Master Recherche «Sciences, Technologies, Santé»

Mention «Ingénierie de Conception»

Spécialité «Mécanique, Matériaux et Génie Civil»
M2R MMGC

COURS

METHODES NUMERIQUES
Problèmes d’Optimisation

Adinel GAVRUS

Maître de Conférences des Universités HC HDR

 Master 2R MMGC – UE3

METHODES NUMERIQUES
PARTIE A. GAVRUS
PROBLEMES D’OPTIMISATION

1 PRINCIPES ET METHODES NUMERIQUES DE RESOLUTION

1.1 Introduction

1.1.1 Utilisation des techniques d’optimisation dans le domaine de Génie

Mécanique et Génie Civil
1.1.2 Illustrations avec des exemples qui utilisent la simulation numérique
1.1.3 Principes de base pour la formulation d’un problème d’optimisation

1.2 Formulation mathématique d’un problème d’optimisation

1.2.1 Critères et conditions du point optimal

1.2.2 Analyse fonctionnelle et Control Optimal
1.2.3 Traitement des problèmes d’optimisation sans et avec contraintes
1.2.4 Exemples, Extensions …

1.3 Méthodes numériques de résolution

1.3.1 Méthodes d’ordre zéro : simplex, génétiques, réseaux neuronales, …

1.3.2 Méthodes d’ordre un : gradient, gradient conjugué, …
1.3.3 Méthodes d’ordre deux : Newton-Raphson, BFGS,…
1.3.4 Méthodes mixtes et couplages

1.4 Algorithmes d’optimisation : conditionnement, stabilité et précision

1.5 Conclusions

2 PROBLEMES INVERSE : PRINCIPES ET APPLICATIONS DANS LE DOMAINE

DE L’INGENIERIE ET D’IDENTIFICATION PARAMETRIQUE

2.1 Introduction

2.1.1 Principes
2.1.2 Généralités et exemples

2.2 Problèmes directes et problèmes inverses

2.2.1 Formulation générale

2.2.2 Problèmes d’ingénierie : optimisation de structures, recherche de
conditions limites, calibration des modèles, identification des paramètres par
analyse inverse
2.3 Résolution des problèmes inverses

2.3.1 Formulation mathématique

2.3.2 Résolution d’un système linéaire par inversion analytique : régression
linéaire
2.3.3 Résolution d’un système non-linéaire. Technique de régression non-
linéaire : méthode des moindres carrées
2.3.4 Résolution des systèmes complexes par analyse inverse

2.4 Identification de paramètres par analyse inverse

2.4.1 Formulation dans un espace euclidien : réduction à un problème de

minimisation
2.4.2 Diagramme général de résolution
2.4.3 Technique de résolution par analyse inverse : Définition de l’espace
expérimental, du modèle de calcul et du couplage d’optimisation

2.5 Algorithmes numériques

2.5.1 Algorithmes d’ordre zéro : simplex, Monte-Carlo, …

2.5.2 Algorithmes basés sur la méthode de descente : gradient simple,
gradient conjugué, Newton-Raphson, quasi-Newton (BFGS, Gauss-Newton),
…
2.5.3 Techniques de régularisation : Tikhonov, Levenberg, Levenberg-
Marquard, …
2.5.4 Traitement numériques des conditions paramétriques et fonction
contraintes

2.6 Analyse de sensibilité paramétrique

2.6.1 Définition et principes

2.6.2 Méthode de différentiation par différences finis
2.6.3 Méthode de différentiation analytique
2.6.4 Méthode adjointe
2.6.5 Méthode mixte ou semi-analytique

2.7 Estimation statistique des paramètres

2.8 Applications dans le domaine de la recherche et d’ingénierie des matériaux et

structures

2.8.1 Techniques d’optimisation par analyse inverse. Application de

l’analyse inverse aux tests mécaniques.
2.8.2 Quelques exemples sur l’optimisation de la masse dans les opérations
de mise en forme, sur l’optimisation de la forme, sur l’identification des
variables thermiques dans le formage à chaud, sur la distribution optimale
des particules chargés, …
2.8.3 Identification des paramètres rhéologiques des matériaux par l’analyse
inverse des tests de traction, compression, torsion et des essais en
dynamique rapide et impact
2.8.4 Logiciels industriels et commerciaux
2.8.5 Perspectives

2.9 Conclusions, extensions et couplages

 Adinel GAVRUS

Scientific Course
Laboratory of GCGM (EA3913)

OPTIMIZATION PROBLEMS
PART I
PRINCIPLES AND NUMERICAL METHODS

Assoc. Prof. HDR Dr. Adinel GAVRUS

GMA/PSM – LGCGM (EA3913)
National Institut of Applied Sciences INSA Rennes,
University of Loire Brittany (ULB), France

Adinel GAVRUS
 Adinel GAVRUS

« Les outils numériques disponibles dans les logiciels de

calcul sont nombreux et sophistiqués. Développés par des
numériciens de haut vol – qui ont parfois du mal à faire
passer le message, même via les étapes commerciale, ils sont
de ce fait, au mieux sous-employés, au pire mal utilisés ».

« Fort heureusement, les choix sont déjà en partie effectués

par le concepteur ; mais, quand tant de possibilités sont
offertes, il reste malgré tout du ressort de l'utilisateur d'opter
pour telle méthode ou tel schéma ».

Reading Notes of Dr. Philippe PASQUET (CISI - Bulletin XXI-2 de Φ2AS)

concerning the book « Analyse Numérique pour Ingénieurs » of André FORTIN
(Ecole Polytechnique de Montreal, 1st Edition 1995)

Adinel GAVRUS

CONTENTS

INTRODUCTION

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
NUMERICAL METHODS

OPTMIZATION ALGORITHMS

CONCLUSIONS
 Adinel GAVRUS

BIBLIOGRAPHIE
[1] J. BARANGER– « Analyse Numérique », Hermann Edition, 1991.
[2] A. FORTIN – « Analyse Numérique pour Ingénieurs », Editions de l’Ecole
Polytechnique de Montréal, 1995.
[2] S. CHARNOZ and A. DAERR – « Algorithmes de Minimisation»,
Université Paris 6 Denis Diderot – CEA Saclay.
[3] J.-M. HURE and D. PELAT – « Méthodes Numériques. Eléments d’un
première parcours », Ecole Doctorale d’Ille de France, Université Paris 7 et 11 –
Observatoire de Paris, 2002/2003.
[4] V. BARRA, J. KOKO and P. MAHEY – « Cours d’Analyse Numérique.
Analyse Numérique Matricielle et Optimisation », Institut Supérieur
d’Informatique, de Modélisation et leurs Applications (ISIMA).
[5] C.R. BECTOR, S. CHANDRA and J. DUTTA – « Principles of
optimization theory», Alpha Science International Ltd., 2005.
[6] J. F. BONNANS, J. C. GILBERT, C. LEMARECHAL... et al. – « Numerical
optimization : theoretical et practical aspects », Springer Ed., 2006.

Adinel GAVRUS

CONTENTS

INTRODUCTION

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
NUMERICAL METHODS

OPTMIZATION ALGORITHMS

CONCLUSIONS
 Adinel GAVRUS

INTRODUCTION
Many problems requires to minimize or maximize
functions or functional:
- Maximization of performance, minimization of mass, price,
energy consuming, optimal distribution of internal state variables
- Find the equilibrium state of a mechanical system
(minimization of the potential energy)
- Find the equilibrium state of a fluid or a mixture
(maximization of the entropy)
- Minimization of the distance between geometrical points or
between the points of an experimental curve and a computed one
...
All these previous problems fall into a broad category:
OPTIMIZATION
Remark : minimization of f(x) = maximization of –f(x)

Adinel GAVRUS

OPTIMIZATION TECHNICS APPLIED IN CIVIL

AND MECHANICAL ENGINEERING

FORMULATION OF THE PHYSICAL

PROBLEM
• Preliminary definition, design and modeling of a structure, piece
or material objet knowing the extern loadings, the boundary
conditions, the intern and the other environment interactions

FORMULATION OF THE OPTIMIZATION

PROBLEM
• Search of conditions and of some variables values which take
correct predictions, maximize the performances, minimize the
mass, the energy consuming or the total manufacturing price
 Adinel GAVRUS

EXAMPLE OF A SIMPLE PROBLEM

CIRCULAR SECTION BEAM IN CONSOLE

Optimization Problem: Mass minimization

under structural constraints
Data: L = 1m, D = 50 mm, F = 1000 N, AU4G (E = 74000 MPa)
Optimization Constraints: wmax ≤ 10 mm, σmax ≤ 150 MPa
Optimization Results from RDM5 software: Dsol = Dmin = 55.04 mm

Adinel GAVRUS

EXAMPLES OF OPTIMIZATION USING

NUMERICAL MODELS (FINITE ELEMENTS, …)
TOPOLOGIC OPTIMIZATION ( http://www.lanxis.fr )
Positioning a ribs array on a shell structure
 Adinel GAVRUS

TOPOLOGIC OPTIMIZATION
( http://www.lanxis.fr )

Design of a lattice structure

Adinel GAVRUS

TOPOLOGIC OPTIMIZATION
( http://www.lanxis.fr )

Design of a casting process

 Adinel GAVRUS

TOPOGRAPHIC OPTIMIZATION
( http://www.lanxis.fr )

Design of a car door

Topography
Current Design Optimization Results Proposed Optimized
(19/% Displacements Reduction) Design

Adinel GAVRUS

MATERIAL CONSUMING OPTIMIZATION

(http://www.transvalor.com)
 Adinel GAVRUS

3D SHAPE OPTIMIZATION
([Tien Tho DO, 2006], http://www.transvalor.com)
Case of a triaxial wheel forging (modeling of twelfth part)
Initial material specimen Preform material specimen Final piece
(First forging operation) (Second forging operation) (Final forging operation)

Preform?

Initial Preform Optimal Preform

Adinel GAVRUS

BASIC PRINCIPLES FOR AN OPTIMIZATION

PROBLEM FORMULATION
• Choice of the design / structural variables and/or of
the parameter vector
• Definition of constraints and of the domain variation
• Built or choice of the computation model
• Analysis of the output data and definition of the
optimization criteria
• Mathematical formulation of the objective functions
and of the optimization problem
•Verified if it is a well-posed problem (existing solution)
• Choice of the optimal control method and of the
optimization method (or numerical code)
 Adinel GAVRUS

CONTENTS

INTRODUCTION

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
NUMERICAL METHODS

OPTMISATION ALGORITHMS

CONCLUSIONS

Adinel GAVRUS

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
Modeling
Simplification of the reality to define and understand the studied
system: experimental analysis and mathematical modeling
Simulation
“Artificial” representation of the real system working: set of the
specific equations and of the boundary conditions
Optimization
Find of an optimal configuration that is better than any other for
specific defined mathematical criterions:
Objective or Cost Function: min f(x) or max f(x)
Constraint’s Functions: g(x) ≤ a (or g(x) - a ≤ 0 )
g(x) ≥ b (or g(x) - b ≥ 0 )
h(x) = c (or g(x) = h(x) – c = 0 )
Bound Constraints: xmin ≤ x ≤ xmax or Sign Constraints: x ≥ 0 or x ≤ 0
 Adinel GAVRUS

HISTORY

Decision’s Mathematical Combinatory,

Theory Programming Graphs Theory

Pascal (1654) Fourier (1824)

Fermat (1654) Sainte-Laguë (1926)
Bernoulli (1713) D. König (1936)
Op or Opt 2nd world war: First application

Simplex First “big” algorithm (1947)

Mathematical
Improvements
First Computer
First computer First commercial application (1956)

Adinel GAVRUS

GENERAL SCHEME OF AN OPTIMAL

CONTROL PROBLEM
Computation
Reality
Simulation

Observation Prediction Conditions

Optimization

Data Analysis
Decision
Input-Output
Estimation

Modeling
 Adinel GAVRUS

CRITERIA AN CONDITIONS OF OPTIMAL POINTS

Mathematical Forms and Equivalences Rules

( y = α2 )

Convexity
) ≤

Adinel GAVRUS

Conditions of an Optimal Solution i.e. of the

equivalent minimum point
• When the objective function is linear and that there is
no constraints the solution is always infinite.
• If the objective function and the constraints are linear,
there is at least one active constraint at the solution.
(the constraint g(x) ≤ 0 is called active in x* if g(x*)=0)
• If the solution does not exist we say that the problem
has no feasible solution (ill-posed problem).
• If f(x*) = min[f(x)] then f’(x*) = 0 (necessary condition).
• If f’’(x*) > 0 (i.e. locally a convex function) then x*
corresponds to a local minimum.
 Adinel GAVRUS

Resuming of principal properties of the functions

used in an optimization problem
• linear / non-linear

• continuum real variables values / entire ones

• constraints / no constraints

• convex (concave) / non convex (non concave)

• derivative / non-derivative functions

• various local solutions or single global solution

Adinel GAVRUS

Intuitive Illustrations of Non-Convex Functions

Horizontal tangent but Non continuum
not convex function derivative in 0

Concave
Function

)
 Adinel GAVRUS

Application to a general scalar objective function

depending of a parameter’s vector
• State equations of the system Modeling :R[x, v(x)] = 0
where x = {x1, x2, …, xnp} – the design parameters
and v = {v1, v2, …, vnv} – the state variables

• Optimization Problem:
min f[x, v(x)] with g[x, v(x)] = 0 and/or xmin ≤ x ≤ xmax

• For local or global minimum of the cost function f:

Necessary condition: grad(f) = 0 i.e [df/dxi]i=1,np = 0 (Euler Equation)

Local convexity: the hessian matrix H = [d2f/dxidxj]i,j=1,np

must be positive defined around the optimal solution i.e. aTH(x*)a > 0

Adinel GAVRUS

A Simple Analytical Example

• State equations of the system Modeling :R[v] = 0
non depending of design parameters x = {x1, x2}

• Optimization Problem:
min f(x) with f(x1,x2)=x12+x22 (paraboloid)
• For local or global minimum of the cost function f:
Necessary condition: grad(f) = 0

Locally convexity: the hessian matrix H = [d2f/dxidxj]i,j=1,2

Global minimum solution: x*={0, 0}

 Adinel GAVRUS

EXAMPLE OF A SIMPLE PROBLEM

CIRCULAR SECTION BEAM IN CONSOLE

Optimization Problem: Mass minimization

under structural constraints

• Data: L = 1m, D = 50 mm, F = 1000 N, AU4G (E = 74000 MPa)

Opt. Function: f = ρLπD2/4 , Constraints: wmax ≤ 10 mm, σmax ≤ 150 MPa
• Analytical Model: I = πD4/64, wmax = FL3/3EI, σmax = 0.5*FLD/I

• Constraints Functions :FL3/3EI ≤ 10 mm,0.5*FLD/I ≤ 150 MPa

• Analytical Solution: D ≥ 55.04 mm i.e. Dsol = Dmin = 55.04 mm

Adinel GAVRUS

TRAITMENT OF OPTIMIZATION
PROBLEMS WITHOUT AND UNDER
CONSTRAINTS
OPTIMIZATION PROBLEMS
WITHOUT CONSTRAINTS FUNCTIONS
• minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximax

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS
• minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximaxand g(x)=0
or
• minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximax and h(x)≤0
 Adinel GAVRUS

OPTIMIZATION PROBLEMS
WITHOUT CONSTRAINTS FUNCTIONS
• minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximax

• 1° use a transformation of parameters domain:

-definition of x in function of x’ varying between –∝ and + ∝
(transformation of [xmin, xmax] domain in [–∝ , +∝ ] domain)
• solve or verify grad[f(x)] = 0
• verify if aTH(x*)a > 0

• 2°possible to add the constraints functions defined by:

h1 = xi -ximax ≤ 0 and h2 = ximin-xi ≤ 0 and to use methods of
optimization problems under constraints

Adinel GAVRUS

OPTIMIZATION PROBLEMS
WITHOUT CONSTRAINTS FUNCTIONS
EXERCICE

1° Résoudre le problème d’optimisation en utilisant grad[f(x,y)] = 0

2° Partant de la matrice Hessienne H(x,y) calculée pour le point de gradient nul
vérifier bien la convexité et donc que la solution obtenue représente bien un point de
minimum.
3° Discussions sur la solution trouvée par rapport au minimum global de la fonction.
 Adinel GAVRUS

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS (g(x)= 0)
minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximaxand g(x)=0
Classical Method
• use transformation of parameters
• solve or verify: grad[f(x)] = 0 and g(x) = 0 and verify aTH(x*)a > 0
Method of Lagrange multipliers
• use transformation of parameters
Min L(x,λ) where L(x,λ) = f(x) + λg(x)
• solve or verify: gradL(x,λ) = 0
grad[f(x)] + λgrad[g(x)] = 0 and g(x) = 0 and verify if aTH(x*)a > 0
Penalty Method
• use transformation of parameters
Min F(x) where F(x) = f(x) + pg2(x) with p>> (p = 105 or 108)
or Min F(x) where F(x) = f(x) + 0.5pg2(x)-rg(x) with p>>, r r-pg(x)

Adinel GAVRUS

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS (g(x)= 0)
EXERCICE

Method of Lagrange
 Adinel GAVRUS

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS (g(x)= 0)
EXERCICE

1° A l’aide d’un graphique indiquer dans le plan xy le domaine admissible de la

solution.

2° Résoudre le problème d’optimisation en utilisant l’écriture du Lagrangien L(x,y,λ)

3° Partant de la matrice Hessienne H(x,y) de L(x,y, λ), pour des valeurs bien précises
du multiplicateur de Lagrange λ, donner les coordonnées des points optimales et
indiquer la solution finale du problème. Discussions sur les solutions trouvées.

Adinel GAVRUS

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS (h(x) ≤ 0)
minf(x) with x = {x1, x2, …, xnp} where ximin ≤ xi ≤ ximaxand h(x)≤0

Khun and Tucker Method (similar to Lagrange multipliers)

• use transformation of parameters
• solve or verify: grad[f(x)] + µ.grad[h(x)] = 0, µh(x) = 0 with µ ≥ 0
and verify if aTH(x*)a > 0 and the minimum of the function
Barrier Method
• use transformation of parameters
Min F(x) where F(x) = f(x) + ω/h(x) with ω > 0
• necessary to updated ω to obtain h(x*)<0
Lagrange multipliers
• use transformation of parameters
Min L(x,λ) where L(x,λ,µ’) = f(x) + λ[h(x)+µ’] with µ’ ≥ 0
• see precedent slides
 Adinel GAVRUS

OPTIMIZATION PROBLEMS
UNDER CONSTRAINTS FUNCTIONS (h(x) ≤ 0)
EXERCICE

Khun and Tucker Method

Adinel GAVRUS

EXAMPLE OF A SIMPLE PROBLEM

CIRCULAR SECTION BEAM IN CONSOLE

Optimization Problem: Mass minimization

under structural constraints
• Data: L = 1m, D = 50 mm, F = 1000 N, AU4G (E = 74000 MPa)
Opt. Function: f = ρLπD2/4 Constraints: wmax ≤ 10 mm, σmax ≤ 150 MPa
• Analytical Model:
I = πD4/64, wmax = FL3/3EI, σmax = 0.5*FLD/I

• Optimization Constraints:
h1(D) = FL3/3EI − 10 mm ≤ 0 ; h2(D) = 0.5*FLD/I − 150 MPa ≤ 0

• Solution: Dsol = 55.04 mm

 Adinel GAVRUS

FUNCTIONAL ANALYSIS AND OPTIMAL

CONTROL
Definition of an integral functional

Definition of the functional derivative following the function v*(x)

Minimization of Φ requires the following necessary conditions

(Euler-Poisson)

Adinel GAVRUS

FUNCTIONAL ANALYSIS AND OPTIMAL

CONTROL
Functional optimization under functional constraints

Functional depending on first order partial derivatives

Minimization of Φ requires the following necessary condition

(Euler-Ostrogradski)
 Adinel GAVRUS

Example for the equilibrium position of a flexible and heavy inextensible

string with a constant length “L” between two fixed points
Equilibrium is reached when the gravity center of the string takes the
lowest position i.e. Min YG

According to the optimal functional theory it is necessary to minimize

the Lagrangian Functional Ψ(y) defined by

The problem solution can be obtained from the Euler-Poisson equation using a
variable change defined by

Adinel GAVRUS

REMARKS, EXTENSIONS, …
• An optimization method can be used to solve scalar non-
linear equations:
- it is possible to verify that to solve f(x) = 0 is equivalent to
search the global minima of f2(x)
- consequently some numerical methods used to solve non-
linear equations can be employed to solve optimization
problems
• It is possible to use multi-objective functions:
- the first strategy suggest to optimize the sum of all the
objective functions
• Majority of physical systems have the equilibrium
defined by the optimal (minimal) value of a potential
Example: The Finite Elements modeling of the viscoplastic
material’s behavior (Annexe 1)
 Adinel GAVRUS

APPLICATION
LINEAR REGRESSION METHOD

Adinel GAVRUS

APPLICATION
FINITE ELEMENT MODELING OF A VISCOPLASTIC
MATERIAL BEHAVIOR
(ANNEXE I)
• The mechanical equilibrium can be written in a weak form in
terms of the velocities starting from the local equilibrium
equations
• It is the obtained the virtual power principle
• Using a Norton viscoplastic behavior, the previous equations are
equivalent to the minimum of a potential written as a functional of
the unknown velocities functions
• existence and uniqueness of the mechanical solution can be
proved from the convexity of the potential using theory of
functional optimization
 Adinel GAVRUS

CONTENTS

INTRODUCTION

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
NUMERICAL METHODS

OPTMISATION ALGORITHMS

CONCLUSIONS

Adinel GAVRUS

NUMERICAL METHODS
(Minimization Problems)
ZERO ORDER METHODS
• Only evaluation of the objective function f for an “intelligent” choice of a set of
the design values x
• f(xnext) ≤ f(xprevious)
• Possible to find global minimum solution but high times consuming methods
ONE ORDER METHODS
• Evaluation of the objective function f and of its gradient f’ = grad(f)
• initial a priori information for x i.e. set of x0
• Computation of the parameter variation ∆x using the grad(f)
• verify if f(x+∆x) ≤ f(x)
• If ∆x ≤ ε possibility to have a local or global minimum solution

SECOND ORDER METHODS

• Similar with one order method but ∆x is computed using Euler Equation
f’ = grad(f) = 0 with a some form of the hessian matrix f’’ = H of f
COUPLED METHODS
 Adinel GAVRUS

ZERO ORDER METHODS

Golden Rule Method : only for one dimension (1D) minimization problem
• use the bisection method to define the next parameters values using the
golden number defined by φ = 1.618034 (φ2-φ-1=0)
• some improvement if use of a local approximation of the cost function with
a parabolic form
c Frame the minimum between three points :
a<b<c such as f(b)< f(a), f(b) < f(c)
a 1. Start with a, b and c parameters values
2. Compute f(a), f(b), f(c), the segments AB and BC
3. Choice x1 between a and b (using golden section)
4. Si f(x1)< f(a) then a=x1; Si f(x1)< f(b) then b=x1
5. Choice x2 between b and c (using golden section)
b
Si f(x2)< f(c) then c=x2
Si f(x2)< f(b) then b=x2
6. Return in 2
Remark: Possible also to use the parabolic formula

X B
0.381966 A

Adinel GAVRUS

ZERO ORDER METHODS

Example of the use of “Golden Rule” for the minimization of Sin2(x)

Golden Rule
Error

Parabolic

Iterations Number
In this case the use of golden section has a better converge that the
classical “parabolic” one
 Adinel GAVRUS

ZERO ORDER METHODS

Simplex Method : multi-dimensional linear or non-linear programming

• taking into account the constraints functions: parameters area to avoid or
prohibitive parameters area
• quite complex for non-linear problems and times consuming methods

A Simplex is a set of Np + 1 points which surrounds the minimum

( Np = number of parameter dimension)

A 2D - the simplex is a triangle

A 3D - the simplex is a pyramid
Etc ….
The minimum is in the
pyramid volume

Adinel GAVRUS

ZERO ORDER METHODS

Non-Linear Simplex (Amoeba): Strategy of parameters evolution

Reflection: choice of the point in the inverse

direction

Reflection + Expansion

Contraction

Multiple contraction
 Adinel GAVRUS

ZERO ORDER METHODS

Monte Carlo : probabilistic approach (generally used for gas or crystals theory)

• stochastic strategy to test cost function for a random parameter’s values set
• generally find a global minimum
• requires an important number of cost function evaluation
• times increased exponentially with the parameters vector dimensions

Fuzzy Logic: essentially based on logical functions or fuzzy operators and on a

measure of qualitative or fuzzy variables

• can use non-numerical and discrete variables and parameters

• can take into account a high number of quantitative and qualitative constraints
functions: prohibitive parameters area, sensorial properties, …
• generally use approximate analytical relationships of analyzed physical model

Adinel GAVRUS

ZERO ORDER METHODS

Neuronal Network : inspired by biological neural systems process data.
• artificial intelligence method based on statistical estimation and
multilayer neural processing (nodes network model)
• strongly used for an optimal system control or operational decision
• improvement by trainings and automatic learning
An elementary neural network yi = x1w1i + x2w2i + … + xpwpi
g = h(y), f = g(y’)
 Adinel GAVRUS

ZERO ORDER METHODS

Genetic Method : using principles of the theory of evolution and genetics
• stochastic or combinatory algorithms: generate a sequence sets of parameter
(populations) whose the elements (chromosomes) tend to the optimum
• decimal or binary encoding of the parameters
• require many estimates of the objective function to reach the global minimum
because the rate of convergence is usually relatively small
For each “k+1” increment
1° Generation of an initial parameter set
2° Selection of two “parents” with a
probability proportional to a “fitness” function
3° Crossover to generate a “son”
4° Mutation to modify the specimen with an
aleatory noise
5° Verified the stop criteria and return to 1°

Adinel GAVRUS

ZERO ORDER METHODS

Simulated Annealing (Slow Relaxation) : probabilistic approach using “system
energy” theory (probability P(E) of an energetic state)
• the results depend strongly on the parameters generation strategy
• effective if search of a global minimum
• very times consuming computations
Repeat N times at a constant T

P 1°- Starting from a state number K, energy Ek,

and temperature T
• E will be the cost function f to minimize 2°- It randomly generates a new state, K +1, close
• T will be a fictive control parameter to the state K with an energy Ek+1 (∆E~KT)
which will be decreased slowly 3°- It is now look if accept the state K+1. For this
• The Boltzmann constant K will be state is computed the probability P of jump from
replaced by an arbitrary constant so that K to K+1. It takes a number X between 0 and 1
P(E) keep accessible numerical bounds of
(uniform distribution). If X > P then it accepts it,
the computer precision.
otherwise it rejects
4°- When E does not decrease more, T is reduced
and again in 1°
 Adinel GAVRUS

ZERO ORDER METHODS

Example of use of Slow Relaxation Method for an objective function f(x1,x2)

Start Point : [0,0]

Solution Point : [0.68, -1.42]
Number of iterations: 105
Accepted state change: 310

Adinel GAVRUS

ONE ORDER METHODS

• All these methods can be applied to one or multi-dimensional variables space
• Based on evaluation of objective function f and of its gradient f’ = grad(f)
• Principle: the minimum can be reached following a descent direction d
computed from the opposed directions of function gradient until this one is 0
(find x* such as f’= grad f = 0 and which reaching a minimum of f )
• generally find local minima and can posed problems for narrow “valleys”

k = 0: start from an initial point x0 and of α = α0

For each “k+1” iteration:
1° computation of xk+1 = xk + αkdk where dk =d(xk)
2° Estimation of αk such as f(xk+1) <f(xk) (use of a line
search method or of an zero order ones)
3° Verified the stop criteria:IIdkII≤ε or IIxk+1-xkII/IIxkII≤δ
4° If criteria are not satisfied, return to 1°
 Adinel GAVRUS

ONE ORDER METHODS

Examples of shape variation of two-dimensional functions to be minimized

Adinel GAVRUS

ONE ORDER METHODS

Gradient Method: minimization using the descent gradient
• evaluation of objective function f and of its gradient f’ = grad(f)
• minimum search following the higher descent direction d = h = -grad(f)
k = 0: start from an initial point x0 and of α = α0
For each “k+1” iteration:
1° xk+1 = xk + αkdk where dk = hk = - grad[f(xk)]
2° Estimation of αk such as f(xk+1) <f(xk) (use of a zero
order method or αkopt=(hk.hk)/(H hk).hk where H –hessian
matrix of f)
3° Verified the stop criteria:IIdkII≤ε or IIxk+1-xkII/IIxkII≤δ
4° If criteria are not satisfied, return to 1°
• the convergence depends strongly on the choice of the descent step α
• generally α < characteristic size of the search area
• numerically is better to choice α = α’/IIgrad[f(x)]II where α’ is a scalar
 Adinel GAVRUS

ONE ORDER METHODS

Gradient Method: limitations and difficulties
• problems around narrow “valley

• The arrows are in the direction of the function gradient.

• The points for the yellow and orange gradients are not the best directions!
• Improvement if use of a Conjugate Gradient Method

Adinel GAVRUS

ONE ORDER METHODS

Conjugate Gradient Method: minimization using combinations of
descent gradients at previous steps
• evaluation of objective function f and of its gradient grad(f) for each
search points
• minimum search following an optimal descent direction d computed from
actual step gradient and the previous one

Gradient Method Conjugate Gradient Method

• for each “k+1” iteration, search of dk conjugate to the dk-1 i.e. (Hdk-1).dk=0
• this method browses effectively a narrow "valley“, as compared to a simple
gradient descent
• converges quickly to a minimum and requires preconditioning
• maximum Np iterations if the function is quadratic
• difficult evaluation of the descent step α if not use of the Hessian matrix of f
 Adinel GAVRUS

ONE ORDER METHODS

Strategy of the Conjugate Gradient Method
• for each “k+1” iteration, dk-1 is forced to be orthogonal with gradf(xk) i.e.
gradf(xk).dk-1=0 where dk is a vector of Np dimension and H the hessian matrix of f
k = 0: start from an initial point x0 and of α = α0
For each “k+1” iteration:
1° xk+1 = xk + αkdk where dk = hk + (IIhkII2/IIhk-1II2)dk-1 where hk= -grad[f(xk)]
and d0 = h0
2° Estimation of optimal value for αk such as f(xk+1) <f(xk) using a line search
method, a zero order method (Fletcher-Reeves) or an one-dimensional second
order method: αk=(hk.dk)/(H hk).dk where H is the Hessian matrix of f
3° Verified the stop criteria: IIdkII≤ε or IIxk+1-xkII/IIxkII≤δ
4° If criteria are not satisfied, return to 1°

• often used for serious optimization problems

• if not convergence it is necessary to restart from the last point results

Adinel GAVRUS

ONE ORDER METHODS

Example of use of Conjugate Gradient Method
Minimization of f(x1,x2) =f(x,y) = x4+4y4+4xy
Analytical Solution of X* = (x,y) – solution of min f
X1*=[ - 81/4 / 2, 21/4/2 ] and X2*=[81/4 / 2, -21/4/2 ]
(i.e. X1*=[-0.84, 0.59] and X2*=[0.84, -0.59])
Conjugate Gradient Fletcher-Reeves
using Hessian Matrix using linear search of α
Iter.
 Adinel GAVRUS

SECOND ORDER METHODS

• These methods can be applied to one or multi-dimensional variables space x
• Evaluation of objective function f, of its gradient f’ = grad(f) = [df/dxj], j = 1, Np
and of the hessian matrix f’’= H where Hij = [d2f/dxidxj], i = 1, Np, j = 1, Np
• Principle → the minimum can be reached following a step variation computed
incrementally from the necessary condition of minimum i.e. grad(f) = 0
(find x* solution of equations grad f = 0, reaching the minimum of f )
• Generally find local minima and can posed problems for narrow “valleys”
Principle: solve of equations system grad(f) = 0 using a general Newton method
gradf(x+∆x) = gradf(x) + [H(x)]∆x + O(x)II∆xII2
An estimation of ∆x can be obtained from gradf(x+∆x) = gradf(x)+[H(x)]∆x = 0
Consequently the increment variable ∆x = -[H(x)] -1gradf(x)
• These methods requires computation of the second order derivatives of the
objective functions (prohibitive for problems with very high complexity)
• Initially slow, they have a fast convergence around the local minimal point

Adinel GAVRUS

SECOND ORDER METHODS

NEWTON-RAPHSON Method: use the complete form of the hessian
matrix H
• Evaluation of objective function f, of its gradient f’ = grad(f) and of the
hessian matrix f’’ = H
• Minimum search following an optimal descent direction d computed
from the condition of zero gradient
k = 0: start from an initial point x0 and of α = α0 (generally α0 = 1)
For each “k+1” iteration:
1° xk+1 = xk + αlk dk where dk = ∆xk = - [H(xk)]-1 grad[f(xk)]= [H(xk)]-1 hk
2° Estimation of αlk such as f(xk+1) <f(xk) (use of a zero order method or sub-
iterations i.e. αl+1k = αlk/2 or αl+1k = βαlk (generally β = 1.1 or β = 2)
3° Verified the stop criteria: f ≤ ε or II grad[f(xk)] II≤ε or IIxk+1-xkII/IIxkII≤δ
4° If criteria are not satisfied, return to 1°
• The solution can be computed and it is an optimal point if H is positive defined
• It is easy to verify that d is a descent direction towards a minimum i.e.
grad(f) . d = -h.d = - hT[H]h < 0 if H is positive defined
 Adinel GAVRUS

SECOND ORDER METHODS

QUASI-NEWTON: approximate the hessian matrix f’’ = H with a matrix
form Q in order to obtain a descent direction during the Newton
iterations
• Evaluation of objective function f, gradient f’ = grad(f) and definition of Q

k = 0: start from an initial point x0 and of α = α0 (generally α0 = 1)

For each “k+1” iteration:
1° computation of Qk such as Qk.dk = - grad[f(xk)]
- the matrix Qk must to be positive defined and to represents a “Hessian
means”on the segment [xk-1 , xk ] i.e. Qk . (xk - xk-1) = grad[f(xk)] - grad[f(xk-1)]
2° xk+1 = xk + αlk dk where dk = ∆xk = - [Q(xk)]-1 grad[f(xk)]= [Q(xk)]-1 h(xk)
2° Estimation of αlk such as f(xk+1) <f(xk) (use of a zero order method or sub-
iterations i.e. αl+1k = αlk/2 or αl+1k = βαlk (generally β = 1.1 or β = 2)
3° Verified the stop criteria: f ≤ ε or II grad[f(xk)] II≤ε or IIxk+1-xkII/IIxkII≤δ
4° If criteria are not satisfied, return to 1°

Adinel GAVRUS

SECOND ORDER METHODS

QUASI-NEWTON (BFGS Method): the most common method using
an estimation positive defined form Q of the hessian matrix f’’ = H

• Developed by Broyden-Fletcher-Goldfarb-Shanno
• For the first iterations is very close to the descent gradient method
• Close to the Newton method around the optimum point
• Here λ represents the x parameter vector and φ is the objective function
• for k = 0, Q0 = I
 Adinel GAVRUS

SECOND ORDER METHODS

QUASI-NEWTON (GAUSS-NEWTON Method): approximate the hessian
matrix f’’ = H with a positive defined form Q using only the first order
derivatives of objective function

• Used to solve “quadratic” or least squares objective functions

• Require specific properties of matrix Q close to the optimum point (det[Q] ≠ 0)

Adinel GAVRUS

COUPLED OR MIXED METHODS

Generally, for each type of optimization problems must be chose
the adequate numerical method in function of number of variables
and of the complexity of the studied physical system
• linear or non linear problems, with or without constraints, time consuming
of the used physical model computation
• it is possible to use mixed problems or to made coupling between different
methods
Example of coupled methods
• for a high number of variables describing the objective function the
optimization can be start with a zero order method
• starting from previous solutions, for more precise values of the
parameters with available statistical estimations and rigorous estimation
of the error propagation and of the parameter’s sensitivities it is necessary
to use an one order or a second order method
 Adinel GAVRUS

COUPLED OR MIXED METHODS

Mixed Method (Relaxation): decomposition of the np dimensional
minimization problem in Np one dimensional minimization problem
• choice an initial set of parameters values x0 = (x10, x20, …xp0, …xNp0)
• choice of x = (x1, x20, …xp0, …xNp0) vector direction and minimize f with
respect to x1 (p=1) using one of the previous methods
• keep the optimal values x1 *, preserve the other parameters values and
run minimization for p = p + 1
• if not convergence change repeat from the final parameter values or
change the initial parameters values
Example in a two-dimensional parameter
space (2D)
1°- get on the x1 axis
2°- minimization with respect to x1 variable
3°- get on the x2 axis
4°- minimization with respect to x2 variable
5°- if it has not reached the minimum restart in 1°

Adinel GAVRUS

CONTENTS

INTRODUCTION

MATHEMATICAL FORMULATION OF AN
OPTIMIZATION PROBLEM
NUMERICAL METHODS

OPTMISATION ALGORITHMS

CONCLUSIONS
 Adinel GAVRUS

OPTIMISATION ALGORITHMS
In principle all optimization algorithms have the
following general form

START the initial iteration k = 0 with

a x0 considered that a first estimation of
the solution
REPEAT k = k+1
Find xk+1 from xk, xk-1, …, x0 and
objective function properties/information
UNTIL limk→∞ xk = x* and/or the objective
function is optimal

Adinel GAVRUS

OPTIMISATION ALGORITHMS
SIMPLEX NON-LINEAIRE : Algorithm de Nelder and
Mead [NOUATIN, 2000]
 Adinel GAVRUS

OPTIMISATION ALGORITHMS
SECOND ORDER METHOD : Algorithm based on
Newton-Raphson [GAVRUS, 1996]

Adinel GAVRUS

THE OPTIMIZATION PROBLEM REGARED AS AN

AUTOMATIC DESIGN OR INVERSE ANALYSIS
[CAILLETAUD & PILVIN, 1994 ; GHOUATI, 1994, GAVRUS, 1996;
BALAN, 1997;TILLIER, 1998; BOYER, 1999, ….]
(http://mines-paristech.fr, http://www.univ-fcomte.fr )
Interaction between an Optimization Algorithm and the Physical Model
Simulation : Analytical Model, Numerical Computation, Finite Element Method
 Adinel GAVRUS

THE OPTIMIZATION PROBLEM REGARED AS AN

AUTOMATIC DESIGN OR INVERSE ANALYSIS
[C3M Ljubljana - GRESOVNIK, 2000] (http://www.c3m.si )
( http://www2.arnes.si/~ljc3m2/igor/c3m/inverse/index.html)

Interaction between an Interaction between an Optimization

Optimization Algorithm and the Algorithm and a Finite Element Model
Physical Model Simulation

Optimization
Iterations

Coupling Physical
Optimization Model
Simulation

Adinel GAVRUS

DISCUSSIONS & NUMERICAL REMARKS

• Complex optimization problems requires preliminary theoretical
analysis and several numerical tests
- It is important to begin with a numerical test of a simple analytical problem
having a known optimal solution and to compare this one with those obtained by
the optimization algorithm
- To verify that the physical and numerical model of the system is robust with
respect to large variations of input variables
- To analyze the computation times and to search strategies to diminish this one
- In more complex cases to solve the optimization problem using the theory of
inverse problems (well or ill posed problems)
• Especially for a second order method it is necessary to verify the
optimization and identifiability conditions
This point requires: - To estimate the parameter's sensitivities and the hessian
conditioning number i.e. Cond (H) = µmax/µmin (µ are the eigenvalues of H)
- To search the conditions which maximize det(H)
-To satisfy Cond (H) . ε(computer) ≤ 1 (use preconditioning numerical methods)
 Adinel GAVRUS

CONCLUSIONS
THE OPTIMIZATION PROBLEMS AND THE CORRESPONDING
NUMERICAL METHODS KNOWN TODAY AN IMPORTANT
SCIENTIFIC CHALENGE
- Requires multidisciplinary competencies: theoretical and practical
backgrounds such as physical, mathematical, numerical,
engineering and experimental knowledge
- Permits to give answers of actual requirements for many research
and industrial applications
- Permits to explore sophisticated coupling existing today between
the observed experimental data and the numerical models
predictions
- Requires still researches concerning treatment of local minima
and to solve the question of existence and uniqueness of the
optimal solution
- Leads important improvements for an “Integrated Design”

Adinel GAVRUS

SYNTHESIS APPLICATIONS
Exercice I - Optimisation Non-Linéaire Sous Contraintes

• I.1° Indiquer par un graphique dans le plan (x,y) le domaine admissible

de la solution. Discussions.
• I.2° Chercher la solution du problème d’optimisation en utilisant l’écriture
du Lagrangien L(x,y,λ,µ) suivant les conditions de Kuhn-Trucker où λ et µ
représentent les multiplicateurs de Lagrange.
• I.3° Discuter les solutions obtenues et expliciter leur signification sur le
graphique.
• I.4° Préciser la solution du problème partant de la matrice Hessienne H
(x,y) de L(x,y,λ,µ), pour des valeurs bien précises des multiplicateurs λ et
µ correspondantes à chaque point optimal.
 Adinel GAVRUS

SYNTHESIS APPLICATIONS
Exercice II - Résolution d’un Problème Optimal par Equation
Differentielle (Méthode Numérique des Différences Finies)
Un corps d’aluminium avec section de forme carrée et de coté à la base de 100 m est
positionné sur une plateforme fixe et sollicité à son sommet Hf = 100 m par une
pression de compression de 5 MPa. On cherche la variation de la section avec la
hauteur A(H) pour avoir un rapport minimal entre la masse et la résistance mécanique.
• II.1° Montrer que le problème se réduit à l’étude d’un solide d’égale résistance en compression
soumis à son poids spécifique et justifier l’expression théorique de la variation de la section A(H).
• II.2° Vérifier la valeur théorique exacte de la section A(100) correspondante à 100 m d’hauteur
et la comparer avec la valeur A100T calculée dans la partie « données initiales » de l’algorithme.
• II.3° Ecrire l’équation différentielle caractéristique associée au problème et expliciter à l’aide des
expressions analytiques les termes H, ADH, AH et A utilisés par l’algorithme numérique.
• II.4° Indiquer les valeurs du pas discrétisation DH et de NH et préciser les valeurs de tous les
variables de l’algorithme de différences finies au premier passage dans la boucle de calcul.
• II.5° Donner la valeur numérique AH de la surface de la section en fin de l’algorithme, montrer
qu’elle représente la valeur correspondante au sommet HF et indiquer la valeur de l’erreur
obtenue par rapport à la valeur théorique.
• II.6° Discussions concernant l’amélioration de la précision de la méthode numérique utilisée.
Indiquer sur un graphique la tendance de la variation de l’erreur numérique avec le pas de
discrétisation.

Adinel GAVRUS

SYNTHESIS APPLICATIONS
Exercice II - Algorithme Résolution Eq. Diff. Problème Optimal
* ALGORITHME DES DIFFERENCES FINIES - PROBLEME TYPE SOLIDE DE RESISTANCE OPTIMALE - CAST3M
* DONNEES INITIALES
SIGN = 5.E6; ! Effort spécifique appliquée au sommet du corps exprimé en N/m2
GAMA= 9.81*2700.; ! Poids spécifique exprimé en N/m3 calculé par le produit de l’accélération gravitationnelle avec la masse
volumique
HF = 100.; ! Hauteur du sommet Hf exprimé en m
A0 = 100.**2; ! Surface initiale exprimé en m2 pour H0 = 0 m calculée à partir du carrée de la largeur de la base de 100 m
A100T =76.73057**2; ! Surface de la section correspondante à l’hauteur du sommet Hf exprimé en m2 déterminée par le calcul
théorique à partir du carrée de la largeur de 76.73057 m
* CALCUL PAR DIFFERENCES FINIES DE LA SURFACE A(H)
DH = 25.; ! Pas de discrétisation exprimé en m
NH = ENTI(HF/DH); ! Calcul de la partie entière du rapport HF/DH
HVAR = 0.;
A = A0;
REPETER BOUCLE1 NH; ! Répéter la boucle NH fois
HVAR1 = HVAR + DH;
ADH = -1.*(GAMA/SIGN)*A;
AH = A + (ADH*DH);
HVAR = HVAR1;
A = AH;
FIN BOUCLE1;
A100 = AH;
ERREUR = 100*DABS((A100-A100T)/A100T); ! Erreur du calcul numérique exprimé en %;
ANNEXE 1
 Adinel GAVRUS

Scientific Course
Laboratory of GCGM (EA3913)

OPTIMIZATION PROBLEMS
PART II
INVERSE PROBLEMS
PRINCIPLES AND ENGINEERING APPLICATIONS

Associate Prof. HDR Dr. Adinel GAVRUS

GMA/PSM – LGCGM (EA3913)
National Institut of Applied Sciences INSA Rennes,
University of Loire Brittany (ULB), France

Adinel GAVRUS

APPLICATION & EXAMPLE

Optimization and manufacturing of an auxetic
structure by an inverse homogenization method
Optimal Structure architecture of Compression Test of an auxetic sample
an auxetic material [2011] [S. FOREST, 2012]
 Adinel GAVRUS

CONTENTS
DIRECT AND INVERSE PROBLEM
FORMULATION

GENERAL INVERSE PROBLEM RESOLUTION

PARAMETER IDENTIFICATION BY INVERSE

ANALYSIS
MINIMIZATION ALGORITHMS USED TO SOLVE
INVERSE ANALYSIS PROBLEMS

ENGINEERING APPLICATIONS

Adinel GAVRUS

BIBLIOGRAPHIE
[1] J. V. BECK and J. A. KENNETH – « Parameter Estimation in Engineering and
Science », John Wiley & Sons Ed., 1977.
[2] M. NEWISSEN – « An Inverse Method for the Mechanical Characterisation of
Metals », Technische Univ. Eidhoven Ed., 1998.
[3] A. TARANTOLA – « Inverse Problem Theory and Methods for Model Parameter
Estimation », SIAM Ed., 2005
[4] A. GAVRUS – « Identification automatique de paramètres par Analyse Inverse »,
PhD Thesis, Ecole Nationale Supérieure de Mines de Paris, 1996.
[5] E. MASSONI – « Thermo-mechanical material properties:
new inverse identification methods », Coll. 10ème Anniv. of LAMIH, 23 Mars 2005
[6] R. FORESTIER – « Développement d’une Méthode d’identification de Parametres
par Analyse Inverse couplée avec un Modèle Eléments Finis 3D », PhD Thesis,
Ecole Nationale Supérieure des Mines de Paris, 2004.
[7] S. DIOT – « Caractérisation expérimentale et numérique du comportement
dynamique des matériaux », PhD Thesis, Institut National des Sciences Appliquées
de Rennes, 2003.
 Adinel GAVRUS

DIRECT AND INVERSE PROBLEM

FORMULATION

DIRECT PROBLEM
“well known question leads to correct answer”

INVERSE PROBLEM
“known answer but unknown question”

Adinel GAVRUS

General Direct Problem

CAUSE EFFECT
Input or Known Data Output or Unknown
(Parameters) Data (Observables)

Direct Model : F

General Inverse Problem

CAUSE ? E EFFECT
Output or Unknown Input or Known Data
Data (Parameters)

Inverse Model : φ ou F-1

 Adinel GAVRUS

ENGINEERING PROBLEMS

• Optimization of structures: design, mass, energies, price, ….

* by minimization or maximization of design variables or cost
functions
- requires computation models
-in more complex cases: solve an optimization problem formulated that an inverse
problem

• Search of unknown boundaries conditions or unknown functions

* requires additional or needed observed quantities
-solve of a functional equation
- in more complex case: parameterization of unknown functions and parametric
identification (solve of an inverse problem) starting from the optimal comparison
between data obtained from the direct model and the needed ones

Adinel GAVRUS

ENGINEERING PROBLEMS
Calibration of reduced or simplified models with respect to
complex ones
* requires predictions of reduced or simplified model close to the
complex one
- Introduction of unknown structural parameters for the simplified model
- solve of an inverse problem starting from the comparison between data obtained
from reduced or simplified models and the complex one
Search of unknown design or material parameters
* requires additional measurement data
- solve of an inverse problem starting from the comparison between data obtained
from the direct model and the measured ones

Parameter Identification by Inverse Analysis

(Developed in Mechanical Engineering from 1990 and highly used from 2000)
 Adinel GAVRUS

DIRECT PROBLEM FORMULATION IN

ENGINEERING

Research of the system response undergoing known input data

• Formulation of the computational model R(P,X,U,…)=0
• Definition of the input-output relationships Mc = F (P, U,…)

Adinel GAVRUS

INVERSE PROBLEM FORMULATION IN

ENGINEERING

Definition of the conditions for which the system response Mc is

close to the experimental one Mexp.
I Method: Formulation of inverse relationships ⇒ P= F-1(Mc , Mexp, …)
II Method: Definition of a parametric correction (inverse back up) ⇒ ∆P
 Adinel GAVRUS

GENERAL INVERSE PROBLEM

RESOLUTION

Adinel GAVRUS

MATHEMATICAL FORMULATION OF AN
INVERSE PROBLEM
(TIKHONOV et al., 1974; BECK et al. , 1974; TARANTOLA, 2005)
In the simplest situation, when measuring an n - dimensional data vector M
(considered an element of a linear space D), the observed values Mex are obtained
with some uncertainties generally assumed to be of the Gaussian type, described by a
covariance matrix CD . The associated Gaussian probability ρD(M) is:

ρD(M) = ( (2π) det CD )-1/2 exp[− 0.5 (M − Mex) CD(M − Mex)]

n T

2
In the case where CD=[1/σκ ]I we obtain:
2 2
ρD(M) = [(2π) (Πσκ )]-1/2 exp[− 0.5 (Mk − Mkex) /σκ ]
n 2

In order to have a prediction Mc close to Mex we must to maximize ρD(Mc) i.e. Min Φ:
 Adinel GAVRUS

SOLUTION OF A LINEAR SYSTEM BY AN

ANALYTICAL INVERSION

Adinel GAVRUS

SOLUTION BY LINEAR REGRESSION METHOD

 Adinel GAVRUS

Application of a Linear Regression Method

Adinel GAVRUS

SOLUTION BY NON-LINEAR REGRESSION METHOD

IDENTIFICATION
DE PARAMETRES
(OPTPAR)

P ← P0
i=0

MODELE DE
i ← i+1 CALCUL
c
P ←Pi+1 c dM
M (P) ,
dP
NON
OUI
∆Pi FONCTION COUT
≤δ c ex
M (P)-M 
Pi
REGRESSION
NON
Pi+1←Pi + ∆Pi Error≤ε

(GAUSS-NEWTON) OUI

VALEURS DE
P=(A,B,D,n, )
P
IDENTIFIEES

(Parameters of the constitutive material law)

 Adinel GAVRUS

SOLUTION BY INVERSE ANALYSIS

Input Data Paramètres initiaux

Itération n

Essais Exp Modèle de Calcul

Numerique ou Incremental
Mex Mc, dMc/dP Non
Oui
Stagnation
Optimisation
Φ (P)
Calcul d’un nouveau jeu de
Error ≤ ε
Non
paramètres
P
Oui

RESULTAT : Paramètres identifiés

Adinel GAVRUS

PARAMETER IDENTIFICATION
BY INVERSE ANALYSIS
 Adinel GAVRUS

INVERSE ANALYSIS THEORY &

MATHEMATICAL FORMULATION
• For complex problem the evaluation of the inverse function F-1 is
not possible
• Solution: Minimisation of the euclidian distance between the computed and
the experimental values

– If Mc= F(P) the problem can be then formulated as

– In the case of a least square sense evaluation

Adinel GAVRUS

GENERAL DIAGRAM OF AN INVERSE

PROBLEM RESOLUTION
 Adinel GAVRUS

EXPERIMENTAL SPACE BUILT

- Choice of a representative experimental data space
(Optimal set of operating conditions)
- Analysis of the experimental data sensitivity
(Parameters Sensitivity Computation)

COMPUTATION MODEL
- Analytical Model
(solve of the Inverse Problem by linear or non-linear regression)
- Numerical or Finite Element Model (Inverse Analysis)
OPTIMISATION ALGORITHMS
- Choice of an adequate cost function
- Use of an iterative numerical procedure

Adinel GAVRUS

MINIMIZATION ALGORITHMS
USED TO SOLVE INVERSE
ANALYSIS PROBLEMS
 Adinel GAVRUS

Zero order methods

•Genetic algorithms (Holland, 1975; Ghouati & Gelin, 1998)

•Simplex Methods (Nealder & Mead, 1965; Nouatin, 2000)

• Neural-Networks (Mori & al., 1996), Evolutionary Strategy,

Fuzzy Logic (Aushby, 1997)

Sensitivity analysis not required

It’s not necessary to know very well the direct model

- Generally reach global minimum
- Require lots of cost function evaluations
- Relatively slow for a large number of parameters

Adinel GAVRUS

Descent direction methods

For optimization problem without constrains

Fast Convergence
Require Sensitivity Analysis
Generally provide local minimum
Provide information about the stability of the inverse problem
 Adinel GAVRUS

Basic method
Easy to implement but not very efficient
Rapid convergence far from the optimum
Slower in the neighborhood of the optimum

Conjugate gradient
Need a good choice of βn

Newton-Raphson (NR)
More efficient close to the optimum

Quasi-Newton or BFGS (Mahnken and Stein, 1995)

Idem as NR method but using an approximate evaluation of

Drawback slow for ill posed problems=>stabilisation techniques

Adinel GAVRUS

Gauss-Newton (GAVRUS et al. 1994, 1998)

Based on a NR algorithm, widely used for least squares cost functions

This approximation becomes more and more better arriving close to the optimum
For linear direct model close to the optimum => idem as NR i.e quadratic convergence
For non linear direct model => non convergence => use of line search criteria
 Adinel GAVRUS

Gauss-Newton Algorithm

Adinel GAVRUS

Stability of descent direction methods

Identifiably conditions (Beck & Arnold, 1977)
ex
The set of parameters P is identifiable from the observable M knowledge
if in a neighborhood of the minimum of Φ

Gauss-Newton matrix Q =[S]T[S] => keep curvature of the cost function close
to the optimum

Non identifiably in the Beck sense and Φ is ill-conditioned

Amplitude indicator of parameters cross effects
If change of parameters by P’=P/P0 it is possible to obtain a
well-conditioned Gauss-Newton matrix
 Adinel GAVRUS

Example for a Norton-Hoff thermo-viscoplastic law

• The Gauss-Newton step has an upper bound (Forestier, 2004)

µi eigenvalues of Gauss-Newton matrix

∆P Gauss-Newton step
• If Cond(Q)εM >1 (det(Q) close to 0)=> Gauss-Newton matrix ill-conditioned
=> large values of ∆P => the method may become unstable
• If change of parameters by P’=P/P0 the new matrix Q’ have Cond(Q’)<Cond(Q).
Previous example:

Adinel GAVRUS

Improvement of Stability: Regularization Techniques

TIKHONOV Method
(Tikhonov et al. , 1974)
• Choice of a constant regularization parameter λ
• Search of the optimum values in the neighborhood of the P0

• Well-conditioned matrix
• Very slow convergences
 Adinel GAVRUS

LEVENBERG Method

• Choice of a constant regularization parameter λ

• Search of the optimum in the neighborhood of the previous parameter values

• Good conditioning of the inverse problem

• Fast convergences if a good choice of the regularization parameter

Adinel GAVRUS

LEVENBERG MARQUARDT Method

(Schnur and Zabaras, 1992)
This method filters the lower eigenvalues of Gauss-Newton matrix

Is still a descent direction method if (λ +min(µi)) is positive

The choice of λ is critical (different at each iteration n), but can be chose to
reach a 0 values in the neighbour of the optimum point

EXAMPLE
 Adinel GAVRUS

LEVENBERG Modified Method

(Forestier, 2004)
• In this case the regularization parameter λ is constant
• This method uses the available information about the parameter ranges

• The Matrix D is close to an a priori information about the parameter variation

domain (Bayes estimator)
• Penalises large steps with respect to the constraints imposed by the user

Adinel GAVRUS

Sensitivity Analysis or
How to compute the
sensitivity matrix S
 Adinel GAVRUS

Finite Differences Method

(Cailletaud and Pilvin, 1993; Schnur and Zabaras, 1992;
Khoddam, 1995; Tillier, 1998)
Easy to implement but Slow (for non-linear problems)
Sometimes unstable if remeshing (Forestier et al., 2002)

Adinel GAVRUS

Analytical Method
(Tortorelli, 1991, Gavrus et al., 1994, Gelin and Ghouati, 1996, …)
Very accurate and widely used in the literature
Calculation of some derivatives may be fastidious
Used if the number of model parameters is small as compared to the number of model
internal variables

Direct problem =>

Differentiation of state equation =>

Remark: The number of the functional R describing the direct model is

equal to the number of internal variables U
 Adinel GAVRUS

Adjoint Method
(Pironneau, 1984; Jameson, 1988; Fourment et al. , 1996)
Many important uses in engineering analysis
The number of solution evaluations is in order of the number of input variables to
be varied
Applied if the number of model parameters is more important that the number of
model internal variables

Direct problem =>

Adjoint Solver =>

With

Adinel GAVRUS

Semi-Analytical Method
(Tortorelli and Michaleris, 1994; Massoni et al., 2001)
Stable if remeshing
All terms are computed by the direct model
Easy to update with the direct model

Direct problem =>

Differentiation of state equation =>

With
 Adinel GAVRUS

Parameter Statistics

Adinel GAVRUS

Estimation of the Parameter Identification Error

(Beck, 1977)
• Computation of the parameter confidence interval
• Using the covariance matrix of measurement data CD
• Computation of the parameter covariance matrix CP
 Adinel GAVRUS

Numerical Analysis of an Inverse

Problem Robustness

Adinel GAVRUS

CONVERGENCE ANALYSIS
-Choice of a virtual parameter set
- Generate virtual experimental data from numerical model
- Starting from far the virtual parameter values and verify if the
inverse model generates the previous a priori known parameter set

STABILITY ANALYSIS
-Study the behaviour of the inverse problem with respect to
statistic error on experimental data
-Fit virtual experimental data with different statistic noise

PARAMETER SENSITIVITY ANALYSIS

- Computation of the parameter sensitivity matrix
- Analysis of parameter correlation factor
- Search to maximize Det {[S]T[S]}
 Adinel GAVRUS

Exercice: Identification of a Hardening Law

It is proposed to identify the parameters P = {K, n} directly in the constraints space
of a law describing the behavior of work hardening curve of a metallic material in
the form given by the Ludwick law:

where represents the Von Mises stress shown than the value computed from the
model (F (P) = ), is the elastic yield stress, a priori assumed known, K is
called the consistency of the material and n the work hardening coefficient. The
cumulative plastic deformation can be seen as the main variable of the model
defining the dependence of the stress .

Make the identifiability study of the system compared to the parameter vector
P = {K, n} based on the stress measurement for a known value of plastic
deformation . Give the minimal requirement experimental data space so that
the problem becomes identifiable (give the minimum number of experimental
stress points on the hardening curve). In this case indicate the values of
parameters K and n according to the previous precised measurement points.
Obs : the least squares cost function is written in this case as follows:

Adinel GAVRUS
 Adinel GAVRUS

Engineering Applications

Adinel GAVRUS

IMPROVEMENT OF ENGINEERING PROBLEMS

• Development of predictive numerical tools
• The most employed: the Finite Element Model (FEM)
Example: analysis and optimization of material forming processes
(collaboration LGCGM/PSM /INSA- GEVELOT Extrusion Laval)
 Adinel GAVRUS

MATERIAL FORMING PROCESS DESIGN

Plastic
Strain

Tool/piece contact,
Geometry friction
1.5
1.3
Initial Temperatures Thermal Transfert 1.1
0.9
0.7
0.5
Tools Velocities Material 0.3
0.1
Behavior

Min = 0.14665
Max = 4.1782

Adinel GAVRUS

GENERAL NUMERICAL MODEL FOR THE

SIMULATION OF A MATERIAL FORMING PROCESS

A lot of unknown input parameters

 Adinel GAVRUS

PARAMETER IDENTIFICATION BY
INVERSE ANALYSIS (OPTPAR SOFTWARE)

GAUSS-NEWTON

Identified Parameters

Adinel GAVRUS

OPTIMIZATION AND INVERSE ANALYSIS OF

FORMING PROCESSES
(Design Parameters Identification by Inverse Analysis)

• Weight Minimization of the Material Billet and Minimization of Forging

Loads
(TRANSVALOR & CEMEF/ENSMP, 2009)

• Tool Shape Optimization for Many-Steps Materials Forging

(Balan, 1996 – PhD Thesis CEMEF/ENSMP; Gresovnik, 2000 - PhD Thesis University of
Wales Swansea - UK)

• Optimization of Heating Variables for Forming Operations

(Gresovnik, 2000; Rodicz, 2003 - C3M University of Ljubljana - Slovenia)

• Identification of Cracks and Damage Optimization

• Optimal Arrangements of Physical Entities Defining the Material

Microstructure
(Jelovsek, 2003 - C3M University of Ljubljana - Slovenia)
 Adinel GAVRUS

INVERSE ANALYSIS OF MECHANICAL TESTS

(Material Parameters Identification)
• Tensile and Torsion Test (Cold or Hot)
(Gavrus, 1996 – PhD Thesis CEMEF/ENSMP; Diot, 2003 - PhD Thesis LGCGM/PSM -
collaboration CEMEF)

• Cylindrical Compression Test

(Gavrus, 2000; Caestecker, 2003 - PhD Thesis LGCGM/PSM -
collaboration Ecole Centrale de Nantes)

• « Dumbbell » Specimen Upsetting Test (Cold or Hot)

(Gavrus, 2003; Diot et al. , 2009 - collaboration GEVELOT Extrusion Laval)
• Dynamic Upsetting Test (Cold or Hot)
(Gavrus, 2008 - HDR; Caestecker, 2003; Diot, 2003; Davoodi, 2006 - PhD Thesis
LGCGM/PSM - collaboration Tampere University - Finlande)
• Forward Extrusion Friction Test
(Gavrus, 2002; Pham, 2011 - PhD Thesis LGCGM/PSM)
• Erichsen Test
(Gavrus, 2010 - collaboration ULBS & UGAL – Romania)

Adinel GAVRUS

TENSILE TEST
(Gavrus, 1996; Diot, 2003 – PhD Thesis LGCGM/PSM)

F
 Adinel GAVRUS

Experimental data corresponding to AA5083 alloy and the

description of the Finite Element Model
V

Adinel GAVRUS

V Parameter Identification by INVERSE ANALYSIS

 Adinel GAVRUS

Improvement of the Tensile Test

(Gavrus, 1996, Tillier, 1997 – CEMEF; 2010 – LGCGM/PSM)

Adinel GAVRUS

Parameter Identification for a Solid Polyethylene

by Non-Linear REGRESSION and by INVERSE ANALYSIS
Parameters Identification

Initial Values Analytical Inverse Analysis I Inverse Analysis II

Method applied of the loads applied of the loads and
(Regression) (TORRAO) radius
(TORRAO)
K0[MPasm] 10. 11.014 6.031 6.039

w 30. 30.786 286.53 322.06

h 0.5 0.515 0.896 1.001

n 2. 2. 1.255 1.215

m 0.1 0.156 0.103 0.121

R2 - 0.89 0.945 0.954

(F, Rmin) (0.864, 0.907) (0.971,0.908) (0.953, 0.956)
 Adinel GAVRUS

Comparison of the estimated True STRESS-STRAIN

Flow Curves

Adinel GAVRUS

Comparison of the predicted loads

FE from
FE from Classical
Parameter Identification INVERSE ANALYSIS II
 Adinel GAVRUS

Comparison of the predicted minimal radius

FE from Classical FE from

Parameter Identification INVERSE ANALYSIS II

Adinel GAVRUS

UPSETTING TEST
(Gavrus, 2000; Caestecker, 2003 – PhD Thesis LGCGM/PSM)
Experimental Setup Finite Element Simulation
(FORGE2)

• Taking into account the contact

Analytical Model friction and the heterogeneity of
• Neglecting the contact friction the deformation
and the heterogeneity of the • Computation of the load
deformation variation
• Inverse Analysis applied of the
loads
 Adinel GAVRUS

Identified Parameters for Johnson-Cook law and

Comparisons of Methods and Results

(m=1)

Adinel GAVRUS

Prediction of the Von Mises Stress for the two Methods

(m=1)

• non-linear regression • Inverse Analysis • Inverse Analysis

(without friction) (without friction) (with friction)
 Adinel GAVRUS

Improvement of the Upsetting Test

(Gavrus, 2006; Diot et al, 2009; Davoodi, 2006; Pham, 2011 – PhD Thesis LGCGM)

“DUMBELL” Specimen

d
h/d < 1.5
• ‘‘Dumbbell’’ Specimen – equivalent (d x h) cylindrical specimen
• Advantage : avoid the influence of friction between tool and work-
piece on the compression load
• Possibility to conduct tests at high temperatures and to control
the initial specimen temperature by induction heating
• Analytical Model not exist

Adinel GAVRUS

Finite Element Analysis (FORGE2)

⇒ The compression load –

independent of the friction
between tool & workpiece
 Adinel GAVRUS

Inverse Analysis for an AA5083 aluminium alloy

‘‘Dumbbell’’ upsetting test: (for ε < 30%) =>

The compression load is almost independent
of the compression velocity
Identification by Inverse Analysis

The predicted FE compression load curve is very close to the

experimental one (maximum error around of 5%)

Adinel GAVRUS

SHPB TEST
(Gavrus, 2000-2008 – HDR; Caestecker, 2003; Davoodi, 2006 – PhD Thesis LGCGM)
(collaboration with Mechanical Laboratory of Tampere University - Finlande)

CLASSICAL PRINCIPLE OF COLD OR HOT SHPB UPSETTING

Classical Analytical
Analysis using Elastic
Wave Theory to estimate

• Neglecting the contact friction and the •Use of non-linear REGRESSION to

heterogeneity of the deformation identify law parameters
 Adinel GAVRUS

FEM SIMULATION AND INVERSE ANALYSIS OF SHPB TEST

COMPLETE NUMERICAL SIMULATION

THERMAL SIMULATION INVERSE ANALYSIS OF A REDUCED MODEL

(CAST3M) (FORGE2®)
(Initial Heating: 500°C and Cooling 10 s)

(Tmax = 483.72 °C, Tmin = 482.723°C)

Adinel GAVRUS

PARAMETER IDENTIFICATION BY INVERSE ANALYSIS OF

THE SHPB REDUCED MODEL
Parameters Identified Values using different constitutive laws (Johnson-Cook, Cowper-Symonds,
ASH, …) for the AA5083 aluminium alloy (A= σ00 = 142 MPa, Tr=293 °K, Tm=933 °K)
Parameters P
Q (%) Nb
B n0 n1 n C0 C1 C m
Iter
Law Init. 500, - - 0,250 - - 0,001 2,00 29
I (JC) final 118, - - 0,281 - - 0,093 1,935 7,07
10
Law Init. 118, 0,280 0,00 - 0,090 0,00 - 1,930 7,17
II final 122, 0,504 -0,0005 - 0,166 -0,0002 - 2,996 6,99
7
Law Init. 500, - - 1,00 - - 0,001 2,00 13,9
III final 98,6 - - 14,382 - - 0,073 1,969 6,61
17
Law Init. 98,6 14,380 0,00 - 0,073 0,00 - 1,970 6,62
IV final 99,3 1,936 0,026 - 0,148 -0,0002 - 4,692 6,29
16
Law Init. 120, - - 0,300 - - 0,100 2,000 15,9
V final 82,4 - - 0,401 - - 0,104 1,970 7,47
11
Law Init. 82,4 0,400 0,00 - 0,100 0,00 - 1,970 7,53
VI final 108 1,083 -0,0015 - 0,224 -0,0003 - 1,975 7,12
10
 Adinel GAVRUS

COMPLETE SIMULATION USING AN IMPLICITE

NUMERICAL FEM (FORGE2®)

COMPARAISON BETWEEN COMPLETE MODEL (a) AND REDUCED MODEL (b)

a) b)

Adinel GAVRUS

COMPLETE SIMULATION USING AN EXPLICITE

NUMERICAL FEM (ABAQUS)

V=14,5 m/s, T=200 °C

 Adinel GAVRUS

TORSION TEST
(Gavrus, 1996 – PhD Thesis CEMEF/MinesPrisTech)

FINITE ELEMENT SIMULATION WITH TORRAO software

(developed at CEMEF, Ecole des Mines de Paris - ENSMP)

Adinel GAVRUS

Classical Analysis
(FIELDS & BACKOFEN)
• all mechanical variables depend only on the radius coordinate
• neglecting the self heating of the specimen
 Adinel GAVRUS

Analysis of the torsion test by Finite Element Simulation

Adinel GAVRUS

Parameter Identification for a Solid Polycarbonate

Non-linear REGRESSION and INVERSE ANALYSIS
Parameters Identification

Initial Values Analytical Model Inverse Analysis

(Regression) applied on the torques

K0[MPasm] 5.8 8.07 4.502

w 1. 0.84 3.505

c 10. 9.28 7.395

d 10. 9.60 13.470

h 0. -0.075 0.328

n 2. 2. 2.

m 0.001 0.002 0.033

R2 - 0.884 0.972
 Adinel GAVRUS

Comparison of the estimated True STRESS-STRAIN Flow

Curves

Adinel GAVRUS

Comparison of the predicted TORQUES

FE from Classical FE from

Parameter Identification INVERSE ANALYSIS
 Adinel GAVRUS

Improvement of the Torsion Test Analysis

Torsion test applied to plastic metallic materials
Identification of complex laws
(Gavrus et al., 1996, 1999; Dal Negro et al., 2001 – PhD Thesis CEMEF)

SICO test

Parameter Identification of a generalized Norton-Hoff law

(Dal Negro et al., 2002)

Adinel GAVRUS

Hot Torsion Tests for an INCONEL Alloy

Torque (Nmm)
12000

10000

8000

6000

4000

2000

0 Rotations number
0 0.5 1 1.5 2 2.5 3 3.5 4

T=900°C,Np=1 /s T=900°C,Np=0.2 /s T=900°C,Np=0.06 /s

T=1050°C,Np=1 /s T=1050°C,Np=0.2 /s T=1050°C,Np=0.06 /s
T=1200°C,Np=1 /s T=1200°C,Np=0.2 /s T=1200°C,Np=0.06 /s
 Adinel GAVRUS

More Complex Constitutive Models

(Gavrus, 1996 – PhD Thesis CEMEF MinesParisTech)
“Constitutive equations are the vehicle by which our knowledge of material
behavior enters engineering design” (U.F. Kocks)
W
Softening
Work Saturation
hardening
1

Adinel GAVRUS
 Adinel GAVRUS

Identification results obtained by Inverse Analysis

Torque (Nmm)
Law A Torque (Nmm)
Law D

Negative
work hardening !

Rotations
Error = 5.0% Rotations

Error = 9.4% Error = 8.7%

Torque (Nmm)
Law G

experiment Bad modeling of

computation
work-hardening

Rotations

Adinel GAVRUS

MORE COMPLEX PROBLEMS: UPSETTING IMPACT TESTS

(Diot, 2003 – PhD Thesis LGCGM/PSM)

Material : 50CrMo4
Jack

Furnace

Specimen

Extension
Bar

Velocity : 3 m/s
Cooler
Load cell
 Adinel GAVRUS

Experimentl upsetting test for the 50CrMo4 steel

* Different configurations

a) b) c) d)
Extension Bar

Cooler
Load cell

* Two velocities : 3 et 4 m/s

Adinel GAVRUS

Effort (N) Effort (N)

Configuration A Configuration B

Temps (s) Temps (s)

Effort (N) Effort (N)

Configuration C Configuration D

Temps (s) Temps (s)

 Adinel GAVRUS

FFT of experimental results Study of the natural frequencies of

the entire mechanical system by
finite element simulations
Load Cell Data:

Stiffness K
Mass m
Jam Dimensions SxH

Young Modulus

Density
The three first own
frequencies [1] G. HAUGOU et al, « Mise en œuvre d ’une méthode originale pour
l ’identification des modes propres et la conception d ’un montage
corresponding to d ’essais dynamiques ». International Conference on Multidisciplinary
Design in engineering, CSME-MDE 2001, 21-22 November 2001
each configuration

Adinel GAVRUS

FE Simulations with CASTEM2000

Computation of the
three first own
frequencies
corresponding to each
configuration
 Adinel GAVRUS

Comparisons FE Computation / Experiment

B
A)
A
Extension Bar C
D
Cooler
Load Cell

Own Fréquencies (kHz)

First Second Third

A
B
C
D

Adinel GAVRUS
 Adinel GAVRUS

Numerical FE Model

Rheological law of the 50CrMo4 steel:

avec et

Friction specimen/clamping tool : skintight

Thermal conditions : isothermal à 20°C

Frequency offset between

the model and experience:

Modele Calibration

Adinel GAVRUS

Calibration of the Numerical Model

Initial Parameter Set

4 configurations Eeq, ρeq, Erallonge, ρrallonge, ti

Code
Experiences Cast3M
NO
Fexp(t) Fcal(t)
YES
Stagnation
Comparison :
Least Squares
Function Gauss-Newton Algorithm :
New Parameter Set
NO
Error ≤

YES

Results : Optimal Parameter Ste

 Adinel GAVRUS

Young Modulus

Density

Adinel GAVRUS

The used mechanical setup

Effort (N) Effort (N)
Résultats

Temps (s) Temps (s)

Effort (N) Effort (N)
 Adinel GAVRUS

Rheological Parameter Identification

Goal : to define with P2 the rheological parameters

1) inelastic response of the material

Fspecimen STEP II
V

One
Dynamic ∆L Step
Test
2)Vibration responses of the mechanical setup
STEP I
Fcell
capteur

Computation
Time
∆L

Adinel GAVRUS

First Step: Solve an Inverse Problem

Fspecimen(t) Fspecimen(∆L)
∆L

Approximation :
What is known
- polinomial of degree 6
Fcell(∆L) -
Fcell

-piecewice linear
∆L
Inverse Problem
Search the parameter’s values P1 of
minimising in a least squares sens the error
between and
 Adinel GAVRUS

Second Step: Parameter Identification by Inverse Analysis

What is known

Inverse Analysis
Search of parametres values P2 of
minimising in a least squares sens the error
between and

Adinel GAVRUS

Example: Dynamic Compression for a pseudo-real material

Rheological behavior Law and the initial parameter set:

avec
et
 Adinel GAVRUS

Step I: Identification of the load at the specimen interface

(Inverse Problem)
Fspecimen(t) Fspecimen(∆L)

Fcell(∆L)

Déplacement (mm) Déplacement (mm) Déplacement (mm)

Piecewice linear Polynomial of degree 6

Adinel GAVRUS

Step II: Identification of the rhéological parameters

(Inverse Analysis)

Piecewice Linear Polyomial of degree 6

A = 702.7407 MPa A = 701.6244 MPa A = 658.2955 MPa

0.391 % 0.232 % 5.96 %
n = 0.3046295 n = 0.3034776 n = 0.2817769
1.54 % 1.16 % 6.07 %

V
A = 701.6244 MPa
n = 0.3035

< 1.5 %
n = 0.3
 Adinel GAVRUS

Real Application
(XC10 Steel)
Load (N)

Compression tests

Displacement (mm)

Adinel GAVRUS

Step I: Identification of Specimen Load

T = 20°C T = 200°C T = 400°C
Load (N) Load (N) Load (N)

Displacement (mm)
Displacement (mm) Displacement (mm)
Load (N) Load (N) Load (N)

Displacement (mm) Displacement (mm) Displacement (mm)

 Adinel GAVRUS

Experimental Loads at Specimen Interfaces

Load (N)

Displacement (mm)

Adinel GAVRUS
Titre de diapositive de contenu

Step II: Material Law Identification

K = 324.931062 MPa m0 = 0.08975

Identified Parameters
n0 = 0.215397 m1 = -1.17e-4 K-1
n1 = -8.7e-5 K-1 β = 186.400672 K
T = 20°C T = 200°C T = 400°C
Load (N) Load (N) Load (N)

Displacement (mm) Displacement (mm) Displacement (mm)

 Adinel GAVRUS

COMMERCIAL SOFTWARE
TODAY EXIST A LOT OF COMMERCIAL NUMERICAL SIMULATION
CODES WHICH INTEGRATE POWER OPTIMIZATION AND
IDENTIFICATION TOOLS
- ANSYS
- COMSOL Multiphysics
- LS-Dyna (LS-Opt)
- ALTAIR OptStruct
- FORGE, Zebulon, Identify
IF NECESSARY IS POSSIBLE TO COUPLING THE COMMERCIAL
NUMERICAL CODE WITH AN OPTIMIZATION SOFTWARE
- Mode Frontiere
- MATHLAB
- SiDoLo, …

Adinel GAVRUS

CONCLUSIONS
THE INVERSE ANALYSIS IS TODAY AN IMPORTANT AND
POWERFUL TOOL
- Modular Structure
- Openness, Flexibility, Portability
- Permits to explore more sophisticated mathematical models
- Find a more accurate parameter identification for more
complex experimental investigations

POSSIBILITY TO IMPROVE THE PARAMETER PRECISION

KNOWLEDGE TOGETHER WITH THE IMPROVEMENT OF THE
DIRECT NUMERICAL MODEL

PERMITS TO SOLVE AND IMPROVE ACADEMIC SCIENTIFIC

RESEARCHS AND INDUSTRIAL ENGINEERING DESIGN PROBLEMS