λc “ 800 µm
Specimen ndef ects λc “ 800 µm rcurv “ 50 µm
rcurv “ 20 µm rcurv “ 20 µm
Z-1 725 323 234 215 51
Z-2 654 287 140 200 55
Z-3 710 181 198 187 67
Z-4 732 316 217 271 59
Z-5 752 26 11 1 100
Z-6 719 318 200 242 71
Z-7 739 301 87 176 67
Z-8 709 49 20 5 93
Z-9 730 53 31 28 82
Z-10 761 114 60 57 84
Median 728 234 114 182 69

Table 5.3 | Killer defect prediction scores (rank & percentage) for as-built Z speci-
mens using roughness, κσ and Kt˚ parameters. The Specimen column refers to the name
of individual specimens and the ndef ects column gives the total number of defects that were au-
tomatically segmented. Colors are given for guidance, the color tending to green for a successful
killer defect prediction while it tends to red for a failed prediction.

First, it can be noticed that the number of segmented defects is much higher for
as-built specimens (« 700) in comparison with specimens polished by PeP (ă 50). This is,
of course, linked to the much higher density of surface defects in as-built specimens.
A second observation is that ranks are much higher for as-built Z specimens compared to those
polished by PeP. Thus, for most as-built Z specimens, the method clearly fails to properly
predict the killer defect. In this context, it seems hazardous to try determining which
parameter among the roughness, κσ , and Kt˚ is the most relevant.
Nonetheless, results for as-built 45° specimens are much more encouraging than
those for as-built Z specimens. They are shown in the form of PDFs for one specimen in
Fig.5.22, followed by a more comprehensive analysis for all specimens in Tab.5.4.
As-built 45° specimens consistently achieve much lower ranks compared to as-built Z spec-
imens. For instance, the median rank for the Kt˚ parameter across 10 analyzed 45° specimens
is 21. The relatively better results obtained for the 45° specimens might be because defects in
45° specimens are larger. This may facilitate their identification using XCT with the resolution
employed here (voxel = 3 µm). The difference between Z and 45° specimens might also arise due
to the very different valley morphologies in either case.

Killer defect prediction | 179

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

1.0

0.5

0 0
Probability density
(defects max Kt*)

0.5

1 2 3 4 5

Figure 5.22 | Distributions of surface points and defects Kt˚ values for as-built 45°-7
specimen. The blue line shows the PDF for Kt˚ values of surface points, considering only
points with negative heights and κσ . The orange points and curve represent the maximum Kt˚
value for each segmented defect and its PDF. The killer defect is marked in purple. PDFs are
calculated using the KDE method with Silverman’s rule of thumb for bandwidth selection.

As for PeP specimens, it is also worth noting that although very high ranks are obtained for
some specimens, the percentage score can remain reasonably high. For instance, a rank
of 242 is obtained but a relatively high percentage of 71% is found for the Z-6 specimen.
Given all these observations, the following points might at least partially explain the inability
of the model to provide an accurate prediction of the killer defects in as-built specimens:
• The resolution of XCT scans is insufficient to properly capture surface defects.
• Due to the higher density of defects in as-built specimens, it seems logical that it is more
challenging to find the killer defect among «700 segmented defects than it was to find it
among «30 defects in the case of specimens polished by PeP.
• Other factors such as the microstructure may play a role and promote crack initiation at
spots where the stress concentration is not maximal.
• The model does not take into account the defects’ morphology. Yet, the particularly
elongated shape of surface valleys may play a significant role in crack initiation and/or
propagation. Properly taking into account the defects’ shape would probably require
improving the method used for defect segmentation together with the resolution of XCT
scans.

180 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
45° Rank Percentage
Height κσ Kt˚ Kt˚
λc “ 800 µm λc “ 800 µm
Specimen ndef ects λc “ 800 µm rcurv “ 50 µm
rcurv “ 20 µm rcurv “ 20 µm
45°-1 799 1 32 1 100
45°-2 788 12 30 3 93
45°-3 774 23 37 22 87
45°-4 706 97 2 78 64
45°-5 685 130 81 73 60
45°-6 670 85 68 51 68
45°-7 692 35 2 20 79
45°-8 766 22 15 9 77
45°-9 724 35 15 14 82
45°-10 655 58 5 41 75
Median 715 35 23 21 78

Table 5.4 | Killer defect prediction scores (rank & percentage) for as-built 45°
specimens using roughness, κσ and Kt˚ parameters. The Specimen column refers to the
name of individual specimens and the ndef ects column gives the total number of defects that
automatically were segmented. Colors are given for guidance, the color tending to green for a
successful killer defect prediction while it tends to red for a failed prediction.

Killer defect prediction | 181

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Regarding the issue of lack of resolution, it has already been discussed extensively for spec-
imens polished by PeP since it was possible in such a case to compare with the defect shape
observed by SEM. The task is much more difficult for as-built specimens because the killer
defects cannot be clearly identified on fracture surfaces.
Still, simply looking at the XCT scan and the surface extracted from it already shows that
resolution is a significant issue. For instance, Fig.5.23c-e present the segmentation for one of
the two killer defects found in the Z-3 specimen. Although the valley can be identified on
the 3D minimum curvature map in Fig.5.23a, Fig.5.23c reveals that the resolution is not good
enough to properly capture its root. In such a case, regardless of the threshold used for surface
segmentation, it leads to a significant smoothing of the valley root, inevitably altering the
roughness, curvature, and Kt˚ measurements. Thus, here again, improvements in the resolution
of XCT scans seem to be a key axis for improvement.

3D minimum curvature map b XCT radial slice

a
External view

κmin (µm-1)
< -0.09 -0.05 0 > 0.06

60 µm

c d Otsu (129/255) e TTBH (198/255)

30 µm

Figure 5.23 | Detailed inspection of the surface segmentation on a radial slice of

a killer defect in the Z-3 specimen. (a) Valley responsible for failure, observed on a 3D
minimum curvature map. (b) XCT radial slice at a given point of the valley, marked by a blue
cross in (a). (c) Enlarged view of (b) centered on the killer defect. (d) Segmentation of the killer
defect using Otsu’s threshold. (e) Segmentation of the defect using the TTBH.

182 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Intermediate summary
This chapter aimed at testing to which extent the roughness, curvature and Kt˚ measure-
ments are representative of the harmfulness of surface defects on fatigue properties. First,
a method to segment defects and in particular surface valleys from the surface extracted
from XCT scans was described. Although it can segment valleys with particularly elongated
shapes, valleys are often only partially captured. A major reason is the limited resolution of
XCT scans, which does not always enable full capture of them.

Qualitatively, roughness parameters such as Sa and Sv , along with the maximum estimated
˚
stress concentration factor Kt,max , exhibited some correlation with the fatigue resistance
of as-built and PeP-polished specimens. In terms of the fatigue life reached at a given
maximum stress level, these parameters effectively differentiate between significantly
different surface states (like as-built vs PeP). However, they fall short in explaining the
variations in fatigue life among specimens with similar surface states, such as between
different as-built Z specimens. Surprisingly, a better correlation was still found for the Sa
parameter. Conversely, at a fixed fatigue lifetime, the maximum stress that specimens can
endure appears to correlate better with the parameters tested, particularly Kt˚ .

A perspective to improve these results for as-built specimens would be to take into account
the length of the surface valleys. Indeed, a correlation was also found between the fatigue
resistance of as-built specimens and the length of the initiation zone measured on the
fracture surface observed by SEM. The latter is related to the length of the valley at the
origin of fatigue failure.

The results obtained for the prediction of the killer defects using the Kt˚ parameter calculated
from XCT scans were encouraging for specimens polished by PeP. For 4 specimens out of the
8 where the killer defect could be identified on the XCT scan before failure, the killer defect
could be successfully predicted. For the 4 other specimens, the killer defect was among the
10 defects considered to be the most harmful based on their Kt˚ value. However, due to
the lack of resolution of XCT scans, the killer defect could not be identified on the surface
extracted from XCT scans for 6 specimens (out of the 14 that were analyzed).
The killer defect predictions have proved to be a far more complex task for specimens
with an as-built surface. Defects have much more complex geometries than in the case
of PeP specimens, making accurate segmentation more difficult. The high density of defects
(«700/specimen compared to «30/specimen for PeP specimens) also contributes to making
the task more difficult.
Hence, the killer defect predictions clearly failed for as-built Z specimens. The median rank
of the killer defect is thus 182 when using the Kt˚ parameter. Even though the killer defect
could not be successfully predicted in most cases, the results are much better for as-built
45° specimens. This could be due to the larger depth and different shape of valleys in 45°
specimens.

Killer defect prediction | 183

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
In all cases, the ability of the roughness, curvature, and Kt˚ measurements to quantify
surface defects’ harmfulness is thought to be strongly limited by the resolution of XCT
scans. Several examples were presented to show that even with a 3 µm voxel size, the XCT
characterization could lead to a significant blurring which can affect its depth and curvature
measurement. Although the use of the Triangle Threshold for Bimodal Histograms (TTBH)
instead of a more conventional Otsu’s threshold for surface segmentation helps to better
capture critical defects, it seems difficult to achieve significantly better results without
having access to higher resolution scans.

Still, further improvements could be achieved, for instance by taking into account the length
of the surface valleys. This would require improvements in the method used for the surface
defects segmentation, which has been done here essentially by thresholding surface points
based on the local curvature. Also, if higher resolutions are diﬃcult to obtain, it would be
interesting to apply this methodology to specimens with larger surface defects which would
be better captured by XCT at an equivalent resolution. This could, for instance, be the case
by studying as-built samples manufactured by other AM processes such as E-PBF or Wire
Arc Additive Manufacturing (WAAM).

184 | Chapter 5 – Quantitative analysis of surface defects harmfulness

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Conclusion
This work aimed at better understanding the impact of defects on fatigue properties of the
Ti64 alloy manufactured by Laser Powder Bed Fusion (L-PBF). The extensive literature on this
subject reveals that the large surface roughness inherited from this process signiﬁcantly reduces
the fatigue properties. This roughness is caused by surface defects such as valleys that generate
stress concentrations and promote crack initiation under cyclic loading. Characterizing these
defects is challenging because additively manufactured surfaces can show complex 3D shapes
and topography. Traditional surface characterization methods often fall short in these scenarios,
requiring new developments.

A methodology for a thorough 3D surface characterization

To address these challenges, Chapter 2 describes the methodology developed for a thor-
ough 3D characterization of surfaces based on X-ray Computed Tomography (XCT) data. This
approach enables the calculation of roughness and curvature in 3D. Thus, the methodology
could be applied on various components that would have been impossible to characterize using
conventional methods (e.g. a 2.5 characterization using a white light interferometers), namely
additively manufactured architected structures and an impeller produced by L-PBF. This 3D
characterization also makes possible to capture critical defects such as notches hidden by par-
tially melted powder particles at the surface. The methodology proposed makes use of volumetric
operations (e.g. 3D Gaussian filters) and has thus the advantage of being possible to perform
rather easily on a software such as Image. Thus, an ImageJ implementation of the 3D rough-
ness measurement is available in an online repository [STE 23]. A Python implementation is
provided as well.
This methodology developed here could be used in a wide variety of contexts, such as studying
surfaces inherited from different processes (conventional casting, etc.) and in different applica-
tions (mechanical, biomedical, etc.). Here, we illustrate how this methodology can be applied
to investigate the effect of notch-like defects on fatigue properties. For this purpose, a simple
model has been proposed to combine roughness and curvature so as to derive a parameter rep-
resentative of the stress concentration induced by defects, denoted Kt˚ . Enhancements in XCT
data analysis were also explored to improve the identification of critical defects, including the
use of a particular threshold for surface segmentation and the use of 3D curvature analysis to
specifically highlight notches oriented perpendicularly to the loading direction.

Conclusion and Perspectives | 185

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
A case study was conducted on specimens with rather simple surface states (i.e. few isolated
defects) to test the ability of the methodology to predict the defect responsible for the fatigue
failure (referred to as the killer defect). It showed promising results and has the advantage of
being fully automated, with the potential for further improvements. Additionally, this method-
ology applies to more complex surfaces, as demonstrated in subsequent chapters on L-PBF Ti64
specimens with as-built surface states.

Material characterization
Chapter 3 describes the samples used in this work. These are Ti64 mechanical specimens
manufactured by L-PBF using several build orientations: vertical (Z), at 45°, and horizontal
(for tensile specimens only). Samples were stress relieved for 2 h at 720 °C to remove all residual
stresses. After stress relief, the microstructure is composed of a mixture of ﬁne α laths (with an
average thickness of 0.5 µm) along with a small fraction of β phase (2.2 ˘ 0.3%).
It results in a rather high yield strength (Y¯S “ 1098 MPa for Z specimens) but a rather
low elongation to failure (7.8% for Z specimens, and only 0.8% for horizontal specimens). The
strong anisotropy in elongation to failure seems to be partly due to the preferential failure at
prior β0 columnar grain boundaries.
The defects characterization showed that samples have a high density (ą 99.998%) and
that residual pores are both small (over 80% smaller than 30 µm) and have a high sphericity.
Most pores are located in a sub-surface layer at approximately 140 µm from the surface, which
corresponds to the boundary between contour laser scans and the bulk material (hatching).
Several surface defects could be identiﬁed on the 3D surface characterization, the defects that
are expected to cause the most damage in fatigue being valleys. Valleys were found to be deeper
and more tortuous in down-skin areas of specimens built at a 45° angle with respect to the built
plate, in comparison with vertically built specimens.

Fatigue properties of as-built specimens

Chapter 4 assesses the effect of surface defects on fatigue properties. Specimens were first
studied with a stress relief treatment only. The objective was to preserve the as-built surface.
100% of crack initiations take place from surface defects, which emphasizes their crucial role,
much more than internal pores. Conventional fracture surface observations by Scanning Electron
Microscopy (SEM) did not enable the systematic identification of the killer defects, the latter
being shallow and difficult to delineate. The use of the 3D XCT characterization presented in
Chapter 2 allowed to overcome this limitation and demonstrated that most killer defects are
shallow valleys running along the specimens’ surface. Furthermore, it showed that cracks tend
to follow surface valleys over large distances, which is thought to play a role in their detrimental
impact on fatigue resistance.
Quantitatively speaking, the fatigue strength at 106 cycles of as-built specimens is estimated
to be approximately 250 MPa. It is the case for both Z and 45° orientations, although 45° spec-
imens show on average slightly lower fatigue lives than Z specimens. In any case, the fatigue
strength of as-built specimens is much lower than that of Machined and manually Polished spec-
imens (M&P), which is around 850 MPa. This confirms the critical impact of surface defects, as
well as the possibility of reaching a very high fatigue resistance if defects are properly controlled.

186 | Conclusion and Perspectives

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fatigue properties after a heat treatment leading to surface oxygen contami-
nation
Some vertically built specimens were subjected to a subsequent higher temperature heat
treatment (860 °C for 4 h). The latter leads to the formation of an oxygen-contaminated surface
layer, referred to as α-case, of about 70 µm in depth. This layer was found to be responsible for
a slight decrease in fatigue resistance. This is most probably caused by the brittleness of the
α-case layer which promotes crack initiation, as suggested by the high density of initiated cracks
observed in those specimens.

Fatigue properties improvements after a surface ﬁnishing treatment

Some as-built specimens underwent Plasma electrolytic Polishing (PeP) to enhance fatigue
properties. Several polishing conditions led to diﬀerent polishing depths, either around 180 µm or
340 µm. A 180 µm removal improved fatigue strength at 2 ˆ 106 cycles to approximately 400 to
550 MPa. This improvement was limited due to the exposure of sub-surface pores (particularly
numerous at about 140 µm from the surface), creating pit-like defects that led to fatigue failure.
In contrast, a more extensive 340 µm removal that goes beyond the subsurface layer containing a
high density of pores signiﬁcantly increased fatigue strength at 2 ˆ 106 cycles, above 1000 MPa.

Correlation between fatigue resistance and parameters derived from the 3D

surface characterization
In Chapter 5, the quantitative and predictive capabilities of the surface characterization
method introduced in Chapter 2 were evaluated. Specifically, it tested whether this method
could give interesting indicators of a specimen’s fatigue resistance and accurately predict the
killer defect.
Conventional roughness parameters such as Sa capture the difference between very different
surface states, e.g. they capture the fact that as-built specimens have lower fatigue lives on
average than specimens polished by PeP. However, they are unable to account for the fatigue
lives scatter observed between similar surface states, e.g. between two as-built Z specimens.
These parameters seem better correlated with the maximum stress that can be endured by a
specimen for a targeted lifetime, the Kt˚ parameter yielding, in this case, the most promising
results. Results could be improved by taking into account other factors such as the length of
valleys.

Prediction of the killer defect

The prediction of killer defects using the Kt˚ parameter showed good results for specimens
polished by PeP. However, it seems strongly limited by the resolution of the XCT scans. Thus,
for 6 out of 14 specimens, XCT could not detect the killer defect due to an insuﬃcient resolution.
Over the remaining 8 specimens, the method accurately identiﬁed the killer defect in 50% of the
cases.
Predicting killer defects in as-built Z specimens turned out to be more challenging and did
not lead to conclusive results. However, results for as-built 45° specimens were much more
encouraging.

Conclusion and Perspectives | 187

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Improving XCT resolution could potentially yield better predictions, as valleys in as-built
specimens were often too shallow to be adequately captured with the 3 µm voxel size used in this
study. Considering the shape and length of surface valleys might also enhance the predictions.
Furthermore, as-built surfaces present a large population of valleys with similar shapes and
mechanical harmfulness. Thus, the exact location of the killer defect within this population
could depend on other factors such as the local surface microstructure.

Perspectives
This work opens various avenues for future developments and a deeper understanding of
the impact of surface defects on fatigue properties. Below are some key areas for potential
improvements.

Enhancing XCT scan resolution

One of the main challenges in correlating 3D roughness and curvature characterization with
fatigue resistance or killer defect location is the insufficient resolution of XCT scans. Thus, the
most obvious (and maybe the most promising) lever for improvement is the access to higher-
resolution scans. These could be obtained with advanced laboratory X-ray tomographs capable
of smaller voxel sizes or by using synchrotron XCT facilities.
An experiment was conducted to evaluate the impact of improved resolution. The Z-10
(as-built Z) specimen, was scanned at the I12 Beamline of the Diamond synchrotron (UK),
thanks to Thomas Connoley, James Ball, and David Collins. The killer defect was identified,
and its detection using laboratory XCT (with a voxel size of 3 µm) was compared to that using
synchrotron XCT (with a voxel size of 1.3 µm), see Fig.5.24a-e.
The 3D κmin map in Fig.5.24a illustrates that a better resolution of the surface is achieved
with synchrotron XCT compared to what could be obtained using laboratory XCT in this
work. The radial slices (Fig.5.24b-c) centered on the killer defect, a valley, further highlight this
difference. Enlarged views (Fig.5.24d-e) provide a closer look at the killer defect, with surface
segmentation performed using the TTBH.
Although laboratory XCT partially captures the killer defect, synchrotron XCT offers a
more detailed and accurate description of the defect. This suggests that improved resolution
could significantly enhance surface characterization, especially in measuring curvature at the
notch root. Hence, pursuing higher-resolution scans could potentially yield better results and
predictions in future studies. However, it is worth keeping in mind that accessing synchrotron
XCT can be difficult and that higher resolutions may lead to large amounts of data that require
significant computing resources for analysis.

188 | Conclusion and Perspectives

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.24 | Comparaison of killer defect identiﬁcation using laboratory XCT (vox.
= 3 µm) and synchrotron XCT (vox. = 1.3 µm) on Z-10 specimen. (a) 3D minimum
curvature map centered on the killer defect. (b) and (c) show radial slices centered on the killer
defect using either laboratory or synchrotron XCT. (d) and (e) show enlarged views of (b) and
(e) with the surface segmentation achieved in both cases using the TTBH.

Conclusion and Perspectives | 189

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Comparing 2.5D and 3D characterization methods
Given the challenges in obtaining high-resolution XCT scans, a practical alternative could in-
volve using high-resolution 2.5D instruments like white light interferometry. These instruments
typically oﬀer better resolutions than XCT [VIL 19]. This suggestion might appear counterin-
tuitive, given that part of this research underscores the necessity of true 3D characterization
for adequately capturing critical surface defects typical of as-built surfaces in additive manu-
facturing. However, the higher resolution that can be achieved using 2.5D instruments may
justify the trade-oﬀ depending on the case. Consequently, conducting further studies to com-
pare the predictive accuracy (e.g. regarding fatigue properties) of 3D XCT measurements and
higher-resolution 2.5D measurements could be worthwhile.

Figure 5.25 | Surface roughness and curvature characterization from 2.5D data. (a)
2.5D roughness measurement obtained using a chromatic confocal sensor (HiroX NPS) and a
lateral spacing for measurement of 1 µm. (b) Minimum curvature map, computed using rcurv “
10 µm. (c) Superposition of (a) and (b).

190 | Conclusion and Perspectives

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
As an example, Fig.5.25a-c present a 2.5D surface characterization of an as-built Z speci-
men. Fig.5.25a displays the roughness map, from which curvature is computed using the same
integral invariant method described in Chapter 2. Fig.5.25b shows the resulting minimum cur-
vature map, and Fig.5.25c combines the roughness and curvature maps for a comprehensive
surface representation of both details and larger scale variations. The lateral measurement
spacing here is 1 µm, achieving a detailed surface representation comparable to the one obtained
with synchrotron XCT, see Fig.5.24a. Thus, the accuracy of the surface characterization may
be improved accordingly. Of course, despite the improved resolution, the 2.5D characteriza-
tion might miss or only partially capture hidden features, potentially leading to inaccuracies,
particularly in curvature estimation.

Enhancing defects segmentation and considering defects’ shape and neighbor-

hood
If enhancing the spatial resolution of surface characterization is not an option, alternative
avenues for improvement remain. Thus, Chapter 5 highlights the potential for more accurately
assessing the impact of valley-type defects by considering their shape, particularly their length.
Additionally, the neighborhood of a defect, and in particular nearby valleys, can also influence
its harmfulness. A well-known phenomenon, the shielding effect, suggests that a single valley
can generate higher stress concentration compared to multiple valleys close to one another
[BAR 23, BAR 21].
Improving the defect segmentation methodology could lead to a better understanding and
quantification of these effects. While enhanced XCT scan resolution, as discussed in Chapter
5, would most probably be of great help, exploring other segmentation techniques could also
be effective. The watershed method, combined with Wolf pruning, is the method outlined in
ISO standards [Int 21, LOU 20] and could be an interesting choice. One of the few examples
of watershed-based segmentation on a free-form surface found in the literature is shown in
Fig.5.26a-d.
However, it might be necessary to adapt the watershed and Wolf pruning algorithms to better
take into account the elongated shape of valley-type defects and break the valleys network
mentioned in Section 5.1.1, for example by using directional curvature measurement besides
roughness.
Alternative techniques could be explored for more precise segmentation of valley roots, as the
watershed method typically segments the entire surface, categorizing all points as part of a defect
(see Fig.5.26d). For instance, the works of [COE 16, LEV 15] introduce segmentation methods
that rely on estimations of curvature or normal fields. These approaches could potentially offer a
more targeted segmentation, specifically focusing on the critical areas of the valleys. An example
of sharp feature segmentation (here, edges instead of notches) using this method is shown in
Fig.5.27.

Conclusion and Perspectives | 191

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure 5.26 | Surface valleys segmentation on a free-form surface using watershed
method together with Wolf pruning, adapted from [LOU 20]. (a) Initial surface mesh,
derived from the Rosetta comet model developed by the European Space Agency. (b) Smoothed
mesh obtained by Laplacian smoothing and used as a reference for roughness computation. (c)
3D roughness map obtained by computing the distance between (a) and (b). (d) Result of the
watershed segmentation after Wolf pruning.

Figure 5.27 | Sharp features extraction using the approach of [COE 16], based on
normal vector ﬁeld regularization. Sharp edges are successfully segmented in red, even
when some noise is added to the data (see the lower right half of the image).

192 | Conclusion and Perspectives

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Selecting the appropriate scale for surface analysis
Another option to improve results regardless of the spatial resolution of XCT scans would
be to gain a deeper understanding of the role of the scale at which the analysis is conducted.
In our study, the roughness measurement scale is defined by the cut-off wavelength λc , and
the curvature measurement scale by the convolution radius rcurv . The results in Section 2.4
and Section 5.3 have shown that the prediction ability of our model heavily depends on these
parameters.
Setting these parameters appropriately can be challenging, a complexity only partially ad-
dressed in this work. As highlighted in Section 2.4, fully understanding and describing surface
features may even require characterizing the surface at multiple scales. This concept of multi-
scale analysis has been explored in various studies for both roughness [JIA 20a, GO 16] and
curvature [BRO 18, BAR 19, VUL 14].
In fact, our comparison of killer defect predictions for various values of rcurv in Sections
2.4 and 5.3 already ludes elements of multi-scale analysis. This is somewhat analogous to
the approach of [VUL 14], who calculated surface curvature at multiple scales on machined
steel specimens to find the scale where the maximal mean curvature correlates best with the
experimental fatigue resistance. This approach could be further studied in our case by computing
curvature (and/or roughness) at an increased number of scales and trying to link the most
relevant scales to the observed surface defects, particularly those causing fatigue failure.
Further work could also focus on integrating measurements from different scales, possibly
requiring more complex modeling approaches. For instance, [LEV 15] suggests a method for
segmenting surface features (like sharp edges) using the slope of curvature change with respect
to the convolution radius rcurv , rather than the curvature at a fixed scale. This method can
identify features existing at varying scales on the same surface, offering a more comprehensive
characterization.

Applying the 3D characterization method to various materials and processes

Besides the influence of the resolution of XCT scans and the methodology used for data
analysis, the accuracy of predictions may also depend on the manufacturing processes used
for sample manufacturing. Exploring the fatigue properties of samples made using different
additive manufacturing techniques, such as Electron Beam Powder Bed Fusion (E-PBF), Wire
Arc Additive Manufacturing (WAAM), or conventional casting, could therefore be insightful.
Notably, these methods often result in larger defects, potentially simplifying the characterization
and providing more conclusive results even with a limited spatial resolution.
Investigating materials different from Ti64, with different properties (e.g. different notch
sensitivities) could also provide valuable insights, as they might respond differently to fatigue
loading with a similar defect population. A preliminary study on IN718 bending fatigue speci-
mens produced by L-PBF has shown promising results, see Fig.5.28a-d.
The Kt˚ maps before and after the fatigue test are displayed in Fig.5.28b and Fig.5.28c,
respectively. They are shown from an internal point of view (i.e. from under the surface). This
allows to better observe hidden features such as the notch at point A shown in the enlarged
view in Fig.5.28a.

Conclusion and Perspectives | 193

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Fig.5.28c shows the propagated crack’s location, and comparison with Fig.5.28b helps identify
major defects along the crack path. The defect at point A shows the highest Kt˚ value along this
path. Fig.5.28d is an observation of this defect on an XCT scan slice. Only one other surface
location, point B in Fig.5.28b, exhibits higher Kt˚ values. Therefore, the actual killer defect was
the second most critical based on the Kt˚ parameter, indicating a reasonably good prediction.

Figure 5.28 | Comparing the killer defect prediction to the actual crack location
after failure in an IN718 bending fatigue specimen manufactured by L-PBF. The
surface subjected to fatigue loading is the down-skin part of a specimen built a 45°. The
characterization was done using XCT scans with a voxel size of 4 µm. (a) Enlarged view of the
Kt˚ map of the defect at point A, belonging to the crack path. Point B is the only location of
the surface where Kt˚ values superior to those in point A were found. (b) Internal view of the
surface Kt˚ map before fatigue testing. (c) Internal view of the surface Kt˚ map after fatigue
testing. The propagated crack is characterized by high Kt˚ values and can thus be seen in red.
(d) XCT slice revealing the notch defect at point A.

194 | Conclusion and Perspectives

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Combining Hot Isostatic Pressing (HIP) with surface finishing
In parallel to enhancing our understanding and predictive capabilities regarding the impact
of defects on fatigue properties, exploring ways to improve the fatigue properties themselves
through various post-treatments would be interesting. Although not explored in this study, Hot
Isostatic Pressing (HIP) treatment has for example the potential to offer significant benefits,
particularly if properly combined with a subsequent surface finishing step.
Chapter 4 demonstrated that surface finishing significantly enhances the fatigue properties
of L-PBF Ti64, provided it does not expose internal pores to the surface. A polishing depth of
340 µm was particularly effective, completely removing the subsurface layer where most pores
were found to be located. To avoid the issues associated with such extensive polishing (like
dimensional inaccuracies), a preliminary HIP treatment to close internal pores might be more
advantageous. Following HIP, a milder polishing with a reduced polishing depth of around
200 µm could be sufficient to remove surface valleys and achieve excellent material properties.
However, it is important to note that HIP, typically conducted at high temperatures (e.g.,
920 °C for 2 hours for Ti64 alloy), can also alter the microstructure, potentially affecting me-
chanical properties or causing surface contamination. Additionally, HIP significantly adds to
the manufacturing cost. Thus, a comprehensive study assessing the pros and cons and finding
an optimal balance for this treatment could be valuable.

Conclusion and Perspectives | 195

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Comparison of diﬀerent distance

calculation methods

Measuring roughness on free-form surfaces involves calculating the distance between the
original rough surface and a reference surface. In this study, the original surface is defined by
what we call the sample mask or surface mask, as introduced in Chapter 2. The reference surface
is identified as the smooth mask in the same chapter. Additionally, we can introduce here the
concept of the smooth surface mask, which is to the smooth mask what the surface mask is to
the sample mask.
Different methods can calculate this distance, and while they often yield similar results, some
may introduce more errors, especially in specific scenarios. This appendix focuses on comparing
two distinct methods.
The method presented in Chapter 2 involves the use of a Signed Euclidean Distance Transform
(SEDT). Its advantage is that it is a 3D operation, directly available on software such as ImageJ.
Alternatively, a Nearest Neighbor Search (NNS) technique can be more appropriate in certain
cases. NNS seeks the closest point(s) in a point cloud for a given point. In our context, it can be
used to identify the nearest point on the smooth surface for each point on the (rough) surface.
The local height, or roughness, is then the distance between these two points. Practically, NNS
is often executed using a kd-tree data structure [SKR 19], as used in this study. In terms of
computing time, the NNS approach is similar to SEDT, but better than SEDT in terms of
memory usage. However, unlike SEDT or Gaussian filters, it is not a volumetric operation and
is not readily available in software like ImageJ without custom implementation. In this work,
the NNS computation was done using the scipy-spatial Python library. The SEDT method was
also conducted using Python, utilizing the edt library, which offers an efficient multi-threaded
implementation. Note that the Python implementation for 3D roughness measurement provided
in [STE 23] proposes the two methods for distance calculation. In the script, the SEDT method
is referred to as the ’edt’ method, while the NNS method is referred to as the ’kdtree’ method.
Fig.A.1a-e presents a comparison of roughness measurements obtained using SEDT and NNS
methods on a straightforward example: a perfect cylinder with a radius of 3.5 voxels. In this
scenario, both the sample mask and the smooth mask are identical, as they represent the same
perfect cylinder (shown in Fig.A.1a). Similarly, the surface mask and the smooth surface mask
are also the same (see Fig.A.1b).

| 197

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Theoretical
sample shape

Background voxel Foreground voxel Surface voxel

c d e
Theoretical result SEDT method NNS method
2.8 2.2 1.4 1 1 1 1.4 2.2 2.8

0 0 0 2.2 1.4 1 -1 -1 -1 1 1.4 2.2 0 0 0

0 0 1.4 1 -1 -1.4 -2 -.9 -1 1 1.4 0 0
0 0 1 -1 -1.4 -2.2 -2.8 -1.7 -.9 -1 1 0 0
0 0 1 -1 -2 -2.8 -3.6 -2.3 -1.5 -1 1 0 0
0 0 1 -1 -.9 -1.7 -2.3 -1.7 -.9 -1 1 0 0
0 0 1.4 1 -1 -.9 -1.5 -.9 -1 1 1.4 0 0
0 0 0 2.2 1.4 1 -1 -1 -1 1 1.4 2.2 0 0 0
2.8 2.2 1.4 1 1 1 1.4 2.2 2.8

d = 0 for all surface voxels d = distance to d = distance to

since smooth mask is nearest background voxel nearest foreground voxel
identical to sample mask of smooth sample mask of smooth surface mask
Mean error = 1 Mean error = 0
Figure A.1 | Comparing mean errors in roughness measurement between SEDT
and NNS methods for a perfect cylinder. Each subﬁgure displays a cross-sectional
slice of the cylinder. (a) Sample mask showing the cylinder’s theoretical shape and surface
voxels. This also represents the ideal sample mask as the smooth version of a perfect cylinder
is identical. (b) Surface mask, identical to the smooth surface mask in this case. (c) Expected
theoretical roughness values for each surface voxel. (d) SEDT results, with roughness values of
-1 for all surface voxels. (e) Roughness values, determined by the NNS method for each surface
voxel.

198 | Appendix A – Comparison of diﬀerent distance calculation methods

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
In this example, where the surface and its smooth reference are the same, the anticipated
local height for all surface voxels should be zero (as seen in Fig.A.1c). However, SEDT calculates
the distance from each surface voxel to the nearest background voxel of the smooth sample mask,
resulting in a -1 value. Consequently, this leads to an error of 1 for each voxel, and an overall
mean error of 1, deﬁned as:

Mean error “ mean |height ´ theoretical height| (A.1)

surface voxels

In this speciﬁc case, the error can be corrected by subtracting the mean height from all surface
voxels post-calculation, a process that will be referred to as zero-centering. This correction step
is recommended in Chapter 2, notably for situations like this.
For a more in-depth comparison, we used a cylinder with an artiﬁcial sinusoidal roughness
on its surface. The cylinder is 4 mm in height and has an average diameter of 1 mm, with the
sinusoidal roughness featuring a wavelength of λ “ 100 µm and amplitude A varying from 0 to
30 µm. Figure A.2 provides a schematic representation of a 2D segment of this surface. The
sample mask, used for calculations, is generated from the analytical formula of this shape, using
a voxel size of 2.5 µm.

λ = 100 µm A ∈ [0,30] µm

Bulk material

Figure A.2 | Illustration of a section of the sinusoidal roughness used for the example.

3D roughness maps for this cylinder with A “ 15 µm are depicted in Fig.A.3a-c.

Fig.A.3a serves as a reference, showing the ideal results one would get by directly computing
the local height at each surface voxel based on the analytical formula. Figures A.3b and A.3c
display the results from the SEDT and NNS methods, respectively. Both methods produce
roughness maps that closely resemble the theoretical ones. However, enlarged views reveal that
the discrete nature of distance calculations results in abrupt local height changes, unlike the
smoother theoretical results in Fig.A.3a.
Fig.A.4a-b quantitatively compare the mean error and the average arithmetic roughness Sa
for amplitude values A from 0 to 30 µm. These ﬁgures assess which method, SEDT or NNS,
yields more accurate results compared to the theoretical values. Additionally, both methods are
evaluated with and without the zero-centering step to determine its eﬀectiveness in reducing
errors.
Fig.A.4a indicates that generally, the NNS method outperforms the SEDT method, partic-
ularly for smaller amplitude values (A), similar to the smooth cylinder case seen in Fig.A.1.
However, applying the zero-centering step allows to reduce to some extent the error generated
by the SEDT method. For the NNS method, at least in our example of a sinusoidally roughened
cylinder, zero-centering does not improve accuracy.

| 199

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure A.3 | 3D roughness maps of the cylinder with sinusoidal roughness. (a)
Displays the ideal, analytically derived roughness values. (b) Shows results using the SEDT
method. (c) Presents results obtained with the NNS method.

These ﬁndings suggest that while the choice between SEDT and NNS methods may not be
critical for surfaces with signiﬁcant roughness, it becomes crucial for smoother surfaces. This is
for example the case for PeP-polished samples from Chapter 4, where the SEDT method was
found to cause substantial errors in some roughness measurements.

The error from using SEDT, as opposed to a method like NNS, will typically be around
the voxel size. While this may not seem substantial for single voxel height measurements or for
calculating roughness parameters like Sz or Sv , it can signiﬁcantly aﬀect average parameters such
as Sa . This is because a consistent 2.5 µm negative error (as seen in Fig.A.1d) will substantially
skew the Sa value.

200 | Appendix A – Comparison of diﬀerent distance calculation methods

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

2.5
Mean error (µm)

2.0

1.5

1.0

0.5

0.0 Theory
0 5 10 15 20 25 30 EDT
A (µm) EDT | Zero-centered
b
NNS
NNS | Zero-centered
15
Sa (µm)

0
0 5 10 15 20 25 30
A (µm)

Figure A.4 | Comparing roughness measurements from SEDT and NNS methods
with theoretical calculations. (a) Theoretical results with analytically calculated roughness
values. (b) Roughness measurements using the SEDT method. (c) Measurements from the NNS
method.

This eﬀect is visible in Fig.A.4b, where using the SEDT method without zero-centering
results in Sa values markedly higher than theoretical ones, particularly at small A values. This
discrepancy persists even at larger A values. In contrast, the NNS method delivers Sa values
very close to the theoretical results.
In summary, the NNS method is preferable to SEDT where possible. If SEDT must be used,
the zero-centering step can reduce errors. However, caution is advised when characterizing very
smooth surfaces, particularly if Sa values are close to the voxel size.

| 201

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

3D Gaussian S-Filter using

normalized convolution

The following methodology can be used to apply any 3D linear filter on a set of voxels. In
particular, it can be used to apply a Gaussian S-filter (or L-filter) to roughness values calcu-
lated following the workflow described in Section 2.2.2. It can be considered as a particular
application case of the concept of normalized convolution introduced by [KNU 93]. Unlike a
conventional convolution where all voxels are taken into account, normalized convolution can be
used to ignore certain voxels. Here, it will be used to ignore background voxels and thus filter
only the ones that carry actual information − i.e. the surface voxels that carry roughness values.

A linear filter is an operation where the value at a given voxel is replaced by a linear combi-
nation of the value at the given point and its neighbors. Each neighbor has a specific weight w,
i.e. the weight of its contribution to the final filtered value. The matrix assigning the weights
for the central voxel and its neighbors is called the filter kernel. For a uniform (= mean) filter,
all weights in the kernel will have the same weight. In the case of a Gaussian filter, weights
decrease as the distance to the central voxel increases following a Gaussian law, see Fig.B.1. In
both cases, weights are generally normalized, meaning the sum of all kernel weights equals 1.
For example, this ensures that applying a Gaussian filter to a uniform volume made of 1 will
result in a uniform volume made of 1. This is the standard "global" normalization illustrated in
Fig.B.1 and employed in common Gaussian filter implementations.

However, the situation is a bit different when applying a S-filter to a set of surface voxels. In
such a case, to compute the filtered value at a given voxel, only neighbor surface voxels should be
taken into account instead of all neighbor voxels. The solution would be to explicitly loop over
surface voxels only, and never visit background voxels. This can be done in programming lan-
guages such as C++. However, this still requires some programming skills, especially to achieve
reasonable computation times. Conversely, there are already many efficient implementations of
the standard Gaussian filter, including ones accelerated via GPU of FFT computations. The
idea here is thus to use such optimized implementations ingeniously to compute indirectly the
S-filter.

| 203

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

1 2 1
Discrete Gaussian kernel 2 4 2 Sum = 16
without normalization
1 2 1

"Local" normalization
Only surface voxels are taken into account

1 2
Apply standard Gaussian filter Apply standard Gaussian filter
to roughness to normalization volume
Pixel being
processed 0 1 0 0 0 0
0 6.7 0 0 0 0
0 8.9 0 0 0 0 0 1 0 0 0 0

0 0 8.6 7.9 0 0
Surface 0 0 1 1 0 0
pixel
0 0 0 0 7.2 0 0 0 0 0 1 0

0 0 0 0 8.1 8.9 0 0 0 0 1 1
Background
0 0 0 0 0 0 pixel 0 0 0 0 0 0

Figure B.1 Workflow proposed for the application of a 3D Gaussian S-filter. All calculations are
done using operations on digital volumes, just as the rest of the roughness computation workflow.
The example provided in the present figure is restricted to 2D, which means voxels are replaced
by pixels. This choice is made to facilitate understanding and visualization, considering the
adaptation to a 3D case is straightforward. Note that in steps 1 and 2 of the "local normalization"
in Fig.B.1, a non-normalized Gaussian filter is used for the sake of clarity only. The same steps
can therefore be followed using a conventional Gaussian filter.

204 | Appendix B – 3D Gaussian S-Filter using normalized convolution

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
To do so, the first step consists of applying a conventional Gaussian filter to the volume
where surface voxels contain roughness values, while background voxels contain 0 − see step 1
in Fig.B.1. Thus, background voxels do not contribute to the final filtered value. Even though,
the obtained filter is not properly normalized. Indeed, the sum of the weights of all neighbors
that are taken into account (i.e. all neighbor surface voxels) should equal 1. However, this
sum will be in general much lower than one, because most neighbor voxels are background ones
(equal to 0). To correct this bias, one simply needs to compute the same conventional Gaussian
filter to a volume containing 0 at background voxels and 1 at surface voxels, see step 2 in Fig.B.1.

The resulting value at each point is the sum of all the Gaussian kernel weights that effectively
come across surface voxels. It can be considered as a "local normalization factor", since by
dividing the filtered value obtained in step 1 by this value, we obtain a properly normalized
S-filter, see step 3 in Fig.B.1. This method is equivalent to use at each voxel a specific kernel.
The latter contains 0 at background voxels, and at other voxels, a weight decreasing according to
a Gaussian law with the distance to the central voxel. To be normalized, the sum of all weights
must be equal to one. The trick here is to achieve such a (non-linear) computation using only
linear filtration steps − i.e. using the same unique kernel for all voxels.

| 205

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Near-surface microstructure after

PeP treatment

The following ﬁgures show EBSD characterizations on vertical specimens before and after
PeP treatment. Average α-laths thickness for the as-built (Z), Z|Vertical Vertical specimens
40min and Z|1h20
characterized were measured to be 0.609 µm, 0.593 µm and 0.594 µm respectively.
Acquisition and analysis were done by Quentin Gaillard. For a more detailed analysis of the
near-surface microstructure, see Section 3.2.2 or the PhD thesis of Quentin Gaillard [GAI 23a].

| 207

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure C.1 | EBSD analysis of the microstructure near the surface in a transverse
cross-section of a vertical fatigue specimen before PeP treatment. Crystallographic
orientation maps of (a) α and (e) prior β0 phases. Pole figures of (b) α and (f) prior β0 phases in
the border area. Inverse pole figures color maps for (c) α and (g) prior β0 phases. Inverse pole
figures of (d) α and (h) prior β0 phases in the border area. The build direction Z is taken as a
reference. The maps are 600 µm ˆ 445 µm wide and are collected with a step size of 0.25 µm.

208 | Appendix C – Near-surface microstructure after PeP treatment

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure C.2 | EBSD analysis of the microstructure near the surface in a transverse
cross-section of a vertical fatigue specimen after a 40 min PeP treatment. Crystallo-
graphic orientation maps of (a,e) α and (i) prior β0 phases. Pole figures of (b,f) α and (j) prior
β0 phases in the border area. Inverse pole figures color maps for (c,g) α and (k) prior β0 phases.
Inverse pole figures of (d,h) α and (l) prior β0 phases in the border area. The build direction Z
is taken as a reference. The map in (a) is 75 µm ˆ 75 µm wide and is done with a step size of
0.06 µm. The map in (e) and (i) are 400 µm ˆ 400 µm wide and are collected with a step size of
0.25 µm.

| 209

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Figure C.3 | EBSD analysis of the microstructure near the surface in a transverse
cross-section of a vertical fatigue specimen after a 1 h 20 min PeP treatment. Crys-
tallographic orientation maps of (a) α and (e) prior β0 phases. Pole figures of (b) α and (f) prior
β0 phases in the border area. Inverse pole figures color maps for (c) α and (g) prior β0 phases.
Inverse pole figures of (d) α and (h) prior β0 phases in the border area. The build direction Z
is taken as a reference. The maps are 400 µm ˆ 400 µm wide and are collected with a step size
of 0.25 µm.

210 | Appendix C – Near-surface microstructure after PeP treatment

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

[ABD 13] Abdul-Rahman H. S., Jiang J., Scott P. J.

A Gaussian-like Filtering algorithm for Freeform Surfaces Represented by Triangular Meshes.
, 2013. Publisher: Unpublished. 35
[ABO 14] Aboulkhair N. T.
Reducing porosity in AlSi10Mg parts processed by selective laser melting. Additive Manufac-
turing, , 2014, Page 10. 15
[AHM 20] Ahmed Obeidi M., Mussatto A., Groarke R., Vijayaraghavan R. K., Con-
way A., Rossi Kaschel F., McCarthy E., Clarkin O., OConnor R., Brabazon D.
Comprehensive assessment of spatter material generated during selective laser melting of
stainless steel. Materials Today Communications, vol. 25, 2020. 13
[ALE 05] Alekseev Y., Kosobutsky A., Korolev A., Niss V., Kucheryavy V.,
Povszhik A.
Features of the processes of dimensional processing of metal products by the electrolyte-plasma
method. Litye. Metall., vol. 4, 2005, p. 188–195. 133, 137
[ALI 20] Ali U.
Internal surface roughness enhancement of parts made by laser powder-bed fusion additive
manufacturing. Vacuum, vol. 177, 2020. 35
[ALL 11] Alla R. K., Ginjupalli K., Upadhya N., Shammas M., Krishna R., Sekhar
R.
Surface Roughness of Implants: A Review. Trends in Biomaterials & Artiﬁcial Organs,
vol. 25, no 3, 2011, p. 112–118. 39
[AND 19a] Andreau O.
Nocivité en fatigue et contrôle de défauts produits par fabrication additive. Doctoral thesis,
Ecole Nationale Supérieure d’Arts et Métiers, 2019. 12, 30, 62
[AND 19b] Andreau O., Pessard E., Koutiri I., Penot J.-D., Dupuy C., Saintier N.,
Peyre P.
A competition between the contour and hatching zones on the high cycle fatigue behaviour
of a 316L stainless steel: Analyzed using X-ray computed tomography. Materials Science and
Engineering: A, vol. 757, 2019, p. 146–159. 18, 28
[ARO 99] Arola D., Ramulu M.
An Examination of the Eﬀects from Surface Texture on the Strength of Fiber Reinforced
Plastics. Journal of Composite Materials, vol. 33, no 2, 1999, p. 102–123. 33, 161
[ARO 02] Arola D., Williams C.
Estimating the fatigue stress concentration factor of machined surfaces. International Journal
of Fatigue, vol. 24, no 9, 2002, p. 923–930. 33

References | 211

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[ATZ 12] Atzeni E., Salmi A.
Economics of additive manufacturing for end-usable metal parts. The International Journal
of Advanced Manufacturing Technology, vol. 62, no 9-12, 2012, p. 1147–1155. 2
[BAG 17] Bagehorn S., Wehr J., Maier H.
Application of mechanical surface finishing processes for roughness reduction and fatigue
improvement of additively manufactured Ti-6Al-4V parts. International Journal of Fatigue,
vol. 102, 2017, p. 135–142. 21, 22
[BAR 19] Bartkowiak T., Brown C. A.
Multiscale 3D Curvature Analysis of Processed Surface Textures of Aluminum Alloy 6061 T6.
Materials, vol. 12, no 2, 2019. 69, 193
[BAR 21] Barricelli L., Beretta S.
Analysis of prospective SIF and shielding effect for cylindrical rough surfaces obtained by
L-PBF. Engineering Fracture Mechanics, vol. 256, 2021. 26, 191
[BAR 23] Barricelli L., Patriarca L., du Plessis A., Beretta S.
Orientation-dependent fatigue assessment of Ti6Al4V manufactured by L-PBF: Size of surface
features and shielding effect. International Journal of Fatigue, vol. 168, 2023. 27, 191
[BAS 22] Basha M., Basha S., Jain V., Sankar M.
State of the art on chemical and electrochemical based finishing processes for additive manu-
factured features. Additive Manufacturing, vol. 58, 2022. 3, 131, 133
[BAT 99] Bathias C.
There is no infinite fatigue life in metallic materials. Fatigue & Fracture of Engineering
Materials & Structures, vol. 22, no 7, 1999, p. 559–565. 9
[BAY 19] Bayat M., Thanki A., Mohanty S., Witvrouw A., Yang S., Thorborg J.,
Tiedje N. S., Hattel J. H.
Keyhole-induced porosities in Laser-based Powder Bed Fusion (L-PBF) of Ti6Al4V: High-
fidelity modelling and experimental validation. Additive Manufacturing, vol. 30, 2019. 15
[BEL 20] Belkin P., Kusmanov S., Parfenov E.
Mechanism and technological opportunity of plasma electrolytic polishing of metals and alloys
surfaces. Applied Surface Science Advances, vol. 1, 2020. 131, 133, 137
[BEN 17] Benedetti M., Torresani E., Leoni M., Fontanari V., Bandini M., Peder-
zolli C., Potrich C.
The effect of post-sintering treatments on the fatigue and biological behavior of Ti-6Al-4V
ELI parts made by selective laser melting. Journal of the Mechanical Behavior of Biomedical
Materials, vol. 71, 2017, p. 295–306. 14, 31
[BEN 18] Benedetti M., Fontanari V., Bandini M., Zanini F., Carmignato S.
Low- and high-cycle fatigue resistance of Ti-6Al-4V ELI additively manufactured via selective
laser melting: Mean stress and defect sensitivity. International Journal of Fatigue, vol. 107,
2018, p. 96–109. 82
[BER 17] Beretta S., Romano S.
A comparison of fatigue strength sensitivity to defects for materials manufactured by AM or
traditional processes. International Journal of Fatigue, vol. 94, 2017, p. 178–191. 12, 27
[BEZ 20] Bezuidenhout M.
The effect of HF-HNO3 chemical polishing on the surface roughness and fatigue life of laser
powder bed fusion produced Ti6Al4V. Materials Today Communications, , 2020, Page 10. 3

212 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[BHA 23] Bhandari L., Gaur V.
Different post-processing methods to improve fatigue properties of additively built Ti-6Al-4V
alloy. International Journal of Fatigue, vol. 176, 2023. 33
[BIA 11] Biamino S., Penna A., Ackelid U., Sabbadini S., Tassa O., Fino P., Pavese
M., Gennaro P., Badini C.
Electron beam melting of Ti48Al2Cr2Nb alloy: Microstructure and mechanical properties
investigation. Intermetallics, vol. 19, no 6, 2011, p. 776–781. 15
[BLU 19] Blunt M. J., Lin Q., Akai T., Bijeljic B.
A thermodynamically consistent characterization of wettability in porous media using high-
resolution imaging. Journal of Colloid and Interface Science, vol. 552, 2019, p. 59–65.
52
[BOB 21] Boban J., Ahmed A.
Improving the surface integrity and mechanical properties of additive manufactured stain-
less steel components by wire electrical discharge polishing. Journal of Materials Processing
Technology, vol. 291, 2021, Page 117013. 131
[BOO 16] Book T. A., Sangid M. D.
Strain localization in Ti-6Al-4V Widmanstätten microstructures produced by additive man-
ufacturing. Materials Characterization, vol. 122, 2016, p. 104–112. 10
[BOR 02] Borbély A., Mughrabi H., Eisenmeier G., Höppel H. W.
A finite element modelling study of strain localization in the vicinity of near-surface cavities
as a cause of subsurface fatigue crack initiation. International Journal of Fatigue, vol. 115,
no 3, 2002, p. 227–232. 17
[BRI 08] Bridier F., Villechaise P., Mendez J.
Slip and fatigue crack formation processes in an / titanium alloy in relation to crystallographic
texture on different scales. Acta Materialia, vol. 56, no 15, 2008, p. 3951–3962. 172
[BRO 18] Brown C. A., Hansen H. N., Jiang X. J., Blateyron F., Berglund J., Senin
N., Bartkowiak T., Dixon B., Le Goïc G., Quinsat Y., Stemp W. J., Thompson
M. K., Ungar P. S., Zahouani E. H.
Multiscale analyses and characterizations of surface topographies. CIRP Annals, vol. 67, no
2, 2018, p. 839–862. 193
[CAB 18] Cabanettes F., Joubert A., Chardon G., Dumas V., Rech J., Grosjean C.,
Dimkovski Z.
Topography of as built surfaces generated in metal additive manufacturing: A multi scale
analysis from form to roughness. Precision Engineering, , 2018, Page 17. 2, 13
[CAI 15] Cain V., Thijs L., Van Humbeeck J., Van Hooreweder B., Knutsen R.
Crack propagation and fracture toughness of Ti6Al4V alloy produced by selective laser melt-
ing. Additive Manufacturing, vol. 5, 2015, p. 68–76. 82
[CAL 14] Calignano F.
Design optimization of supports for overhanging structures in aluminum and titanium alloys
by selective laser melting. Materials & Design, vol. 64, 2014, p. 203–213. 13
[CAR 19] Carrion P. E., Soltani-Tehrani A., Phan N., Shamsaei N.
Powder Recycling Effects on the Tensile and Fatigue Behavior of Additively Manufactured
Ti-6Al-4V Parts. JOM, vol. 71, no 3, 2019, p. 963–973. 16, 19

REFERENCES | 213

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[CAZ 05] Cazals F., Pouget M.
Estimating differential quantities using polynomial fitting of osculating jets. Computer Aided
Geometric Design, vol. 22, no 2, 2005, p. 121–146. 51
[CHA 15] Chan K. S.
Characterization and analysis of surface notches on Ti-alloy plates fabricated by additive
manufacturing techniques. Surface Topography: Metrology and Properties, vol. 3, no 4, 2015.
13
[CHA 17] Chastand V.
Etude du comportement mécanique et des mécanismes d’endommagement de pièces mé-
talliques réalisées par fabrication additive. Doctoral thesis, Centrale Lille, 2017. 19
[CHA 18] Chastand V.
Comparative study of fatigue properties of Ti-6Al-4V specimens built by electron beam melt-
ing (EBM) and selective laser melting (SLM). Materials Characterization, vol. 143, 2018,
p. 76–81. 16, 18, 21
[CHA 22] Charles A., Bayat M., Elkaseer A., Thijs L., Hattel J. H., Scholz S.
Elucidation of dross formation in laser powder bed fusion at down-facing surfaces:
Phenomenon-oriented multiphysics simulation and experimental validation. Additive Man-
ufacturing, vol. 50, 2022. 13, 34
[CHI 21] Childerhouse T., Hernandez-Nava E., Tapoglou N.
The influence of finish machining depth and hot isostatic pressing on defect distribution
and fatigue behaviour of selective electron beam melted Ti-6Al-4V. International Journal of
Fatigue, vol. 147, 2021, Page 106169. 19
[COE 14] Coeurjolly D., Lachaud J.-O., Levallois J.
Multigrid convergent principal curvature estimators in digital geometry. Computer Vision
and Image Understanding, vol. 129, 2014, p. 27–41. 51, 55
[COE 16] Coeurjolly D., Foare M., Gueth P., Lachaud J.-O.
Piecewise smooth reconstruction of normal vector field on digital data. Computer Graphics
Forum, vol. 35, no 7, 2016, p. 157–167. 191, 192
[CUT 22] Cutolo A., Elangeswaran C., Muralidharan G. K., Van Hooreweder B.
On the role of building orientation and surface post-processes on the fatigue life of Ti-6Al-4V
coupons manufactured by laser powder bed fusion. Materials Science and Engineering: A,
vol. 840, 2022. 131
[DEP 18] DePond P. J., Guss G., Ly S., Calta N. P., Deane D., Khairallah S.,
Matthews M. J.
In situ measurements of layer roughness during laser powder bed fusion additive manufac-
turing using low coherence scanning interferometry. Materials & Design, vol. 154, 2018,
p. 347–359. 13, 15
[DGt 22] DGtal Team
. DGtal, 2022. 51
[DIA 22] Dias Corpa Tardelli J., Duarte Firmino A. C., Ferreira I., Cândido
Dos Reis A.
Influence of the roughness of dental implants obtained by additive manufacturing on os-
teoblastic adhesion and proliferation: A systematic review. Heliyon, vol. 8, no 12, 2022.
39

214 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[DUR 79] Duradzhi V., Bryantsev I., Tokarov A.
Investigation of erosion of the anode under the action of an electrolytic plasma on it. Elek-
tronnaya Obrabotka Materialov, , 1979. 131
[EDW 14] Edwards P., Ramulu M.
Fatigue performance evaluation of selective laser melted Ti6Al4V. Materials Science, , 2014,
Page 11. 21
[EST 96] Ester M., Kriegel H.-P., Sander J., Xu X.
A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise.
Proceedings of the 2nd International Conference on Knowledge Discovery and Data Mining
(KDD-96), , 1996. 63, 155
[FEL 84] Feldkamp L. A., Davis L. C., Kress J. W.
Practical cone-beam algorithm. Journal of the Optical Society of America A, vol. 1, no 6,
1984. 41
[FEN 21] Feng S., Kamat A. M., Sabooni S., Pei Y.
Experimental and numerical investigation of the origin of surface roughness in laser powder
bed fused overhang regions. Virtual and Physical Prototyping, vol. 16, 2021. 13
[FIN 02] Findlay S., Harrison N.
Why aircraft fail. Materials Today, vol. 5, no 11, 2002, p. 18–25. 2
[FOR 16] de Formanoir C., Suard M., Dendievel R., Martin G., Godet S.
Improving the mechanical efficiency of electron beam melted titanium lattice structures by
chemical etching. Additive Manufacturing, vol. 11, 2016, p. 71–76. 34
[FOX 18] Fox J. C., Kim F., Reese Z., Evans C.
Complementary use of optical metrology and X-ray computed tomography for surface finish
and defect detection in laser powder bed fusion additive manufacturing. American Society for
Precision Engineering, vol. 69, 2018, p. 195–200. 35
[GAI 22] Gaillard Q., Cazottes S., Boulnat X., Dancette S., Desrayaud C.
Microstructure, texture and mechanical properties with raw surface states of Ti-6Al-4V parts
built by L-PBF. Procedia CIRP, vol. 108, 2022, p. 698–703. 82, 83
[GAI 23a] Gaillard Q.
Optimisation des propriétés de composants en TA6V élaborés par Laser Beam Melting. Doc-
toral thesis, Ecole des Mines de St Etienne, 2023. 4, 71, 72, 76, 77, 78, 79, 80, 81, 84, 101,
120, 207
[GAI 23b] Gaillard Q., Boulnat X., Cazottes S., Dancette S., Desrayaud C.
Strength/ductility trade-off of Laser Powder Bed Fusion Ti-6Al-4V: Synergetic effect of alpha-
case formation and microstructure evolution upon heat treatments. Additive Manufacturing,
vol. 76, 2023, Page 103772. 76, 101, 107, 120, 121, 122, 128
[GAL 17] Galarraga H., Warren R. J., Lados D. A., Dehoff R. R., Kirka M. M.
Fatigue crack growth mechanisms at the microstructure scale in as-fabricated and heat treated
Ti-6Al-4V ELI manufactured by electron beam melting (EBM). Engineering Fracture Me-
chanics, vol. 176, 2017, p. 263–280. 107
[Gen 18] General Electric
. New manufacturing milestone: 30,000 additive fuel nozzles | GE Additive, 2018. 1
[GIL 22] Gillham B., Yankin A., Shipley H., McNamara F., Tomonto C., O’Donnell
G., Trimble D., Yin S., Taylor D., Lupoi R.

REFERENCES | 215

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
The Analysis of Small-Scale Notches on the Fatigue Performance of SLM Ti-6Al-4V; A Theory
of Critical Distances Approach. Key Engineering Materials, vol. 926, 2022, p. 250–267. 26
[GOC 19] Gockel J., Sheridan L., Koerper B., Whip B.
The influence of additive manufacturing processing parameters on surface roughness and
fatigue life. International Journal of Fatigue, vol. 124, 2019, p. 380–388. 32
[GON 15] Gong H., Rafi K., Gu H., Janaki Ram G., Starr T., Stucker B.
Influence of defects on mechanical properties of Ti6Al4V components produced by selective
laser melting and electron beam melting. Materials & Design, vol. 86, 2015, p. 545–554. 18,
22
[GO 16] Goïc G. L., Bigerelle M., Samper S., Favrelière H., Pillet M.
Multiscale roughness analysis of engineering surfaces: A comparison of methods for the in-
vestigation of functional correlations. Mechanical Systems and Signal Processing, vol. 66-67,
2016, p. 437–457. 69, 193
[GRA 23] Grazia Guerra M., Lavecchia F.
Measurement of additively manufactured freeform artefacts: The influence of surface texture
on measurements carried out with optical techniques. Measurement, vol. 209, 2023. 35, 46
[GRE 17] Greitemeier D., Palm F., Syassen F., Melz T.
Fatigue performance of additive manufactured TiAl6V4 using electron and laser beam melting.
International Journal of Fatigue, vol. 94, 2017, p. 211–217. 21, 22, 82
[GUN 18] Gunenthiram V.
Compréhension de la formation de porosités en fabrication additive (LBM). Analyse expéri-
mentale de linteraction laser - lit de poudre - bain liquide. Doctoral thesis, Ecole Nationale
Supérieure d’Arts et Métiers, 2018. 13, 15
[HAI 13] Haitjema H.
Uncertainty estimation of 2.5-D roughness parameters obtained by mechanical probing. In-
ternational Journal of Precision Technology, vol. 3, no 4, 2013, p. 403–412. 35
[HAM 20] Hamidi Nasab M., Romano S., Gastaldi D., Beretta S., Vedani M.
Combined effect of surface anomalies and volumetric defects on fatigue assessment of AlSi7Mg
fabricated via laser powder bed fusion. Additive Manufacturing, vol. 34, 2020. 26, 35
[HEN 18] Henning Z., Falko B.-H.
Surface Finish of Additively Manufactured Parts using Plasma Electrolytic Polishing.
WCMNM 2018 World Congress on Micro and Nano Manufacturing, Slovenia, 2018 Research
Publishing Services, p. 385–388. 131
[HU 20a] Hu Y., Wu S., Withers P., Zhang J., Bao H., Fu Y., Kang G.
The effect of manufacturing defects on the fatigue life of selective laser melted Ti-6Al-4V
structures. Materials & Design, vol. 192, 2020. 15
[HU 20b] Hu Y., Wu S., Wu Z., Zhong X., Ahmed S., Karabal S., Xiao X., Zhang H.,
Withers P.
A new approach to correlate the defect population with the fatigue life of selective laser melted
Ti-6Al-4V alloy. International Journal of Fatigue, vol. 136, 2020. 28
[HUA 21] Huang Y., Wang C., Ding F., Yang Y., Zhang T., He X., Zheng L., Li N.
Principle, process, and application of metal plasma electrolytic polishing: a review. The
International Journal of Advanced Manufacturing Technology, vol. 114, 2021. 131, 132, 133

216 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[ING 13] Inglis C.
Stresses in plates due to the presence of cracks and sharp corners. Spie Milestone series MS,
vol. 137, 1913, p. 3–17. 24, 54
[Int 12] International Organisation for Standardization
. ISO 25178-3:2012 Geometrical product specifications (GPS) - Surface texture: Areal - Part
3: Specification operators, 2012. 44, 46, 47
[Int 15] International Organisation for Standardization
. ISO 16610-60:2015 Geometrical product specification (GPS) - Filtration - Part 60: Linear
areal filters - Basic concepts, 2015. 46
[Int 16] International Organisation for Standardization
. ISO 16610-28:2016 Geometrical product specifications (GPS) Filtration Part 28: Profile
filters: End effects, 2016. 47
[Int 21] International Organisation for Standardization
. ISO 25178-2:2012 Geometrical product specifications (GPS) Surface texture: Areal Part
2: Terms, definitions and surface texture parameters, 2021. 31, 32, 49, 161, 191
[IRV 74] Irving P. E., Beevers C. J.
The effect of air and vacuum environments on fatigue crack growth rates in Ti-6Al-4V. Met-
allurgical Transactions, vol. 5, no 2, 1974, p. 391–398. 17
[JAI 16] Jain P., Tyagi V.
A survey of edge-preserving image denoising methods. Information Systems Frontiers, vol. 18,
no 1, 2016, p. 159–170. 41
[JAN 12] Janecki D.
Edge effect elimination in the recursive implementation of Gaussian filters. Precision Engi-
neering, vol. 36, no 1, 2012, p. 128–136. 47
[JIA 10] Jiang X., Zhang X., Scott P. J.
Template matching of freeform surfaces based on orthogonal distance fitting for precision
metrology. Measurement Science and Technology, vol. 21, no 4, 2010. 44
[JIA 11] Jiang X., Cooper P., Scott P. J.
Freeform surface filtering using the diffusion equation. Proceedings of the Royal Society A:
Mathematical, Physical and Engineering Sciences, vol. 467, no 2127, 2011, p. 841–859. 44
[JIA 20a] Jiang X. J., Scott P. J.
Advanced metrology: freeform surfaces. Academic Press, London, 2020. 35, 44, 46, 49, 69,
193
[JIA 20b] Jiang X. J., Scott P. J.
Free-form surface filtering using wavelets and multiscale decomposition. Advanced Metrology,
p. 195–246 Elsevier, 2020. 44
[JUN 21] Junet A.
Étude tridimensionnelle de la propagation en fatigue de fissures internes dans les matériaux
métalliques. Doctoral thesis, Institut National des Sciences Appliquées de Lyon, 2021. 17
[KAH 17] Kahlin M., Ansell H., Moverare J.
Fatigue behaviour of notched additive manufactured Ti6Al4V with as-built surfaces. Inter-
national Journal of Fatigue, vol. 101, 2017, p. 51–60. 16, 21, 22
[KAN 18] Kantzos C. A., Cunningham R. W., Tari V., Rollett A. D.
Characterization of metal additive manufacturing surfaces using synchrotron X-ray CT and

REFERENCES | 217

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
micromechanical modeling. Computational Mechanics, vol. 61, no 5, 2018, p. 575–580. 13,
54
[KAR 23] Karimialavijeh H., Ghasri-Khouzani M., Das A., Pröebstle M., Martin .
Effect of laser contour scan parameters on fatigue performance of A20X fabricated by laser
powder bed fusion. International Journal of Fatigue, vol. 175, 2023. 12
[KAS 15] Kasperovich G., Hausmann J.
Improvement of fatigue resistance and ductility of TiAl6V4 processed by selective laser melt-
ing. Journal of Materials Processing Technology, vol. 220, 2015, p. 202–214. 16, 18, 21, 22,
82
[KAS 17] Kastner J., Salaberger D., Heinzl C., Gusenbauer C., Rao G.
High Resolution X-ray Computed Tomography for Non-destructive Characterization and In-
situ Investigations. 2017. 4
[KER 12] Kerckhofs G., Pyka G., Moesen M., Schrooten J., Wevers M.
High-resolution micro-CT as a tool for 3D surface roughness measurement of 3D additive
manufactured porous structures. Proc iCT, , 2012. 35
[KHA 21] Khan H. M., Karabulut Y., Kitay O., Kaynak Y., Jawahir I. S.
Influence of the post-processing operations on surface integrity of metal components pro-
duced by laser powder bed fusion additive manufacturing: a review. Machining Science and
Technology, vol. 25, no 1, 2021, p. 118–176. 131
[KHR 21] Khrapov D., Kozadayeva M., Manabaev K., Panin A., Sjöström W.,
Koptyug A., Mishurova T., Evsevleev S., Meinel D., Bruno G., Cheneler D.,
Surmenev R., Surmeneva M.
Different Approaches for Manufacturing Ti-6Al-4V Alloy with Triply Periodic Minimal Sur-
face Sheet-Based Structures by Electron Beam Melting. Materials, vol. 14, no 17, 2021.
52
[KNU 93] Knutsson H., Westin C.-F.
Normalized and differential convolution. Proceedings of IEEE Conference on Computer Vision
and Pattern Recognition, New York, NY, USA, 1993 IEEE Comput. Soc. Press, p. 515–523.
47, 203
[LAC 17] Lachaud J.-O., Coeurjolly D., Levallois J.
Robust and Convergent Curvature and Normal Estimators with Digital Integral Invariants.
Najman L., Romon P., Eds., Modern Approaches to Discrete Curvature, p. 293–348 Springer
International Publishing, Cham, 2017. 51
[LAD 16] Ladewig A., Schlick G., Fisser M., Schulze V., Glatzel U.
Influence of the shielding gas flow on the removal of process by-products in the selective laser
melting process. Additive Manufacturing, vol. 10, 2016, p. 1–9. 13
[LAR 18] Larrosa N., Wang W., Read N., Loretto M., Evans C., Carr J., Tradowsky
U., Attallah M., Withers P.
Linking microstructure and processing defects to mechanical properties of selectively laser
melted AlSi10Mg alloy. Theoretical and Applied Fracture Mechanics, vol. 98, 2018, p. 123–
133. 18
[LAR 23] Larguier A.
Influence de l’alpha-case sur le comportement en fatigue de pièces en Ti-6Al-4V élaborées par
fusion laser sur lit de poudre. Rapport de stage, 2023, INSA Lyon. 129

218 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[LE 21] Le V.-D., Pessard E., Morel F., Prigent S.
Effect of the surface roughness on the high cycle fatigue behaviour of Ti-6Al-4V alloy obtained
by additive manufacturing process. Engineering Integrity, vol. 53, 2021. 33
[LEE 20] Lee S., Pegues J. W., Shamsaei N.
Fatigue behavior and modeling for additive manufactured 304L stainless steel: The effect of
surface roughness. International Journal of Fatigue, vol. 141, 2020, Page 105856. 33
[LES 22] Lesseur J., Tranchand B., Mancier T., Montauzier A., Larignon C., Pe-
rusin S.
On the use of X-ray microtomography to control artificial defect geometries produced by
metal additive manufacturing. Nondestructive Testing and Evaluation, vol. 37, no 5, 2022,
p. 611–630. 16
[LEU 13] Leuders S., Thöne M., Riemer A., Niendorf T., Tröster T., Richard H.,
Maier H.
On the mechanical behaviour of titanium alloy TiAl6V4 manufactured by selective laser melt-
ing: Fatigue resistance and crack growth performance. International Journal of Fatigue,
vol. 48, 2013, p. 300–307. 22
[LEV 15] Levallois J., Coeurjolly D., Lachaud J.-O.
Scale-space feature extraction on digital surfaces. Computers & Graphics, vol. 51, 2015,
p. 177–189. 69, 191, 193
[LI 19] Li P., Warner D., Pegues J., Roach M., Shamsaei N., Phan N.
Investigation of the mechanisms by which hot isostatic pressing improves the fatigue perfor-
mance of powder bed fused Ti-6Al-4V. International Journal of Fatigue, vol. 120, 2019,
p. 342–352. 19
[LI 22] Li Z., Li H., Yin J., Li Y., Nie Z., Li X., You D., Guan K., Duan W., Cao L.,
Wang D., Ke L., Liu Y., Zhao P., Wang L., Zhu K., Zhang Z., Gao L., Hao L.
A Review of Spatter in Laser Powder Bed Fusion Additive Manufacturing: In Situ Detection,
Generation, Effects, and Countermeasures. Micromachines, vol. 13, no 8, 2022, Page 1366.
13
[Lie 17] Liebherr
. Technologies for the future, 2017. 1
[LIF 21] Lifton J. J., Liu Y., Tan Z. J., Mutiargo B., Goh X. Q., Malcolm A. A.
Internal surface roughness measurement of metal additively manufactured samples via x-ray
CT: the influence of surrounding material thickness. Surface Topography: Metrology and
Properties, vol. 9, no 3, 2021. 35
[LIN 23] Lin P., Wang M., Trofimov V. A., Yang Y., Song C.
Research on the Warping and Dross Formation of an Overhang Structure Manufactured by
Laser Powder Bed Fusion. Applied Sciences, vol. 13, no 6, 2023, Page 3460. 13
[LIU 14] Liu Q. C., Elambasseril J., Sun S. J., Leary M., Brandt M., Sharp P. K.
The Effect of Manufacturing Defects on the Fatigue Behaviour of Ti-6Al-4V Specimens Fab-
ricated Using Selective Laser Melting. Advanced Materials Research, vol. 891-892, 2014,
p. 1519–1524. 18
[LIU 17] Liu Y., Zhang J., Li S.-J., Hou W.-T., Wang H., Xu Q.-S., Hao Y.-L., Yang
R.
Effect of HIP Treatment on Fatigue Crack Growth Behavior of Ti6Al4V Alloy Fabricated by

REFERENCES | 219

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Electron Beam Melting. Acta Metallurgica Sinica (English Letters), vol. 30, no 12, 2017,
p. 1163–1168. 107
[LIU 19] Liu S., Shin Y. C.
Additive manufacturing of Ti6Al4V alloy: A review. Materials & Design, vol. 164, 2019. 107
[LIU 20] Liu W., Chen C., Shuai S., Zhao R., Liu L., Wang X., Hu T., Xuan W., Li C.,
Yu J., Wang J., Ren Z.
Study of pore defect and mechanical properties in selective laser melted Ti6Al4V alloy based
on X-ray computed tomography. Materials Science and Engineering: A, vol. 797, 2020. 15
[LOR 87] Lorensen W. E., Cline H. E.
Marching cubes: A high resolution 3D surface construction algorithm. ACM SIGGRAPH
Computer Graphics, vol. 21, no 4, 1987, p. 163–169. 85
[LOU 13a] Lou S., Jiang X., Scott P. J.
Geometric computation theory for morphological filtering on freeform surfaces. Proceedings
of the Royal Society A: Mathematical, Physical and Engineering Sciences, , 2013. 46
[LOU 13b] Lou S., Zeng W.-H., Jiang X.-Q., Scott P. J.
Robust Filtration Techniques in Geometrical Metrology and Their Comparison. International
Journal of Automation and Computing, , 2013, p. 1–8. 44
[LOU 19a] Lou S., Jiang X., Sun W., Zeng W., Pagani L., Scott P.
Characterisation methods for powder bed fusion processed surface topography. Precision
Engineering, vol. 57, 2019, p. 1–15. 46
[LOU 19b] Lou S., Pagani L., Zeng W., Ghori M. U., Jiang X., Scott P. J.
Surface texture evaluation of additively manufactured metallic cellular scaffolds for acetabular
implants using X-ray computed tomography. Bio-Design and Manufacturing, vol. 2, no 2,
2019, p. 55–64. 35, 44
[LOU 20] Lou S., Pagani L., Zeng W., Jiang X., Scott P.
Watershed segmentation of topographical features on freeform surfaces and its application to
additively manufactured surfaces. Precision Engineering, vol. 63, 2020, p. 177–186. 191, 192
[MAH 22] Mahtabi M., Yadollahi A., Stokes R., Morgan-Barnes C., Doude H., Bian
L.
Effect of Powder Reuse on Microstructural and Fatigue Properties of Ti-6Al-4V Fabricated
via Directed Energy Deposition. Proceedings of the 33rd Annual International Solid Freeform
Fabrication Symposium, , 2022. 72
[MAJ 10] Majzoobi G. H., Daemi N.
The effects of notch geometry on fatigue life using notch sensitivity factor. Transactions of
the Indian Institute of Metals, vol. 63, no 2-3, 2010, p. 547–552. 25
[MAL 21] Maleki E., Bagherifard S., Bandini M., Guagliano M.
Surface post-treatments for metal additive manufacturing: Progress, challenges, and oppor-
tunities. Additive Manufacturing, vol. 37, 2021, Page 101619. 131
[MAN 16] Mancisidor A. M.
Reduction of the Residual Porosity in Parts Manufactured by Selective Laser Melting Using
Skywriting and High Focus Offset Strategies. Physics Procedia, , 2016, Page 10. 12
[MAR 19] Martin A. A., Calta N. P., Khairallah S. A., Wang J., Depond P. J.,
Fong A. Y., Thampy V., Guss G. M., Kiss A. M., Stone K. H., Tassone C. J.,
Nelson Weker J., Toney M. F., van Buuren T., Matthews M. J.

220 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Dynamics of pore formation during laser powder bed fusion additive manufacturing. Nature
Communications, vol. 10, no 1, 2019, Page 1987. 12
[MAS 18] Masuo H., Tanaka Y., Morokoshi S., Yagura H., Uchida T., Yamamoto Y.,
Murakami Y.
Influence of defects, surface roughness and HIP on the fatigue strength of Ti-6Al-4V manufac-
tured by additive manufacturing. International Journal of Fatigue, vol. 117, 2018, p. 163–179.
19, 20, 22, 26, 27, 28, 29
[MAT 03] Matsunaga H., Murakami Y., Kubota M., Lee J.-H.
Fatigue Strength of Ti-6Al-4V Alloys Containing Small Artificial Defects. Journal of the
Society of Materials Science, Japan, vol. 52, 2003, p. 263–269. 27
[MCB 20] McBride J. W., Cross K. J.
The surface area and localised 3D roughness of a highly structured surface using X-Ray
Computed tomography (XCT). Metrology Letters, , 2020. 35
[MEA 17] Mearian L.
. Boeing turns to 3D-printed parts to save millions on its 787 Dreamliner, 2017. 1
[MIA 21] Mian M. J., Razmi J., Ladani L.
Defect Analysis and Fatigue Strength Prediction of As-built Ti6Al4V Parts, Produced Using
Electron Beam Melting (EBM) AM Technology. Materialia, , 2021. 32
[MIL 22] Milhomme S.
Influence du procédé sur les propriétés mécaniques et microstructurales de pièces en Ti-6Al-
4V issues de fabrication additive (LMD et SLM). Doctoral thesis, Université de Bordeaux,
2022. 16
[MIN 14] Mingareev I., Bonhoff T., El-Sherif A. F., Meiners W., Kelbassa I., Bier-
mann T., Richardson M.
Femtosecond laser post-processing of metal parts produced by laser additive manufacturing.
J. Laser Appl., vol. 25, no 5, 2014, Page 5. 35
[MUR 83] Murakami Y., Masahiro E.
Quantitative evaluation of fatigue strength of metals containing various small defects or cracks.
Engineering Fracture Mechanics, vol. 17, no 1, 1983, p. 1–15. 26
[MUR 02] Murakami Y.
Metal Fatigue - Effects of Small Defects and Nonmetallic Inclusions. Elsevier, 1st édition,
2002. 27, 31, 33, 64
[NAJ 19] Najmon J. C., Raeisi S., Tovar A.
Review of additive manufacturing technologies and applications in the aerospace industry.
Additive Manufacturing for the Aerospace Industry, p. 7–31 Elsevier, 2019. 1
[NAK 19] Nakatani M., Masuo H., Tanaka Y., Murakami Y.
Effect of Surface Roughness on Fatigue Strength of Ti-6Al-4V Alloy Manufactured by Additive
Manufacturing. Procedia Structural Integrity, vol. 19, 2019, p. 294–301. 2, 31
[NAV 22] Navickait K., Nestler K., Böttger-Hiller F., Matias C., Diskin A., Golan
O., Garkun A., Strokin E., Biletskiy R., Safranchik D., Zeidler H.
Efficient polishing of additive manufactured titanium alloys. Procedia CIRP, vol. 108, 2022,
p. 346–351. 131, 132, 133
[NGU 22] Nguyen H. D., Pramanik A., Basak A., Dong Y., Prakash C., Debnath S.,
Shankar S., Jawahir I., Dixit S., Buddhi D.

REFERENCES | 221

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
A critical review on additive manufacturing of Ti-6Al-4V alloy: microstructure and mechanical
properties. Journal of Materials Research and Technology, vol. 18, 2022. 82, 101
[NUT 19] Nutal N.
. Finition de surface de pièces produites par fabrication additive, 2019. 131
[OGD 55] Ogden H., Jaffee R.
The effect of carbon, oxygen, and nitrogen on the mechanical properties of titanium and
titanium alloys. rapport no TML-20, 4370612, 1955. 118
[OTS 79] Otsu N.
A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems,
Man, and Cybernetics, , 1979. 41
[PAG 18] Pagani L., Lou S., Zeng W., Jiang X., Scott P. J.
Effect of the form estimation on the areal texture parameters for X-ray computed tomography
measurement. Proc iCT 2018, , 2018. 35, 44
[PAL 18] Pal S., Lojen G., Kokol V., Drstvensek I.
Evolution of metallurgical properties of Ti-6Al-4V alloy fabricated in different energy densities
in the Selective Laser Melting technique. Journal of Manufacturing Processes, vol. 35, 2018,
p. 538–546. 15
[PAR 61] Paris P., Gomez M., Anderson W.
A rational analytic theory of fatigue. The Trend in Engineering, , no 13, 1961, p. 9–14. 26
[PAU 23] Pauzon C., Raza A., Hanif I., Dubiez-Le Goff S., Moverare J., Hryha E.
Effect of layer thickness on spatter properties during laser powder bed fusion of Ti6Al4V.
Powder Metallurgy, vol. 66, no 4, 2023, p. 333–342. 13
[PEG 18] Pegues J., Roach M., Scott Williamson R., Shamsaei N.
Surface roughness effects on the fatigue strength of additively manufactured Ti-6Al-4V. In-
ternational Journal of Fatigue, vol. 116, 2018, p. 543–552. 108, 114
[PEG 19] Pegues J. W., Shamsaei N., Roach M. D., Williamson R. S.
Fatigue life estimation of additive manufactured parts in the as-built surface condition. Ma-
terial Design & Processing Communications, vol. 1, no 3, 2019. 23, 33
[PER 18] Persenot T.
Fatigue of Ti-6Al-4V thin parts made by electron beam melting. Doctoral thesis, Université
de Lyon, Lyon, 2018. 21, 40
[PER 19] Persenot T., Burr A., Martin G., Buffiere J.-Y., Dendievel R., Maire E.
Effect of build orientation on the fatigue properties of as-built Electron Beam Melted Ti-
6Al-4V alloy. International Journal of Fatigue, vol. 118, 2019, p. 65–76. 2, 13, 14, 34, 40,
54
[PER 20] Persenot T., Burr A., Dendievel R., Buère J.-Y., Maire E., Lachambre J.,
Martin G.
Fatigue performances of chemically etched thin struts built by selective electron beam melting:
Experiments and predictions. Materialia, , 2020. 3, 12, 27, 29, 30, 62, 63, 64, 65, 67, 70
[PES 20] Pessard E.
Contribution à la prise en compte des effets du procédé de mise en forme sur la tenue en fatigue
des matériaux métalliques. Habilitation à Diriger des Recherches, Université d’Angers, 2020.
13, 14

222 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[PET 59] Peterson R.
Notch sensitivity. Metal Fatigue, , 1959, p. 293–306. 25
[PLE 20a] du Plessis A., Beretta S.
Killer notches: The effect of as-built surface roughness on fatigue failure in AlSi10Mg produced
by laser powder bed fusion. Additive Manufacturing, vol. 35, 2020. 2, 23
[PLE 20b] du Plessis A., Macdonald E.
Hot isostatic pressing in metal additive manufacturing_ X-ray tomography reveals details of
pore closure. Additive Manufacturing, vol. 34, 2020. 19, 54, 62
[PLE 20c] du Plessis A., Yadroitsava I., Yadroitsev I.
Effects of defects on mechanical properties in metal additive manufacturing: A review focusing
on X-ray tomography insights. Materials & Design, vol. 187, 2020. 18
[POL 14] Polasik A.
Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in
the Graduate School of The Ohio State University. Doctoral thesis, The Ohio State University,
OH, USA, 2014. 107
[POL 22] Polley C., Radlof W., Hauschulz F., Benz C., Sander M., Seitz H.
Morphological and mechanical characterisation of three-dimensional gyroid structures fabri-
cated by electron beam melting for the use as a porous biomaterial. Journal of the Mechanical
Behavior of Biomedical Materials, vol. 125, 2022. 52
[POM 19] Pomberger S., Stoschka M., Leitner M.
Cast surface texture characterisation via areal roughness. Precision Engineering, vol. 60,
2019, p. 465–481. 44
[POM 20] Pomberger S., Stoschka M., Aigner R., Leitner M., Ehart R.
Areal fatigue strength assessment of cast aluminium surface layers. International Journal of
Fatigue, vol. 133, 2020. 33
[PYK 13] Pyka G., Kerckhofs G., Papantoniou I., Speirs M., Schrooten J., Wevers
M.
Surface Roughness and Morphology Customization of Additive Manufactured Open Porous
Ti6Al4V Structures. Materials, vol. 6, no 10, 2013, p. 4737–4757. 13
[RA 13] Ra H. K., Karthik N. V., Gong H., Starr T. L., Stucker B. E.
Microstructures and Mechanical Properties of Ti6Al4V Parts Fabricated by Selective Laser
Melting and Electron Beam Melting. Journal of Materials Engineering and Performance, ,
2013, Page 13. 22
[REK 15] Rekedal K., Liu D.
Fatigue Life of Selective Laser Melted and Hot Isostatically Pressed Ti-6Al-4v Absent of
Surface Machining. 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and
Materials Conference, Kissimmee, Florida, 2015 American Institute of Aeronautics and As-
tronautics. 22
[ROM 17] Romano S., Brandão A., Gumpinger J., Gschweitl M., Beretta S.
Qualification of AM parts: Extreme value statistics applied to tomographic measurements.
Materials & Design, vol. 131, 2017, p. 32–48. 27
[ROM 18] Romano S., Brückner-Foit A., Brandão A., Gumpinger J., Ghidini T.,
Beretta S.
Fatigue properties of AlSi10Mg obtained by additive manufacturing: Defect-based modelling

REFERENCES | 223

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
and prediction of fatigue strength. Engineering Fracture Mechanics, vol. 187, 2018, p. 165–
189. 23
[SAN 20] Sanaei N., Fatemi A.
Analysis of the effect of surface roughness on fatigue performance of powder bed fusion additive
manufactured metals. Theoretical and Applied Fracture Mechanics, vol. 108, 2020. 18
[SAN 21] Sanaei N., Fatemi A.
Defects in additive manufactured metals and their effect on fatigue performance: A state-of-
the-art review. Progress in Materials Science, vol. 117, 2021. 11, 34
[SAV 07] Savio E., De Chiffre L., Schmitt R.
Metrology of freeform shaped parts. CIRP Annals, vol. 56, no 2, 2007, p. 810–835. 44
[SCH 70] Schoen A. H.
Infinite periodic minimal surfaces without self-intersections. rapport no D-5541, 1970, NASA.
40
[SCH 09] Schijve J.
Fatigue of structures and materials. Springer, Dordrecht, 2nd édition, 2009. 10, 11, 23, 24,
27, 118, 130, 167
[SEI 17] Seifi M., Gorelik M., Waller J., Hrabe N., Shamsaei N., Daniewicz S.,
Lewandowski J. J.
Progress Towards Metal Additive Manufacturing Standardization to Support Qualification
and Certification. JOM, vol. 69, no 3, 2017, p. 439–455. 1
[SEO 21] Seo B., Park H.-K., Kim H. G., Kim W. R., Park K.
Corrosion behavior of additive manufactured CoCr parts polished with plasma electrolytic
polishing. Surface and Coatings Technology, vol. 406, 2021. 131
[SHA 21] Shahpaski M., Sapaico L. R., Süsstrunk S.
Surface roughness estimation using structured light projection. Electronic Imaging, vol. 33,
no 5, 2021. 35
[SIL 98] Silverman B. W.
Density estimation for statistics and data analysis. No 26 Monographs on statistics and
applied probability Chapman & Hall/CRC, Boca Raton, 1998. 86
[SKR 19] Skrodzki M.
. The k-d tree data structure and a proof for neighborhood computation in expected loga-
rithmic time, 2019. Preprint available at http://arxiv.org/abs/1903.04936. 197
[SOR 21] Soro N., Saintier N., Merzeau J., Veidt M., Dargusch M. S.
Quasi-static and fatigue properties of graded Ti6Al4V lattices produced by Laser Powder Bed
Fusion (LPBF). Additive Manufacturing, vol. 37, 2021. 34, 40
[STE 23] Steinhilber F.
3D roughness computation from XCT data - Data and Python & ImageJ implementations.
Zenodo repository, , 2023. Available at https://doi.org/10.5281/zenodo.10276253. 43, 45, 49,
185, 197
[STO 08] Stock S. R.
Recent advances in X-ray microtomography applied to materials. International Materials
Reviews, vol. 53, no 3, 2008, p. 129–181. 57

224 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[TAM 17] Tammas-Williams S., Withers P. J., Todd I., Prangnell P. B.
The Influence of Porosity on Fatigue Crack Initiation in Additively Manufactured Titanium
Components. Scientific Reports, vol. 7, no 1, 2017. 18, 23, 27, 62
[TAS 02] Tasdizen T., Whitaker R., Burchard P., Osher S.
Geometric surface smoothing via anisotropic diffusion of normals. Boston, MA, USA, 2002
IEEE, p. 125–132. 44, 60
[TAU 95] Taubin G.
Estimating the tensor of curvature of a surface from a polyhedral approximation. Proceedings
of IEEE International Conference on Computer Vision, Cambridge, MA, USA, 1995 p. 902–
907. 50
[TAY 91] Taylor D., Clancy O. M.
THE FATIGUE PERFORMANCE OF MACHINED SURFACES. Fatigue & Fracture of
Engineering Materials and Structures, vol. 14, no 2-3, 1991, p. 329–336. 32
[TAY 08] Taylor D.
The theory of critical distances. Engineering Fracture Mechanics, vol. 75, no 7, 2008, p. 1696–
1705. 25, 26
[THO 18] Thompson A.
Surface texture measurement of metal additively manufactured parts by X-ray computed
tomography. Doctoral thesis, University of Nottingham, 2018. 35
[THO 19] Thomas M., Davoine C., Drawin S.
Fabrication additive en aéronautique et en spatial. Techniques de l’ingénieur, , 2019. 1, 2, 3
[TIL 23] Tilton M., Borjali A., Griffis J. C., Varadarajan K. M., Manogharan G. P.
Fatigue properties of Ti-6Al-4V TPMS scaffolds fabricated via laser powder bed fusion. Man-
ufacturing Letters, vol. 37, 2023, p. 32–38. 52
[TOW 16] Townsend A.
Surface texture metrology for metal additive manufacturing: a review. Precision Engineering,
vol. 46, 2016, p. 34–47. 35
[TUB 23] Tubei V., Toda H., Ketanond W., Fujihara H., Takakuwa O., Takeuchi A.,
Uesugi M.
Direct observation of three-dimensional short fatigue crack closure behavior in Ti-6Al-4V
alloy using ultra-high-resolution X-ray microtomography. International Journal of Fatigue,
vol. 168, 2023. 10
[VAF 21] Vafadar A., Guzzomi F., Rassau A., Hayward K.
Advances in Metal Additive Manufacturing: A Review of Common Processes, Industrial Ap-
plications, and Current Challenges. Applied Sciences, vol. 11, no 3, 2021, Page 1213. 1
[Val 00] Vallellano, Navarro, Dominguez
Fatigue crack growth threshold conditions at notches. Part I: theory. Fracture of Engineering
Materials and Structures, vol. 23, no 2, 2000, p. 113–121. 25
[VAN 12] Van Hooreweder B., Boonen R., Moens D., Kruth J.-P., Sas P.
On the Determination of Fatigue Properties of Ti6Al4V Produced by Selective Laser Melting.
Honolulu, Hawaii, 2012 American Institute of Aeronautics and Astronautics. 82
[VAN 22] Vanderesse N., Bocher P., Nuño N., Yánez A., Hof L. A.
On the characterization of roughness and geometrical irregularities of additively manufactured
single titanium-alloy struts. Additive Manufacturing, vol. 54, 2022. 41

REFERENCES | 225

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[VAY 19] Vayssette B., Saintier N., Brugger C., El May M., Pessard E.
Numerical modelling of surface roughness effect on the fatigue behavior of Ti-6Al-4V obtained
by additive manufacturing. International Journal of Fatigue, vol. 123, 2019, p. 180–195. 21,
22, 26
[VAY 20] Vayssette B.
Comportement en fatigue de pièces de Ti-6Al-4V obtenues par SLM et EBM: effet de la
rugosité. Doctoral thesis, Ecole Nationale Supérieure d’Arts et Métiers, 2020. 15
[VET 14] Vetterli M., Schmid M., Wegener K.
Comprehensive investigation of surface characterization for laser sintered parts. DDMC 2014
: Fraunhofer Direct Digital Manufacturing Conference : proceedings, 2014. Artwork Size: 6
p. Medium: application/pdf Publisher: ETH Zurich. 46
[VIA 22] Viale V., Stavridis J., Salmi A., Bondioli F., Saboori A.
Optimisation of downskin parameters to produce metallic parts via laser powder bed fusion
process: an overview. The International Journal of Advanced Manufacturing Technology,
vol. 123, no 7-8, 2022, p. 2159–2182. 13
[VIL 18a] Villarraga-Gómez H.
Studies of Dimensional Metrology with X-ray CAT Scan. Doctoral thesis, University of North
Carolina, 2018. 3
[VIL 18b] Villarraga-Gómez H., Peitsch C. M., Ramsey A., Smith S. T.
THE ROLE OF COMPUTED TOMOGRAPHY IN ADDITIVE MANUFACTURING.
vol. 69, 2018. 44
[VIL 19] Villarraga-Gómez H., Herazo E. L., Smith S. T.
X-ray computed tomography: from medical imaging to dimensional metrology. Precision
Engineering, vol. 60, 2019, p. 544–569. 3, 190
[VI 23] Vié T., Deschanel S., Godin N., Normand B.
On the effect of coatings on the tensile and fatigue properties of 7075-T6 aluminum alloy
monitored with acoustic Emission (AE): Towards lifetime estimation. International Journal
of Fatigue, vol. 171, 2023, Page 107578. 126, 127, 130
[VUL 14] Vulliez M., Gleason M. A., Souto-Lebel A., Quinsat Y., Lartigue C., Ko-
rdell S. P., Lemoine A. C., Brown C. A.
Multi-scale Curvature Analysis and Correlations with the Fatigue Limit on Steel Surfaces
after Milling. Procedia CIRP, vol. 13, 2014, p. 308–313. 193
[WAN 21] Wang R., Law A. C., Garcia D., Yang S., Kong Z.
Development of structured light 3D-scanner with high spatial resolution and its applications
for additive manufacturing quality assurance. The International Journal of Advanced Manu-
facturing Technology, vol. 117, no 3-4, 2021, p. 845–862. 35
[WAN 22] Wang Z., Yang S., Peng Z., Gao Z.
Effect of defects in laser selective melting of Ti-6Al-4V alloy on microstructure and mechanical
properties after heat treatment. Optics and Laser Technology, vol. 156, 2022, Page 108522.
13, 15
[WAN 23] Wang F., Lei L., Fu X., Shi L., Luo X., Song Z., Zhang G.
Toward developing Ti alloys with high fatigue crack growth resistance by additive manufac-
turing. Journal of Materials Science and Technology, vol. 132, 2023, p. 166–178. 107

226 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[WEI 95] Weixing Y., Kaiquan X., Yi G.
On the fatigue notch factor, Kf. International Journal of Fatigue, vol. 17, no 4, 1995,
p. 245–251. 25
[WU 23] Wu M.-W., Ni K., Yen H.-W., Chen J.-K., Wang P., Tseng Y.-J., Tsai M.-K.,
Wang S.-H., Lai P.-H., Ku M.-H.
Revealing the intensified preferred orientation and factors dominating the anisotropic me-
chanical properties of laser powder bed fusion Ti6Al4V alloy after heat treatment. Journal of
Alloys and Compounds, vol. 949, 2023, Page 169494. 10
[WYC 13] Wycisk E., Emmelmann C., Siddique S., Walther F.
High Cycle Fatigue (HCF) Performance of Ti-6Al-4V Alloy Processed by Selective Laser
Melting. Advanced Materials Research, vol. 816-817, 2013, p. 134–139. 20, 21, 22
[XU 21] Xu Z., Liu A., Wang X., Liu B., Guo M.
Fatigue limit prediction model and fatigue crack growth mechanism for selective laser melting
Ti6Al4V samples with inherent defects. International Journal of Fatigue, vol. 143, 2021. 17,
23
[YAN 19] Yang L., Yan C., Cao W., Liu Z., Song B., Wen S., Zhang C., Shi Y., Yang
S.
Compressioncompression fatigue behaviour of gyroid-type triply periodic minimal surface
porous structures fabricated by selective laser melting. Acta Materialia, vol. 181, 2019,
p. 49–66. 40
[YIN 23] Yin H., Li P.
Micropore-propagation-based model of fatigue life analysis of SLM manufactured Ti-6Al-4V.
International Journal of Fatigue, vol. 167, 2023. 26
[YU 19] Yu H., Li F., Wang Z., Zeng X.
Fatigue performances of selective laser melted Ti-6Al-4V alloy: Influence of surface finishing,
hot isostatic pressing and heat treatments. International Journal of Fatigue, vol. 120, 2019,
p. 175–183. 16, 18
[YUA 19] Yuan L., Ding S., Wen C.
Additive manufacturing technology for porous metal implant applications and triple minimal
surface structures: A review. Bioactive Materials, vol. 4, 2019, p. 56–70. 52
[ZAB 18] Zabala A., Blunt L., Tato W., Aginagalde A., Gomez X., Llavori I.
The use of areal surface topography characterisation in relationto fatigue performance.
MATEC Web of Conferences, vol. 165, 2018. 32
[ZAC 77] Zack G. W., Rogers W. E., Latt S. A.
Automatic measurement of sister chromatid exchange frequency. Journal of Histochemistry
& Cytochemistry, vol. 25, no 7, 1977, p. 741–753. 43
[ZAN 19] Zanini F., Pagani L., Savio E., Carmignato S.
Characterisation of additively manufactured metal surfaces by means of X-ray computed
tomography and generalised surface texture parameters. CIRP Annals, vol. 68, no 1, 2019,
p. 515–518. 35
[ZEI 22] Zeidler H., Böttger-Hiller F.
Plasma-Electrolytic Polishing as a Post-Processing Technology for Additively Manufactured
Parts. Chemie Ingenieur Technik, vol. 94, no 7, 2022, p. 1024–1029. 131, 133

REFERENCES | 227

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
[ZER 19] Zerbst U., Madia M., Klinger C., Bettge D., Murakami Y.
Defects as a root cause of fatigue failure of metallic components. I: Basic aspects. Engineering
Failure Analysis, vol. 97, 2019, p. 777–792. 10
[ZHA 17] Zhao C., Fezzaa K., Cunningham R. W., Wen H., De Carlo F., Chen L.,
Rollett A. D., Sun T.
Real-time monitoring of laser powder bed fusion process using high-speed X-ray imaging and
diffraction. Scientific Reports, vol. 7, no 1, 2017, Page 3602. 13
[ZHA 18] Zhang L., Feih S., Daynes S., Chang S., Wang M. Y., Wei J., Lu W. F.
Energy absorption characteristics of metallic triply periodic minimal surface sheet structures
under compressive loading. Additive Manufacturing, vol. 23, 2018, p. 505–515. 52
[ZHA 19] Zhang J., Chaudhari A., Wang H.
Surface quality and material removal in magnetic abrasive finishing of selective laser melted
316L stainless steel. Journal of Manufacturing Processes, vol. 45, 2019, p. 710–719. 13
[ZHA 21] Zhao C., Qu N., Tang X.
Removal of adhesive powders from additive-manufactured internal surface via electrochemical
machining with flexible cathode. Precision Engineering, vol. 67, 2021, p. 438–452. 13

228 | REFERENCES

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf
© [F. Steinhilber], [2024], INSA Lyon, tous droits réservés
FOLIO ADMINISTRATIF
THÈSE DE L’INSA LYON, MEMBRE DE L’UNIVERSITÉ DE LYON

NOM : Steinhilber Date de soutenance : 01/02/2024

Prénom : Florian
TITRE : On the use of X-ray computed tomography to investigate the inﬂuence of surface integrity on the
fatigue properties of additively manufactured Ti64
NATURE : Doctorat Numéro d’ordre : 2024ISAL0014
École doctorale : ED 34 - Matériaux de Lyon
Spécialité : Matériaux

RÉSUMÉ :
Manufacturing defects are known to significantly impact the fatigue properties of additively manufactured (AM)
components. Notably, the large surface roughness, typical of AM processes, leads to increased stress concentrations
that promote crack initiation, thereby reducing fatigue resistance.
Traditionally, surface roughness and related defects are evaluated using tools like white light interferometers. How-
ever, these instruments offer a limited, single-perspective and only partially three-dimensional analysis. Those limita-
tions do not enable the thorough characterization of the complex surfaces and hidden defects typical in AM components.
This study describes a methodology for performing a 3D surface analysis using X-ray Computed Tomography
(XCT) data. The method is illustrated on various samples, ranging from simple cylinders to more intricate architected
structures. It turns out to be very efficient at detecting critical surface defects, such as notches hidden by partially
melted powder particles.
The methodology is then applied to examine the effect of surface defects on the fatigue properties of Ti64 pro-
duced by Laser Powder Bed Fusion (L-PBF). This analysis includes both as-built surfaces and those subjected to
post-treatments, specifically investigating the impact of Plasma electrolytic Polishing (PeP) and surface oxygen con-
tamination (presence of an α-case layer) resulting from high-temperature heat treatment (860 °C).
Using XCT for 3D characterization, defects responsible for fatigue failure are identified, the latter being predom-
inantly surface valleys. The method’s ability to predict crack initiation locations is also evaluated, as well as its
potential to estimate the fatigue resistance of a specimen before testing.

MOTS-CLÉS: 3D surface analysis, Fatigue, Additive Manufacturing, X-ray Computed Tomography, Ti64,
Plasma electolytic Polishing, Alpha-case.

Laboratoires de recherche : MatéIS − UMR CNRS 5510 | SIMaP − UMR CNRS 5266

Président du jury : Stefano Beretta (Professeur)

Composition du jury : Fabien Szmytka, ENSTA Paris (Professeur associé, Rapporteur)
Sabine Rolland du Roscoat 3SR (Maître de Conférences, Rapporteur)
Stefano Beretta, PoliMI (Professeur, Examinateur)
Jean-Yves Buﬃère, INSA LYON (Professeur, Directeur de thèse)
Rémy Dendievel, Grenoble INP (Professeur, Co-encadrant)
Guilhem Martin, Grenoble INP (Maître de Conférences, Co-encadrant)
Victor Chastand, Dassault Aviation (Ingénieur, Invité)

Cette thèse est accessible à l'adresse : https://theses.insa-lyon.fr/publication/2024ISAL0014/these.pdf

Vous aimerez peut-être aussi

These - Étude Tridimensionnelle de La Propagation en Fatigue de Fissures Internes Dans Les Matériaux Métalliques
Pas encore d'évaluation
These - Étude Tridimensionnelle de La Propagation en Fatigue de Fissures Internes Dans Les Matériaux Métalliques
193 pages
These
Pas encore d'évaluation
These
179 pages
These
Pas encore d'évaluation
These
219 pages
Caractérisation des matériaux ferroélectriques
Pas encore d'évaluation
Caractérisation des matériaux ferroélectriques
172 pages
Microstructure des Aciers Lean Duplex
Pas encore d'évaluation
Microstructure des Aciers Lean Duplex
223 pages
A Lire001
Pas encore d'évaluation
A Lire001
126 pages
Étude des Lois d'Endommagement en Fatigue
Pas encore d'évaluation
Étude des Lois d'Endommagement en Fatigue
169 pages
Etude Et Modélisation de La Fatigue de Contact en
Pas encore d'évaluation
Etude Et Modélisation de La Fatigue de Contact en
144 pages
These
Pas encore d'évaluation
These
180 pages
These
Pas encore d'évaluation
These
189 pages
These Laura
Pas encore d'évaluation
These Laura
251 pages
These
Pas encore d'évaluation
These
178 pages
These
Pas encore d'évaluation
These
298 pages
HUET Manuscrit These PDF
Pas encore d'évaluation
HUET Manuscrit These PDF
210 pages
Bardel PHD
Pas encore d'évaluation
Bardel PHD
277 pages
These - Etude Des Murs de Soutenement PDF
Pas encore d'évaluation
These - Etude Des Murs de Soutenement PDF
254 pages
These
Pas encore d'évaluation
These
173 pages
These
Pas encore d'évaluation
These
314 pages
Reconstruction 3D de Donnees Echographiques Du Cer
Pas encore d'évaluation
Reconstruction 3D de Donnees Echographiques Du Cer
232 pages
Corrosion des armatures en béton
Pas encore d'évaluation
Corrosion des armatures en béton
205 pages
Propagation Des Ultrasons en Milieu Heterogene Et Anisotrope - Application A Levaluation Des Proprietes D'elasticite Et D'attenuation D'aciers Moules Par Centrifugation Et de Soudures en Inconel PDF
Pas encore d'évaluation
Propagation Des Ultrasons en Milieu Heterogene Et Anisotrope - Application A Levaluation Des Proprietes D'elasticite Et D'attenuation D'aciers Moules Par Centrifugation Et de Soudures en Inconel PDF
169 pages
These
Pas encore d'évaluation
These
225 pages
These
Pas encore d'évaluation
These
308 pages
Comportement thermomécanique des enrobés
Pas encore d'évaluation
Comportement thermomécanique des enrobés
226 pages
These
Pas encore d'évaluation
These
297 pages
These
Pas encore d'évaluation
These
215 pages
These
Pas encore d'évaluation
These
203 pages
These
Pas encore d'évaluation
These
157 pages
These 2
Pas encore d'évaluation
These 2
127 pages
Etude Du Transfer de Chaleur Couple Conduction Rayonnement Dans Les Mousses de Polystyrene Expense Modelisation Et Caracterisation
Pas encore d'évaluation
Etude Du Transfer de Chaleur Couple Conduction Rayonnement Dans Les Mousses de Polystyrene Expense Modelisation Et Caracterisation
300 pages
These C Sebastian2013
Pas encore d'évaluation
These C Sebastian2013
354 pages
Outils de traçabilité agroalimentaire
Pas encore d'évaluation
Outils de traçabilité agroalimentaire
125 pages
Biocapteurs Implantables, These
Pas encore d'évaluation
Biocapteurs Implantables, These
198 pages
These
Pas encore d'évaluation
These
301 pages
Capteurs solaires PV/T hybrides intégrables
Pas encore d'évaluation
Capteurs solaires PV/T hybrides intégrables
200 pages
Propriétés Optiques des Nanoparticules d’Argent
Pas encore d'évaluation
Propriétés Optiques des Nanoparticules d’Argent
113 pages
Caractérisation du transistor JFET en SiC
Pas encore d'évaluation
Caractérisation du transistor JFET en SiC
168 pages
These PDF
50% (2)
These PDF
250 pages
Angular Contact Ball Bearing Modelling With Flexible Cage
Pas encore d'évaluation
Angular Contact Ball Bearing Modelling With Flexible Cage
224 pages
These
Pas encore d'évaluation
These
146 pages
Valorisation Déchets Agricoles Haïti
Pas encore d'évaluation
Valorisation Déchets Agricoles Haïti
217 pages
These - Memoire - Definitif
Pas encore d'évaluation
These - Memoire - Definitif
251 pages
These
Pas encore d'évaluation
These
218 pages
These
Pas encore d'évaluation
These
256 pages
Comportement des sables et argiles
Pas encore d'évaluation
Comportement des sables et argiles
440 pages
Pagegarde 1
Pas encore d'évaluation
Pagegarde 1
5 pages
These T.Ibrahim
Pas encore d'évaluation
These T.Ibrahim
260 pages
These1 Usure Denture Engrenage
Pas encore d'évaluation
These1 Usure Denture Engrenage
221 pages
TH2019TRANVUTUNGLAM
Pas encore d'évaluation
TH2019TRANVUTUNGLAM
139 pages
Tunnels Urbains: Prévision des Tassements
Pas encore d'évaluation
Tunnels Urbains: Prévision des Tassements
267 pages
These
Pas encore d'évaluation
These
149 pages
Recyclage Mélanges PE/PS : Étude 2003
Pas encore d'évaluation
Recyclage Mélanges PE/PS : Étude 2003
208 pages
These Choulak
Pas encore d'évaluation
These Choulak
346 pages
Analyse multi-omique en spectrométrie de masse
Pas encore d'évaluation
Analyse multi-omique en spectrométrie de masse
319 pages
Filtres microondes planaires accordables
Pas encore d'évaluation
Filtres microondes planaires accordables
162 pages
Calcul Haute-Performance Et Mécanique Quantique
Pas encore d'évaluation
Calcul Haute-Performance Et Mécanique Quantique
154 pages
RDM Solution Exercices Cisaillement RDM 230215 083640
100% (1)
RDM Solution Exercices Cisaillement RDM 230215 083640
8 pages
Effet de la Température sur le Béton
Pas encore d'évaluation
Effet de la Température sur le Béton
24 pages
Menu Soir: Spécialités Rôties et Entrées
Pas encore d'évaluation
Menu Soir: Spécialités Rôties et Entrées
2 pages
A2
Pas encore d'évaluation
A2
2 pages
Calcul du pH et solutions tampons
Pas encore d'évaluation
Calcul du pH et solutions tampons
47 pages
Satire de la société du XVIIe siècle
Pas encore d'évaluation
Satire de la société du XVIIe siècle
5 pages
French HW
Pas encore d'évaluation
French HW
2 pages
Ondes Lumineuses Cours PDF
Pas encore d'évaluation
Ondes Lumineuses Cours PDF
7 pages
CpA 44
Pas encore d'évaluation
CpA 44
11 pages
Tableau Comparatif 1929 2025: Effondrement Soudain Et Massif
Pas encore d'évaluation
Tableau Comparatif 1929 2025: Effondrement Soudain Et Massif
3 pages
Questionnaire pré-anesthésique pédiatrique
Pas encore d'évaluation
Questionnaire pré-anesthésique pédiatrique
1 page
Poly Complet
Pas encore d'évaluation
Poly Complet
43 pages
Wembadjonga Wemba Ernest
Pas encore d'évaluation
Wembadjonga Wemba Ernest
12 pages
AGC-4 Data Sheet 4921240402 FR-20180427-J
Pas encore d'évaluation
AGC-4 Data Sheet 4921240402 FR-20180427-J
31 pages
Épreuves CAP Cuisine 2018 : Guide complet
Pas encore d'évaluation
Épreuves CAP Cuisine 2018 : Guide complet
16 pages
Traitement Du Signal - TP2
Pas encore d'évaluation
Traitement Du Signal - TP2
22 pages
French 2am22 2trim d3
Pas encore d'évaluation
French 2am22 2trim d3
2 pages
Acfci Guige PME SST
100% (1)
Acfci Guige PME SST
100 pages
FONDATIONS. Semelles Superficielles Et Semelles Sur Pieux. Juin 2010 Henry THONIER T
100% (2)
FONDATIONS. Semelles Superficielles Et Semelles Sur Pieux. Juin 2010 Henry THONIER T
79 pages
Devis Estimatif Et Quantitatif Prison Principale de Monatele - Copie
Pas encore d'évaluation
Devis Estimatif Et Quantitatif Prison Principale de Monatele - Copie
30 pages
SNC - Protec DEBUT MERCREDI
Pas encore d'évaluation
SNC - Protec DEBUT MERCREDI
144 pages
CHAPII-REDRESSEUR PT
100% (1)
CHAPII-REDRESSEUR PT
83 pages
PRÉDICATION Ce Qui Est Délié Sur Terre Est Délié o Ciel 21 JUIN 2024 - FRJUSTAVIE
Pas encore d'évaluation
PRÉDICATION Ce Qui Est Délié Sur Terre Est Délié o Ciel 21 JUIN 2024 - FRJUSTAVIE
7 pages
3-Cours Spongiaires
Pas encore d'évaluation
3-Cours Spongiaires
39 pages
2023 Normes-Desinfectants CPiasARA
Pas encore d'évaluation
2023 Normes-Desinfectants CPiasARA
6 pages
Plan Stratégique 2019 - 2023: Promouvoir L'accès Des Filles Et Des Femmes À L'éducation Et À La Formation en Afrique
Pas encore d'évaluation
Plan Stratégique 2019 - 2023: Promouvoir L'accès Des Filles Et Des Femmes À L'éducation Et À La Formation en Afrique
64 pages
Série 4 Réactions Spontanées Dans Les Piles
Pas encore d'évaluation
Série 4 Réactions Spontanées Dans Les Piles
5 pages
Corrigé BFEM Mathématiques 2015
Pas encore d'évaluation
Corrigé BFEM Mathématiques 2015
3 pages
Projet Écorce D'orange
Pas encore d'évaluation
Projet Écorce D'orange
15 pages
Exercice Réseaux Électriques HTA-BTA
Pas encore d'évaluation
Exercice Réseaux Électriques HTA-BTA
8 pages