Materialia 26 (2022) 101620

Contents lists available at ScienceDirect

Materialia
journal homepage: [Link]/locate/mtla

Full Length Article

Charge-density based convolutional neural networks for stacking fault

energy prediction in concentrated alloys
Gaurav Arora a, Serveh Kamrava b, Pejman Tahmasebi b, Dilpuneet S. Aidhy a,c,∗
a
Department of Mechanical Engineering, University of Wyoming, Laramie, WY 82071, USA
b
Colorado School of Mines, Golden, CO 80401, USA
c
Department of Materials Science and Engineering, Clemson University, Clemson, SC 29634, USA

a b s t r a c t

A descriptor-less machine learning (ML) model based only on charge density images extracted from density functional theory (DFT) is developed to predict stacking
fault energies (SFE) in concentrated alloys. The model is based on convolutional neural networks (CNNs) as one of the promising ML techniques for dealing with
complex images and data. Identification of correct descriptors is a key bottleneck to develop ML models for predicting materials properties. Often, in most ML models,
textbook physical descriptors such as atomic radius, valence charge and electronegativity are used as descriptors which have limitations because these properties
change in concentrated alloys when multiple elements are mixed to form a solid solution. We illustrate that, within the scope of DFT, the search for descriptors can
be circumvented by electronic charge density, which is the backbone of the Kohn-Sham DFT and describes the system completely. The performance of our model is
demonstrated by predicting SFE of concentrated alloys with an RMSE and R2 of 6.18 mJ/m2 and 0.87, respectively, validating the accuracy of the proposed approach.

1. Introduction [29] used 3D CNN for predicting the microstructure of the materials. In
this work, we exclusively use charge density images obtained by per-
Machine learning (ML) is rapidly becoming an integral part of ma- forming the DFT simulation to train a CNN model and predict stacking
terials research [1–7]. Its contributions from the elucidation of the fun- fault energy (SFE) in Ni-based concentrated alloys.
damental mechanisms to the design of materials are being universally SFE provides a measure of deformation mechanism in alloys [30]. A
recognized [8–16]. Besides the training data, the accuracy of the ML rare combination of high strength and high ductility has been recently
model, especially in computational materials research, hinges on de- observed in multi-principal element alloys (MPEAs) and the underlying
scriptors that act as means to incorporate physics for property predic- reason is attributed to twin formation [31–35]. On the one hand, co-
tion [2,5,17–20]. It is a common practice to assess the suitability of the herent twin boundaries act as barriers to dislocations inducing strength,
descriptors during the ML model development and it is often observed whereas on the other, they provide pathways for the glide of dislocations
that a model relies on a weighted combination of multiple descriptors inducing ductility [30]. Formation of large number of twin boundaries
[17,21–23]. Identification of correct descriptors thus remains a key chal- is correlated to SFE, where it is observed that lower SFE favors twin for-
lenge in the ML materials community. In this work, within the scope of mation [30,36–39]. This lowering has been further tied to alloy compo-
density functional theory (DFT), we demonstrate that the search for de- sition that has opened a route to manipulate SFE [35,40–42]. For exam-
scriptors can be circumvented by electronic charge density, which is the ple, it has been shown that SFE increases with Ni and Mn whereas it de-
backbone of the Kohn-Sham DFT and describes the system completely creases with Co in FeCoNiCrMn [38,40,43]. Similar trends for Ni and Co
and enables materials property predictions. have also been observed in NiCoCr and various other alloys [38,44,45].
Image-recognition based ML approach is used in a wide variety of Despite these observations, due to the structural complexities and large
applications. Common examples include face recognition [24], lane- compositional space in MPEAs, there is very little understanding of the
assisted driving [25], food contaminant detection [26], image recog- guiding principles that drive the coupling between composition and SFE.
nition in Google search, and cancer-cell detection [27]. One of the com- The underlying descriptors remain largely unknown and the SFE predic-
mon image recognition methods is convolutional neural network (CNN) tions in MPEAs from conventional ML is not trivial. As a result, tailoring
which is designed specifically for image classification and computer vi- SFE via composition remains an open challenge. In our recent work [46],
sion. CNNs provide scalable approach for identifying patterns in im- we demonstrated that charge density plays a central role in determin-
ages to make predictions. This approach has also been previously used ing the SFE of alloys. In this work, we leverage the correlation between
in making predictions. For example, Ziletti et al. [28] used CNN for charge density and SFE to develop a CNN-based framework that enables
classifying crystal structures by using diffraction images. Cecen et al. SFE prediction without the need to search for descriptors.

∗
Corresponding author.
E-mail address: daidhy@[Link] (D.S. Aidhy).

[Link]
Received 2 July 2022; Accepted 1 November 2022
Available online 11 November 2022
2589-1529/© 2022 Acta Materialia Inc. Published by Elsevier B.V. All rights reserved.
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

Fig. 1. Charge density distortion due to addition of different elements in Ni. (a) Single doped Sc and Y, (b) double doped Sc & Hg and Y & W, and (c) six Cr atom
placed differently in Ni. Corresponding SFE values are shown; unlabeled atoms represent positions of Ni atoms. The disparity in charge distribution is leveraged to
train CNN model.

The rationale for image recognition approach is anchored in the com- output value. Each time a model is trained, these inputs are updated
positionally driven large charge density variations. Fig. 1 shows the accordingly based on the performance of the model.
charge density profiles of shearing layers in different alloy chemistries The images are converted into arrays that represent the pixel values
in a given supercell of 108 atoms (see description of the supercell in of the color. The convolutional layer performs a dot product operation
the methodology section). A significant difference in the charge density between the two matrices to produce a third matrix of required dimen-
is observed between Sc and Y doped Ni and the associated SFE vari- sions. In CNN, convolution is done over the input matrix by another
ation is also observed. Similar variation is observed in Ni doped with matrix named kernel. This operation is performed over the entire ma-
two different dopants i.e., Sc & Hg and W & Y. Even within the same trix with a predefined step size. The resultant matrix is known as the
alloy chemistry (i.e., Ni-Cr), a different placement of Cr atoms results in feature map as it contains numerical information about the image pat-
different charge densities resulting in widely different SFEs. The CNN terns. The output from the convolutional layer is used as an input for
model is used to learn these patterns of charge density distributions and the pooling layer to downsize the sample from the feature maps with-
the associated SFEs which then enable SFE predictions in MPEAs. out losing the critical information. The flattened layer is used to convert
the resultant matrix into a single continuous linear vector. The flattened
output vector is then fed into the dense layer for further processing of
the data. The output layer is the last layer used to evaluate the perfor-
2. Methodology
mance of the model. The number and placement of each layer can vary
depending upon the model architecture.
2.1. Convolutional neural network
A CNN usually works with an input tensor of order 3 which contains
information about the height, and width of an image, and 3 color chan-
A CNN is a deep learning structure that takes images as input data,
nels (red, green, and blue), each representing pixel values correspond-
learns from the embedded patterns within the images and predicts an
ing to a specific color. Three channels represent the color intensity of
outcome. CNN has been widely used for both classification and regres-
a given pixel. Fig. 2 represents an abstract view of the input data and
sion problems. The ability of CNN to learn and recognize the underlying
various layers, where 𝑥𝑛 represents the input data passed through a layer
patterns in images is due to the types of layers in the architecture of the
𝑙𝑛 resulting in an output 𝑥𝑛+1 . This process is continued until the data is
model. In general, CNN can consist of convolutional, pooling, flatten,
processed through the last layer 𝑙𝐿−1 with an output 𝑥𝐿 . The last layer
fully connected, and output layers. A drop-out layer may also be added
𝑙𝐿 is the loss layer that determines the error with respect to the true
depending upon the performance of the model. A layer is the highest-
value of the input. Loss is calculated using Eq. (1) and is then used to
level building block that contains neurons. A neuron is a function which
update the model parameters. For example, the target value for a given
takes various inputs, performs mathematical operations, and gives an

2
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

Fig. 2. Representation of a mathematical operation that updates the variables. 𝑥1 , 𝑥2 , and so on represent the input while 𝑙1 , 𝑙2 and so on represents the weights.
𝑥𝐿 represents the output while z represents the loss.

Fig. 3. Schematic representation of the architecture of the CNN. Data is fed into model in the form of images.

input 𝑥1 is 𝑡1 but the predicted value by the model is 𝑥𝐿 . Thus, the loss model, the data is split into training, validation, and holdout/test. While
can be defined as 80% of data is used for training, the remaining data is divided equally
1( 1 )2 between validation and holdout. During training, early stopping crite-
𝑧= 𝑡 − 𝑥𝐿 (1)
2 ria is used to avoid overfitting. Mean squared error (mse) is monitored
The calculated loss function is then used as a guiding signal to update during the training process. Training is stopped when accuracy of the
the model parameters (i.e., weights and biases) for further improvement. model does not increase even after 50 epochs. Other hyperparameters
One of the ways to update the model parameters is via gradient descent used for training are learning rate = 0.0007 with Adam optimizer and
defined as: batch size = 8. Default padding option, which is padding = “valid” is
𝑡+1 𝑡 𝑑𝑧 used while developing the model. After training, the model is tested us-
(𝑙 𝑖 ) = (𝑙 𝑖 ) − 𝜂 (2) ing the holdout dataset. The holdout dataset is never shown to the model
𝑑 (𝑙𝑖 )𝑡
during the training phase. Root mean squared error (RMSE) and coeffi-
𝑡+1 𝑡
where, (𝑙𝑖 ) represents updated parameters from (𝑙𝑖 ) with calculated cient of determination (R2 ) are used for evaluating the performance of
gradient ( 𝑑𝑧𝑖 𝑡 ) using 𝜂 as the learning rate. The derivative 𝑑𝑧𝑖 𝑡 measure the model.
𝑑 (𝑙 ) 𝑑 (𝑙 )
𝑡
the rate of change of z with respect to the change in (𝑙𝑖 ) , known as
𝑡+1
the gradient. The value of (𝑙𝑖 ) is changed depending upon the loss 2.2. Stacking fault energy
function. The rate at which the parameters are updated is controlled by
the learning rate (𝜂). Higher learning rate might help to train the model The charge density images and the associated SFEs are calculated
faster but could result in an unstable training process whereas a slower using DFT. The calculations are performed using the Vienna ab-initio
learning rate results in slower training process and the model could be Simulation Package (VASP) [47,48]. The supercell consists of 108 atoms
stuck in a local minimum. An optimum learning rate is defined by the among 9 atomic layers in ABCABCABC stacking sequence in ⟨111⟩ di-
user after testing the model. rection as standard in the face-centered cubic (fcc) lattice. Periodic
Fig. 3 shows the model architecture used in this work. The model boundary conditions are applied in x and z directions, and a vacuum
is trained using the Keras library. A complete description of the model of 6 Å is applied in the y direction to avoid interaction between the
including the size of the kernel and parameters is given in Table 1. After top and bottom layers. The vacuum size convergence is verified. The
trials of optimization, image size of 196 × 73 is used ensuring suffi- ion-electron interaction is described by the full potential frozen-core
cient resolution quality. The image size strikes a balance between the projector-augmented wave (PAW) [48] and GGA-PBE [49] exchange-
visual quality and the time to load the images from the database. The correlation functional is used. The Brillouin zone is sampled using
images are fed after converting them into arrays. For example, convert- Monkhorst-Pack method with the mesh size of 6 × 6 × 1 k-points. After
ing eight images into arrays resulted in a dimension of the array given by performing the convergence tests, the energy cutoff of 350 eV is used
(8 × 196 × 73 × 3), where 8 represents the number of images, 196 × 73 for the plane-wave basis set with energy and force convergence set to
represents image size, and 3 represents the color channels. Since the 10−4 eV and 10−2 eV/Å, respectively. All the calculations are performed
images are used as raw input data, three channels (i.e., red, green, and using spin-polarized approximation. The Methfessel-Paxton technique
blue) are used to feed all information. In total, 137 images are used in with a smearing width of 0.1 eV is employed for the electronic relax-
this study. Feeding the multiple matrices of size 196 × 73 × 3 into a ation [53].
model do not pose a storage problem, as the single matrix of this di- The SFE is calculated using the slab-deformation approach in which
mension only occupies 0.32 MB of the storage space. Given the size and a stacking fault is generated by shearing the supercell along the [112̄ ] di-
capacity of today’s supercomputers, thousands of images can be easily rection. The structure of the supercell before and after shearing is shown
handled by the model. Furthermore, an alternate multiscale approach in Fig. 4. The displacement of atomic layers leads to a change in the
where, a filtered version of large matrices for the initial computation stacking sequence from ABCABCABC to ABCACABCA creating a stacking
and gradually increasing size with smaller patches, can also be used fault. The effect of the alloying element on the SFE is computed by sub-
for handling larger images. To ensure the absolute generalization of the stituting one atom of the host metal with an alloying atom in the stacking

3
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

Table 1
CNN model. First, second, and third columns represent layer type, output shape of the respective layer, and numbers of parameters,
respectively.

Layer (type) Output shape Parameters (#)

Input-1 (InputLayer) [(None, 73, 196, 3)] 0

Conv2d-1 (Conv2D) (None, 72, 195, 32) 416
Conv2d-2 (Conv2D) (None, 71, 194, 64) 8256
Average pooling-1 (AveragePooling) (None, 35, 97, 64) 0
Conv2d-3 (Conv2D) (None, 34, 96, 96) 24,672
Conv2d-4 (Conv2D) (None, 33, 95, 192) 73,920
Average pooling-2 (AveragePooling) (None, 16, 47, 192) 0
Conv2d-5 (Conv2D) (None, 15, 46, 192) 147,648
Dropout-1 (Dropout) (None, 15, 46, 192) 0
Conv2d-6(Conv2D) (None, 14, 45, 256) 196,864
Average pooling-3 (AveragePooling) (None, 7, 22, 256) 0
Conv2d-7 (Conv2D) (None, 6, 21, 256) 262,400
Flatten-1 (Flatten) (None, 32,256) 0
Dense-1 (Dense) (None, 32) 1,032,224
Dense-2 (Dense) (None, 1) 33
Total parametersTrainable parametersNon-trainable parameters 1746,4331746,4330

Fig. 4. Supercell used to calculate SFE. (a) perfect structure without vacuum, (b) perfect structure with vacuum of 6 Å, and (c) sheared structure with vacuum of
6 Å. Green and maroon spheres represent Ni host atoms and a dopant atom, respectively. Gray spheres in (c) represent atoms in the stacking fault layer.

fault following the standard method used in literature [50,51,54]. The 3. Results
initial structure is fully relaxed including atomic positions, cell shape,
and cell volume. The vacuum and stacking fault are then introduced fol- The SFE database is built in Ni binary alloys (one element + Ni),
lowed by DFT relaxation while constraining cell shape and cell volume. ternary alloys (two elements + Ni), and quaternary alloys (three ele-
Fully relaxing the initial structure (without vacuum) ensures capturing ments + Ni). The elements are from 3d (Sc to Zn), 4d (Y to Cd), and 5d
of volumetric effects induced by different size of the dopants and pre- (Hf to Hg) series. The database consists of 28 binary systems containing
vents any artifact or difficulty in relaxation. The energy from the perfect one dopant in the supercell, 56 binary systems containing two atoms
and sheared structures is used to calculate the intrinsic SFE using Eq. (3). of the same element in the supercell, 43 ternaries, and 10 quaternaries
leading to a total of 137 systems. Dopants in ternary and quaternary al-
Esheared − Eperfect
ESFE = (3) loys are placed randomly in the stacking fault plane. The charge density
XSC profiles are plotted for each of these systems and the corresponding SFE
where, Eperfect , Esheared are the energies for perfect (with vacuum) and is used as the target value.
sheared structures, respectively, and XSC is the cross-sectional area of Fig. 5a shows the progress of the model during the training phase.
the xz plane. The computed SFEs of pure Ni and Cu are 136.80 mJ/m2 Green and red lines represent the mean squared error (mse) with respect
and 32.23 mJ/m2 , respectively which are in good agreement with pre- to the number of epochs for training and validation data, respectively.
vious literature validating the methodology used for calculating SFE The mse for both training and validation data drops drastically during
[50,52,55,56]. the first few epochs. This is because the model is initialized with random

4
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

Fig. 5. (a) Change in mean squared error (mse) with respect to the number of epochs. Green and red lines represent mse for training and validation data, respectively.
(b) Correlation between true SFE calculated using DFT vs predicted SFE. (c) Bar graph showing the spread of the true data and the predicted SFE.

Table 2 number of p and d valence electrons, Zhao et al. [44]. used a linear re-
RMSE and R2 values for training, validation, and holdout data. gression model based on bond breaking and bond formation energy, and
RMSE (mJ/m2 ) R2 in our recent work, we used the nearest neighbor bond model as the de-
scriptor to describe SFE in alloys [11]. Recent work by Khan et al. [58].
Training data 2.19 0.98
used basic elemental properties such as metallic radius, melting point,
Validation data 4.23 0.88
Holdout data 6.18 0.87 electronegativity, the total number of electrons, etc. for predicting SFE.
This indicates that the same property could be fitted via different de-
scriptors resulting in a lack of consensus. This work shows that since
charge density is a reliable descriptor of SFE within DFT, its usage via
weights/variables during the training process. After about 100 epochs,
CNN potentially avoids any artifacts. Furthermore, while developing a
the model starts to converge. Log scale for the progress of the model
ML model for various other materials properties, it is observed that ‘text-
is also shown in Fig. S1. Green and red lines represent the log of mean
book’ descriptors, such as metallic radius of an element, valence elec-
squared error (mse) with respect to the number of epochs for training
trons, electronegativity, etc. are often used which could have limitations
and validation data, respectively.
in MPEAs. This is because recent studies have shown that the radius and
The trained model is used to predict SFE for the unseen data. Fig. 5b
the charge state of a given element in MPEAs can differ significantly de-
shows the correlation between the true and predicted SFEs for training,
pending on its nearest neighbor environment at a given lattice site due
validation, and holdout data. Respective RMSE and R2 for training, val-
to chemical randomness [59–63]. By using charge density extracted af-
idation and holdout data are shown in Table 2. RMSE and R2 values
ter DFT relaxation, the effects of nearest neighbor environments on the
of 6.18 mJ/m2 and 0.87 for holdout data, respectively imply that the
atomic and electronic structures are fully captured. In particular, the
model can recognize the patterns from the images and is able to predict
charge density images inherently contain bond length information and
SFE to a very good level of accuracy. Fig. 5c shows the bar representa-
the atomic and electronic distortions that are characteristic features of
tion of the true and the predicted data illustrating that the model can
MPEAs.
predict the range for the SFEs as well. In summary, the model learns
The second benefit of this approach is related to the computational
and recognizes the patterns of charge density images and predicts SFEs
expense. Because of the large phase space in MPEAs [34,64,65] DFT
of complex alloys.
calculations face an uphill battle in terms of computational expense. ML
To validate that the model is learning from the ‘area of interest’,
provides a means to overcome this challenge. Take an example of SFE
class activation mapping (CAM) is used. The CAM images are used to
computation for one supercell for a given alloy composition. The con-
highlight the areas of the image under model’s prime focus. Using CAM,
ventional approach to calculate SFE requires three different DFT calcu-
the output in the form of images is plotted after each convolutional layer.
lations. The first simulation consists of a full relaxation of the supercell
Fig. 6 shows the CAM images of two alloys, i.e., Ni-Sc and Ni-Ta in
(Fig. 4a), which could take ∼15 h for a 108-atom structure on 96 cores.
which one atom of Ni is substituted with Sc or Ta in a supercell of 108
In the second simulation, vacuum is added to the supercell (Fig. 4b)
atoms. The CAM images are shown for each convolutional layer. The
and the relaxation requires another ∼18 h. Finally, the third simulation
CAM images are unique to each dopant and clear differences between
is performed on the sheared structure that requires ∼45 h (Fig. 4c). In
Sc and Ta are observed for all convolutional layers. Since the ‘area of
contrast, the CNN model only requires the first calculation to generate
interest’ is in the vicinity of the dopant, which has a different charge
data and results in a significant reduction of computational time. Fur-
density than the rest of the Ni matrix, the contrast in the central region
thermore, since it is a common practice to calculate SFE of each plane in
of the images for Conv 7 layer illustrates that the model is focused on
the supercell to generate a larger sampling, such approach requires ex-
the region containing the dopant.
plicit calculation of each plane [44,66]. With this approach, the charge
densities of all planes can be obtained in one simulation thereby further
4. Discussion reducing the cost.
Finally, this image-recognition-based approach is generic and can
The ML efforts for SFE prediction interestingly reveal diverse sets of be applied to predict a variety of material properties that rely on charge
descriptors used by different research groups. For example, Shang et al. density. For example, properties related to point defects, defect segrega-
[57]. suggested a linear combination of SFEs of individual elements as tion, line defects, grain boundary, etc. can be predicted. This approach
a descriptor of SFE. Chong et al. [21]. suggested covalent radius and thus opens a new avenue to use ML for property prediction.

5
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

Fig. 6. Class activation mapping (CAM) of all convolutional layers for Ni-Sc and Ni-Ta systems. Contrast between the Ni atoms and the dopant region in the deeper
layers indicates the focus of the model on the ‘area of interest’.

5. Conclusions Acknowledgments

Image-recognition-based ML approach is used in a wide variety of This work is supported by the US Department of Energy, Office of Sci-
applications. Here, we employ this approach to predict materials prop- ence, Basic Energy Sciences, Mechanical Behavior and Radiation Effects
erty. We show that SFEs of concentrated alloys can be predicted by program. We acknowledge the support of computational resources from
learning only from the charge density images obtained from DFT cal- the Advanced Research Computing Center (ARCC) at the University of
culations. Consequently, a descriptor-less ML model is developed that Wyoming.
opens a route to bypass the need to identify descriptors which is often
a bottleneck to ML model development. A significant reduction in com-
putational expense is an additional outcome of the model. References

[1] Y. Liu, T. Zhao, W. Ju, S. Shi, Materials discovery and design using machine learning,
Data availability J. Materiomics 3 (2017) 159–177.
[2] K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for
molecular and materials science, Nature 559 (2018) 547–555.
All the relevant codes and data is available upon request and inquires [3] P. Ball, Using artificial intelligence to accelerate materials development, MRS Bull.
can be directed to the corresponding author. 44 (2019) 335–343.
[4] K.G. Reyes, B. Maruyama, The machine learning revolution in materials? MRS Bull.
44 (2019) 530–537.
Ethics declarations [5] J. Schmidt, M.R.G. Marques, S. Botti, M.A.L. Marques, Recent advances and appli-
cations of machine learning in solid-state materials science, npj Comput. Mater. 5
(2019) 1–36.
Competing Interests: The authors declare no competing interests. [6] N. Huber, S.R. Kalidindi, B. Klusemann, C.J. Cyron, Editorial: machine learning and
data mining in materials science, Front. Mater. 7 (2020) 1–3.
[7] J.E. Saal, A.O. Oliynyk, B. Meredig, Machine learning in materials discovery: con-
Supplementary materials firmed predictions and their underlying approaches, Annu. Rev. Mater. Res. 50
(2020) 49–69.
Please refer to supplementary section for all supplmental informa- [8] A. Chandrasekaran, et al., Solving the electronic structure problem with machine
learning, npj Comput. Mater. 5 (2019) 1–7.
tion. [9] K. Pal, C.W. Park, Y. Xia, J. Shen, C. Wolverton, Scale-invariant machine-learning
model accelerates the discovery of quaternary chalcogenides with ultralow lattice
thermal conductivity, npj Comput. Mater. 8 (2022) 1–12.
Declaration of Competing Interest [10] D. Xue, et al., Accelerated search for materials with targeted properties by adaptive
design, Nat. Commun. 7 (2016) 1–9.
All authors declare no conflict of competing interest. [11] G. Arora, D.S. Aidhy, Machine learning enabled prediction of stacking fault energies
in concentrated alloys, Metals 10 (2020) 1072.
[12] A. Manzoor, et al., Machine learning based methodology to predict point defect
energies in multi-principal element alloys, Front. Mater. 8 (2021) 129.
CRediT authorship contribution statement
[13] A. Manzoor, D.S. Aidhy, Predicting vibrational entropy of fcc solids uniquely from
bond chemistry using machine learning, Materialia 12 (2020) 100804.
Gaurav Arora: Methodology, Conceptualization, Writing – review [14] D. Jha, et al., Enhancing materials property prediction by leveraging computational
& editing. Serveh Kamrava: Methodology, Writing – review & edit- and experimental data using deep transfer learning, Nat. Commun. 10 (2019) 1–12.
[15] R. Ramprasad, R. Batra, G. Pilania, A. Mannodi-Kanakkithodi, C. Kim, Machine
ing. Pejman Tahmasebi: Methodology, Writing – review & editing. learning in materials informatics: recent applications and prospects, npj Comput.
Dilpuneet S. Aidhy: Methodology, Writing – review & editing. Mater. 3 (2017) 1–13.

6
G. Arora, S. Kamrava, P. Tahmasebi et al. Materialia 26 (2022) 101620

[16] D. Jha, et al., ElemNet: deep learning the chemistry of materials from only elemental [42] M. Shih, J. Miao, M. Mills, M. Ghazisaeidi, Stacking fault energy in concentrated
composition, Sci. Rep. 8 (2018) 1–13. alloys, Nat. Commun. 12 (2021) 1–10.
[17] J. Xiong, S.Q. Shi, T.Y. Zhang, Machine learning of phases and mechanical properties [43] N. Hashimoto, T. Fukushi, E. Wada, W.Y. Chen, Effect of stacking fault energy on
in complex concentrated alloys, J. Mater. Sci. Technol. 87 (2021) 133–142. damage microstructure in ion-irradiated CoCrFeNiMnx concentrated solid solution
[18] L. Himanen, et al., DScribe: library of descriptors for machine learning in materials alloys, J. Nucl. Mater. 545 (2021) 152642.
science, Comput. Phys. Commun. 247 (2020) 106949. [44] S. Zhao, Y. Osetsky, G.M. Stocks, Y. Zhang, Local-environment dependence of stack-
[19] S. Kamrava, P. Tahmasebi, M. Sahimi, Linking morphology of porous media to ing fault energies in concentrated solid-solution alloys, npj Comput. Mater. 5 (2019)
their macroscopic permeability by deep learning, Transp. Porous Media 131 (2020) 1–7.
427–448. [45] R. Chang, et al., Microstructure and mechanical properties of CoCrNi-Mo medium
[20] S. Kamrava, M. Sahimi, P. Tahmasebi, Quantifying accuracy of stochastic methods of entropy alloys: experiments and first-principle calculations, J. Mater. Sci. Technol.
reconstructing complex materials by deep learning, Phys. Rev. E 101 (2020) 043301. 62 (2021) 25–33.
[21] X. Chong, et al., Correlation analysis of materials properties by machine learning: [46] G. Arora, A. Manzoor, D.S. Aidhy, Charge-density based evaluation and prediction
illustrated with stacking fault energy from first-principles calculations in dilute fc- of stacking fault energies in Ni alloys from DFT and machine learning, Journal of
c-based alloys, J. Phys. Condens. Matter 33 (2021) 295702. Applied Physics (2022) Under review.
[22] R. Fang, et al., Accelerated discovery of metallic glasses through iteration of machine [47] G. Kresse, J. Furthmüller, Efficiency of ab-initio total energy calculations for met-
learning and high-throughput experiments, Sci. Adv. 4 (2022) eaaq1566. als and semiconductors using a plane-wave basis set, Comput. Mater. Sci. 6 (1996)
[23] S.R. Kalidindi, Feature engineering of material structure for AI-based materials 15–50.
knowledge systems, J. Appl. Phys. 128 (2020) 041103. [48] G. Kresse, D. Joubert, From ultrasoft pseudopotentials to the projector augment-
[24] M. Coşkun, A. Uçar, Ö. Yildirim, Y. Demir, Face recognition based on convolutional ed-wave method, Phys. Rev. B 59 (1999) 1758–1775.
neural network, in: Proceedings of the International Conference on Modern Electrical [49] J.P. Perdew, K. Burke, M. Ernzerhof, Generalized gradient approximation made sim-
and Energy Systems (MEES), 2017, pp. 376–379. ple, Phys. Rev. Lett. 77 (1996) 3865–3868.
[25] J. Kim, J. Kim, G.J. Jang, M. Lee, Fast learning method for convolutional neural net- [50] S.L. Shang, et al., Effects of alloying element and temperature on the stacking fault
works using extreme learning machine and its application to lane detection, Neural energies of dilute Ni-base superalloys, J. Phys. Condens. Matter 24 (2012) 505403.
Netw. 87 (2017) 109–121. [51] A. Goyal, et al., The influence of alloying on the stacking fault energy of
[26] L.D. Medus, M. Saban, J.V. Francés-Víllora, M. Bataller-Mompeán, A. Rosado– gold from density functional theory calculations, Comput. Mater. Sci. 188 (2021
Muñoz, Hyperspectral image classification using CNN: application to industrial food ) 110236.
packaging, Food Control 125 (2021) 107962. [52] S.L. Shang, et al., Temperature-dependent ideal strength and stacking fault energy
[27] X. Xie, et al., Deep convolutional neural network-based classification of cancer cells of fcc Ni: a first-principles study of shear deformation, J. Phys. Condens. Matter 24
on cytological pleural effusion images, Mod. Pathol. 35 (2022) 609–614. (2012) 155402.
[28] A. Ziletti, D. Kumar, M. Scheffler, L.M. Ghiringhelli, Insightful classification of crys- [53] M. Methfessel, A.T Paxton, High-precision sampling for Brillouin-zone integration
tal structures using deep learning, Nat. Commun. 9 (2018) 1–10. in metals, Phys. Rev. B 40 (1989) 3616–3621.
[29] A. Cecen, H. Dai, Y.C. Yabansu, S.R. Kalidindi, L. Song, Material structure-property [54] S.L. Shang, et al., Generalized stacking fault energy, ideal strength and twinnability
linkages using three-dimensional convolutional neural networks, Acta Mater. 146 of dilute Mg-based alloys: a first-principles study of shear deformation, Acta Mater.
(2018) 76–84. 67 (2014) 168–180.
[30] E.P. George, D. Raabe, R.O. Ritchie, High-entropy alloys, Nat. Rev. Mater. 4 (2019) [55] X.Z. Wu, R. Wang, S.F. Wang, Q.Y. Wei, Ab initio calculations of generalized-stack-
515–534. ing-fault energy surfaces and surface energies for FCC metals, Appl. Surf. Sci. 256
[31] G. Laplanche, et al., Reasons for the superior mechanical properties of medium-en- (2010) 6345–6349.
tropy CrCoNi compared to high-entropy CrMnFeCoNi, Acta Mater. 128 (2017) [56] W. Li, et al., Generalized stacking fault energies of alloys, J. Phys. Condens. Matter
292–303. 26 (1–12) (2014) 265005.
[32] J. Liu, et al., Deformation twinning behaviors of the low stacking fault energy high- [57] S. Shang, Y. Wang, Y. Du, M.A. Tschopp, Z.K. Liu, Integrating computa-
-entropy alloy: an in-situ TEM study, Scr. Mater. 137 (2017) 9–12. tional modeling and first-principles calculations to predict stacking fault en-
[33] D. Wei, et al., Novel Co-rich high entropy alloys with superior tensile properties, ergy of dilute multicomponent Ni-base alloys, Comput. Mater. Sci. 91 (2014
Mater. Res. Lett. 7 (2019) 82–88. ) 50–55.
[34] Y.F. Ye, Q. Wang, J. Lu, C.T. Liu, Y. Yang, High-entropy alloy: challenges and [58] T. Khan, et al., Towards stacking fault energy engineering in FCC high entropy alloys,
prospects, Mater. Today 19 (2016) 349–362. Acta Mater. 224 (2022) 117472.
[35] Z. Li, C.C. Tasan, H. Springer, B. Gault, D. Raabe, Interstitial atoms enable joint [59] N.L. Okamoto, K. Yuge, K. Tanaka, H. Inui, E.P. George, Atomic displacement in the
twinning and transformation induced plasticity in strong and ductile high-entropy CrMnFeCoNi high-entropy alloy - a scaling factor to predict solid solution strength-
alloys, Sci. Rep. 7 (2017) 1–7. ening, AIP Adv. 6 (2016) 125008.
[36] G. Frommeyer, U. Brüx, P. Neumann, Supra-ductile and high-strength mangane- [60] Y.F. Ye, et al., Atomic-scale distorted lattice in chemically disordered equimolar
se-TRIP/TWIP steels for high energy absorption purposes, ISIJ Int. 43 (2003) complex alloys, Acta Mater. 150 (2018) 182–194.
438–446. [61] H.S. Oh, et al., Engineering atomic-level complexity in high-entropy and complex
[37] D.T. Pierce, et al., The influence of manganese content on the stacking fault and concentrated alloys, Nat. Commun. 10 (2019) 1–8.
austenite/𝜀-martensite interfacial energies in Fe-Mn-(Al-Si) steels investigated by [62] S. Ishibashi, Y. Ikeda, F. Körmann, B. Grabowski, J. Neugebauer, Correlation analysis
experiment and theory, Acta Mater. 68 (2014) 238–253. of strongly fluctuating atomic volumes, charges, and stresses in body-centered cubic
[38] S.F. Liu, et al., Transformation-reinforced high-entropy alloys with superior me- refractory high-entropy alloys, Phys. Rev. Mater. 4 (2020) 1–15.
chanical properties via tailoring stacking fault energy, J. Alloy. Compd. 792 (2019) [63] R. Zhang, et al., Short-range order and its impact on the CrCoNi medium-entropy
444–455. alloy, Nature 581 (2020) 283–287.
[39] D.T. Pierce, J.A. Jiménez, J. Bentley, D. Raabe, J.E. Wittig, The influence of stacking [64] M.C. Gao, P.K. Liaw, J.W. Yeh, Y. Zhang, High-Entropy Alloys: Fundamentals and
fault energy on the microstructural and strain-hardening evolution of Fe-Mn-Al-Si Applications, Springer, 2016.
steels during tensile deformation, Acta Mater. 100 (2015) 178–190. [65] D.B. Miracle, High entropy alloys as a bold step forward in alloy development, Nat.
[40] S.F. Liu, et al., Stacking fault energy of face-centered-cubic high entropy alloys, In- Commun. 10 (2019) 1–3.
termetallics 93 (2018) 269–273. [66] S. Zhao, G.M. Stocks, Y. Zhang, Stacking fault energies of face-centered cubic con-
[41] P. Yu, et al., The influence of dilute aluminum and molybdenum on stacking fault and centrated solid solution alloys, Acta Mater. 134 (2017) 334–345.
twin formation in FeNiCoCr-based high entropy alloys based on density functional
theory, Sci. Rep. 9 (2019) 1–8.