Journal of Alloys and Compounds 976 (2024) 173144

Contents lists available at ScienceDirect

Journal of Alloys and Compounds

journal homepage: www.elsevier.com/locate/jalcom

Inverse design of high entropy alloys using a deep interpretable scheme for
materials attribution analysis
Chih-Yu Lee a, 1, Chia-Yung Jui a, 1, An-Chou Yeh b, c, Yao-Jen Chang b, c, Wen-Jay Lee a, *
a
National Center for High-Performance Computing, Taiwan, ROC
b
Department of Material Science and Engineering, National Tsing Hua University, Taiwan, ROC
c
High Entropy Materials Center, National Tsing Hua University, Taiwan, ROC

A R T I C L E I N F O A B S T R A C T

High entropy alloys (HEAs) have gained significant attention due to their unique combinations of properties and
Original content: DISMA_Research Dataset
extensive compositional space. However, designing HEAs with specific properties has become increasingly
(High-entropy alloys) (Original data)
complex and challenging. To address this issue, a Deep Interpretable Scheme for Material Attribution Analysis
(DISMAA) was developed to systematically uncover the composition-structure-property (CSP) relationships of
Keywords:
Generative model
HEAs and generate novel compositions. The approach combines a deep generative model, trained with limited
Attribution analysis data, to create a continuous composition space of AlCoCrFeMnNi high entropy alloys, and additional deep
Latent space learning models, trained with more data, to predict the corresponding phase and properties. Attribution and
High entropy alloy sensitivity analysis techniques are used to analyze the contribution of each element to different materials
Material design properties and phases. The distributions of composition, phase, mechanical property, and contribution of ele­
ments are visualized in the latent space, providing a novel and intuitive approach to analyze CSP relationships
using deep learning. The efficacy of the method was validated, demonstrating the interpretability and rationality
of the deep neural network model for material inverse design applications. This approach can be applied to tailor
multiple properties and manufacturing parameters in various materials systems, providing scientists with a
comprehensive context for designing HEAs.

1. Introduction Instead of forming intermetallic compounds, it was proved that HEAs

favor simple solid solution structures (FCC, BCC or mixed) rather than
The concept of High-entropy alloys (HEAs) was proposed by Yeh complex phases based on thermodynamic consideration [12–17]. In
et al. [1,2]. Different from conventional single principal metal based spite of the fact that a strong interest has been made on HEAs with single
alloys, HEAs were initially defined as alloys composed of at least five solid solution structure, the presence of two or more phases is also well
principal metallic elements, each ranging from 5 to 35 at% while in the expected and studied in order to obtain superior combination of me­
past decades many HEA alloy systems accommodated five to nine chanical properties [18]. Studies have shown various elemental contri­
metallic elements [3–5]. Among a great number of elements, Al, Co, Cr, butions in phase formation, i.e., Cu, Ni, Co, Fe and Mn promote the
Cu, Fe, Mn, Ni, V and Ti are by far the most commonly used for HEAs. formation of FCC phases (soft) in the synthesis of HEAs while Al, Cr, V
Within all the alloy families, CoCrFeNi based 3d transition metal alloys and Ti stabilize BCC phases (hard) [19–21]. In most Al-bearing HEA
have garnered most attentions, which can be considered as extensions of systems, attributed to its dualism, the effect of Al addition has been
stainless steels and superalloys e.g., AlxCoCrCuFeNi [6] and TixCoCr­ extensively investigated particularly on the phase transformation from
CuFeNi [7]. Due to the chemical complexity, these alloys possess unique BCC to FCC in many systems [12,22–24]. However, the single element
structures and properties unprecedented in conventional alloys [8–10]. substitution method used in the research limits the degree of freedom for
For instance, the most potential low-density structural materials for alloy design with respect to interpolation and extrapolation. Aside from
automobile, Al-Cr-Fe-Ni-X HEAs have been considered as the potential lacking completeness in alloy design, this approach additionally hinders
low-density structural material for transportation applications [11]. discussions on overall elemental interactions and

* Corresponding author.
E-mail address: [email protected] (W.-J. Lee).
1
Chih-Yu Lee and Chia-Yung Jui contributed equally to the present work and are co-first authors.

https://doi.org/10.1016/j.jallcom.2023.173144
Received 30 August 2023; Received in revised form 2 November 2023; Accepted 11 December 2023
Available online 13 December 2023
0925-8388/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

composition-microstructure-property relationship. new materials that can meet the target properties. With all that being
In order to overcome great amount of trial-and error experiments in said, VAE can help design reliable, synthesizable, and superior mate­
vast HEAs design space, efforts have been made in parametric and rials. Many other successful case studies were found to apply generative
computational approaches [23,25–33]. Although these approaches adversarial networks (GANs) as well, in addition to VAE in materials
accelerate conventional time-consuming alloy design via inverse design [65–68]. For examples, Lee et al. [39] presents a work
high-throughput screening, they also meet some challenges owing to showing that sample volume is augmented through conditional GAN to
oversimplification, high computational cost, insufficient reliability, and improve the accuracy of phase predictions and explain the design pa­
uncertainties. As data of HEAs generated from simulations and experi­ rameters involved in their deep learning network.
ments has become more and more accessible, machine learning (ML) is Since the HEAs comprising of five or more principal elements have
proposed to assist materials design. ML is able to automatically recog­ great combinations of properties, offering greater potential for diverse
nize patterns and features from high dimensional input data via statis­ applications than conventional alloys, the design of new HEAs and the
tical model and subsequently provide certain outputs, most common of associated synthesis recipe become more challenging, particularly with
which are prediction of phase formation, functional property and me­ a lack of understanding on the interaction of different elements. In this
chanical property [34–41]. Fundamentally, ML algorithm can be broken work, a new material design scheme using ML method is developed. In
down into data collection, data preprocessing, feature engineering, addition, the contribution with respect to hardness and each phase is
model selection, and model training. Generally, material data are scar­ investigated through attribution analysis in terms of each element. Mi­
city and costly. Training data can be sourced from various outlets, crostructures and mechanical properties of compositions are success­
including commercial databases, lectures, patterns, and laboratory ex­ fully visualized in a two-dimensional space, and hence the relationship
periments for ML. It’s common to encounter data uncertainty or in­ and contributions among all the elements are revealed and discreetly
consistencies. Hence, the presence of domain experts, responsible for discussed.
evaluating and refining high-quality data, is essential in the data pre­
processing. This ensures that the model, trained on high-quality data, 2. Method
can capture the correct correlations between data. Feature engineering
transforms compositions into physically related descriptors, aiding the Inverse material design using ML has been a promising approach to
model in extracting the connections between descriptors and output discovering new materials. VAE [69] can be used to create a continuous
properties. As phase classification and property regression problems latent space for generating unprecedented composition with continuous
[42–47] involving ML have been well developed, the significance of properties. In order to establish an intuitive design scheme for new
thermodynamic, atomistic, physical and chemical features transforming materials, VAE jointly trained with multi-layer perceptron (MLP) neural
from the compositional input data have also been reported and dis­ network property predictors were applied to establish a variant model of
cussed [42,48–54]. Model selection depends on the objectives of study, VAE(VVAE) as shown in (I) of Fig. 1. Two MLP predictors are trained
whether it involves regression, classification, clustering, or similarity independently and used to predict the property and phase on the
detection in materials. For a more detail discussion, our previous review decoded pseudo compositions given by the VVAE in the following. The
article has offered an in-depth overview of machine learning applica­ prediction can be mapped onto the compositional space, allowing for
tions in the realm of high-entropy materials [5]. visualization and analysis of the relationship between composition,
In the context of material design through machine learning, mate­ property, and phase. VVAE was implemented using Keras package [70],
rials scientists are actively engaged in the pursuit of novel HEAs [55,56]. whereas MLP network was implemented using scikit-learn package [71].
Some researchers still attempt to utilize the forward models for To explore the impact of each elemental component in HEAs, integrated
high-throughput predictions to identify desired materials [57,58] while gradients method implemented in DeepLIFT [72] was used for attribu­
others are focused on optimization strategy. For instance, Rickman et al. tion analysis as shown in (III) of Fig. 1, which helps to identify which
[59,60] team made endeavors in developing advanced data analytical elements are critical for predicting a particular property or phase.
strategies in combination with natured-inspired algorithms, e.g., genetic Autoencoders (AE), a type of generative model, consists of two deep
algorithms and cuckoo search, and finally HEA candidates with high neural networks: an encoder (z = f(x)) and a decoder (x = g(z)). The
hardness were found. Some studies have not only concentrated on encoder maps input features x into a vector z in a lower-dimensional
searching for novel alloys, but also uncovered the underlying property, space known as the latent space, and the decoder maps the latent vec­
structure, performance, processing (PSPP) relationship and the contri­ tor z back to the original representation. The goal is to reconstruct the
bution of each element in multi-component system using ML. Qiao et al. input data as accurately as possible. During the process of compression
[4] found that Cr and Al are sensitive to mechanical performance, and and decompression, AE is expected to learn the implicit information
meanwhile FeCrNiAl0.8 exhibiting high fracture strength and plastic from data by distilling and expanding the information in the latent
strain was predicted based on solidification characteristics; Wu et al. space, which leads the latent space to generate a continuous distribution
[61] employed neural networks to reveal that Al is a critical element for of composition as it is decoded. However, in the case of AE without
eutectic high entropy alloy (EHEA), while Cr is an element that is regularization, latent space could overfit or produce uninterpretable
strongly associated with Al in Al-Co-Cr-Fe-Ni alloy system. However, contents. The corresponding material properties is not guaranteed to be
these studies can only generate and analyze small amount of data, a continuous distribution on the latent space. VAE, an extension of AE,
lacking systematic and comprehensive analysis in a wider compositional introduces probabilistic distributions to the encoded variables (latent
range. vectors). In principle, VAE supposes that the posterior distribution of the
Instead of using the forward perdition model to screen new compo­ latent space p(z|x) possesses a known prior probability distribution, such
sition, deep-learning generative models (GMs) are proved to be adopted as Gaussian distribution p(z)). Technically, this is achieved by mini­
to material inverse design and investigation, generating materials with mizing the Kullbach-Leibler (KL) distance between encoder p(z|x) and the
desired property and structure [62]. Noh et al. [63] demonstrated in­ prior distribution p(z). In other words, VAEs achieve better generality by
verse design of solid-state materials by using variational autoencoders constraining the latent space to follow normal distributions and pro­
(VAE). Gomez-Bombarelli et al. [64] also presented the usage of VAE in moting better interpretability for each data point in the latent space. As a
generation of drug-like molecules. Note that the advantage of VAE is the result, for a well-trained VAE, the latent representation of a data point
latent space of which is represented as continuous and differentiable must consist of the critical information that allows the decoder to ach­
vectors. By jointly training on target properties, VAE allows data with ieve high-quality reconstruction. Thus, the latent space, the condensed
similar properties to spatially cluster on the latent space. Based on this abstract continuous representation of all given data points, is expected
characteristic, gradient-based optimization is able to be adopted to find to reveal the known or even undiscovered relationships between

2
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

Fig. 1. Schematic framework of DISMAA. The method consists of three main steps, as follows. Step (I): Starting with training of VVAE to generate the latent space of
specific HEAs. The decoder network produces the corresponding compositional representations from the latent vectors, serving as composition generator. Step (II):
Independent MLP models (Pproperty and Pphase ) are trained for the prediction of material property and phases using the selected dataset. Once trained, the MLP models
take the pseudo compositions generated from VVAE as inputs and output the corresponding properties and phases of the HEAs. The predictions results are mapped
onto the latent space, allowing for visualization and analysis of the relationship between composition, property, and phase. Step (III): The attribution analysis is
performed to explore the importance of each feature (i.e., element) in the input to the output prediction of MLP models.

existing data points in a continuous manner. For example, neighboring to 6-dimensional vectors representing elemental fractions of HEAs.
data points are supposed to possess similar representation in the latent To study the correlation between material composition, properties,
space. On the contrary, the points away from each other often indicate and phases, an MLP predictor is jointly trained with the VAE model,
that they differ from each other in some or all aspects. Moreover, ac­ which we named VVAE in this work. Therefore, VVAE is used to
cording to the observed characteristics of the latent space, we can generate the material compositions (x(z)) along with their correspond­
generate and explore pseudo data points which are not existing currently ing property (y(z), in our case, Vicker’s hardness). The MLP with 2
with specific intentions by utilizing the decoder as a generative model. hidden layers consists of [2,3] neurons. Tanh is used as activation
VAE was employed to train the latent space of AlCoCrFeMnNi HEAs functions for all neurons, with which the latent space recognizes itself
and generate a distribution of pseudo-material compositions. The conferring the physical meaning and thus materials with similar prop­
elemental fractions of 6 compositional vectors were considered as the erty assemble [73]. In other words, it can generate similar pseudo ma­
inputs and outputs. In order to improve the performance of model, terials by sampling the vicinity of given materials. Conversely, with
feature transformation was made, which transformed the compositional increasing distance of latent coordinates from the origin, the generated
inputs into a list of quantitative features based on relevant physical materials will be more distinctive. In addition, the relationship between
properties, i.e., hardness and density of solid, weighted by the concen­ the concentration distribution of each element and the property distri­
tration of each element in compounds. As a result, 6 elements with 2 bution can be intuitionally visualized on the continuous latent space.
physical quantities are then transformed into 12 physical quantity fea­ The loss function of the VVAE is composed of three parts: recon­
tures. Combined with 6 element fractions, a total of 18 features were set struction loss Lreconst , latent loss LKL , and property prediction loss Lprop ,
as input in the model. Both the encoder and decoder consist of MLPs which can be shown as following equation:
with two layers - the number of neurons of which are [4,5] and [4,5]
Loss = Lreconst (x, x) + α ∗ Lprop + λ ∗ LKL + β ∗ RegL2 (1)
respectively. Tanh is used as activation function for all hidden layers
except for decoder’s output layer which uses a softmax function. The
where Lreconst is defined as the difference between the input composi­
softmax function enforces the sum of every decoded result (elemental
tions x and output compositions x in the VVAE, which measured by
fractions in our case) to be exactly 1.0, satisfying the definition of a valid
cross-entropy loss. Lprop is the difference between the real and predicted
composition. The latent vectors are chosen to be two, which enables a
property in the jointly trained MLP network, which is evaluated by mean
continuous two-dimensional distribution of compositions to be observed
square errors. LKL compares the probability distribution difference be­
in the latent space. Ultimately, the decoder can provide a continuous
tween q(z|x) and p(z), which is calculated by KL divergence. α and λ are
mapping that projects final results from the 2-dimensional latent space

3
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

the parameters for balancing the reconstruction capability, regularity of

latent space, and preciseness of property prediction, which are set to 10 1 ∑M
P= Pj (2)
and 0.5. RegL2 is additional L2 regularization term to constrain the M j=1
margin imposed on the weights of all three MLP architectures for con­
The uncertainty of the prediction is the standard deviation over the
tracting generalization errors. β is set to 0.0001. Adam was utilized as
committee models.
the optimizer during training process with batch size of 5 for 1500
Since the training data of VAE are experimental records obtained
epochs. Due to the probabilistic nature of VAEs, additional Gaussian
from as-cast manufacturing process, we follow the same discipline in
noise is added during the encoding process. Each existing composition is
selecting training data for independent phase predictor. Instead of using
sampled 1000 times, and the mean of these encoded latent vectors is
a large dataset with more compositions to train the phase predictor in
used to represent the corresponding composition. The latent represen­
previous studies, we tend to use a limited dataset and elements for the
tations are then decoded, and we expect the reconstructed compositions
purpose of establishing a strategy to understand the composition-
should be as similar as the original ones.
structure-property relationship. According to previous literature [61,
For the training dataset of VAE model, only 31 data of HEAs with the
64,75], most of AlCoCrFeMnNi systems consist of three type of micro­
6 elements (Al, Co, Cr, Fe, Mn, Ni) are taken into account, which is
structures: disordered FCC (denoted as FCC), disordered BCC (denoted
obtained from vacuum arc melting process to ensure chemical homo­
as BCC) and ordered BCC (denoted as B2). Therefore, we deliberately
geneity. The number of n-nary alloys of dataset are listed in Fig. 2. The
select HEAs whose phases are combinations among them in our training
hardness of training data set ranges from 100Hv to 650Hv, in which 3
data. We note that in spite of the fact that the data collected from lit­
new compositions with hardness over 600Hv were discovered and
eratures are obtained from real experiments, they may still not be well
validated by author’s previous work using a ML based optimization
characterized, because microstructures of HEA in the early stage of
approach [74]. After a well-trained model is obtained, the latent space
development are not completely studied in previous literatures [76–78].
will be represented as continuously differentiable vectors residing on a
Particularly, distinguishing between disordered BCC and ordered BCC
probabilistic manifold. The range of latent space was set between − 4–4
structures was challenging in the 3d transitional HEA system due to
in 2 dimensions, and the corresponding area was partitioned into
limited characterization equipment available at the time. However, we
200 × 200 grid points uniformly for both axes, which indicates that 40,
understand that B2 crystal structure is more than critical in this alloy
000 AlCoCrFeMnNi pseudo compositions can be decoded from latent
system, and a rigorous classification among FCC, BCC and B2 is neces­
space.
sary. After data collecting and cleaning, only 94 data with Al, Co Cr, Cu,
Owing to the small training dataset and additional Gaussian noise
Fe, Mn, Mo, Ni elements remain, 18 of which are newly-released and
added in the encoding process, the accuracy in prediction and the usage
believed to be well characterized [79–83]. In order to validate the ac­
of attribution analysis are constrained. We therefore ultimately make
curacy of our model, 76 training data are used to train and establish
the jointly trained hardness predictor only used for recognizing the
Model 1 and 18 latest data points are used for testing. After validation,
orientation of latent space during training. On the other hand, an in­
the total amount of data is used to train Model 2 for further application.
dependent MLP network with hidden layer structure of [2,3] is trained
Fig. 3 shows the histogram of phase configuration of n-nary HEAs for
with 95 data, where two extra elements Cu and Mo are also included, to
training dataset in Model1 and Model2, where [FCC] indicates the single
practically predict the hardness of HEAs. Without involving complicated
FCC phase, [BCC, B2] indicates the dual phases in a HEA.
physical representations by feature transformation, e.g., ΔSmix, ΔHmix,
The input features for training are the fractions of the 8 elements.
VEC, and Δχ , 8 compositional fractions are considered as the inputs, i.e.,
The model is trained with a MLP multiclass classifier of two hidden
Al, Co Cr, Cu, Fe, Mn, Mo, Ni elemental fractions. The predictor is
layers of [3,5] neurons. The activation function for hidden layers is tanh.
trained by a lbfgs solver with L2 regularization weight of 0.7. To elevate
The solver used for weight optimization is lbfgs. We note that our test
the stability of these models for more reliable predictions, bootstrap
results show that using the compositional fraction as features for model
sampling was employed [74]—a committee size M consisting of 512 NN
training reveals a similar predicted accuracy with using physical fea­
models, labeled j ∈ {1…..M}, was built to provide solutions according to
tures in such a small dataset. The outputs are defined as the probabilities
different randomly sampled training data. The final prediction result P
of FCC, BCC, and B2 phases. Hence, the sigmoid activation function is
was computed as the average over the committee as following:
used in output layer, where the function takes any real value as input
and outputs values in the range 0–1. A value of 0.5 is defined to be the
threshold of classification model. If generate novel HEA compositions

Fig. 2. Number histogram of n-nary alloys in training data for different models: VAE, independent hardness predictor and independent phase predictor.

4
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

Fig. 3. Sample distribution of (a) Phase configurations used in Model1. (b) Phase configuration in dataset 2 used in Model2.

and the predicted outcome ≥0.5, the phase exists. A committee of 512
models are trained as bootstrap sampling. Fig. 3 compares the training 3. Results and analysis
accuracy of Model1 and Model2 for different phase predictions, which
shows that Model1 already has great capability in prediction (0.85 In this work, the latent space trained by the VVAE were introduced to
testing accuracy), while Model2 is greater which will be chosen for final systematically visualize compositions, structures and properties distri­
use. Supplementary Figure 1 shows the training accuracies of each phase butions and to study their correlation. Fig. 4 demonstrates the hardness
in two different models. Finally, test results show that using the distribution of AlCoCrFeMnNi on latent space, where the 2-D latent
compositional fraction as features for model training reveals a better vectors are decoded into 40,000 AlCoCrFeMnNi pseudo compositions
predicted result than that using quantitative features in such a small under 200 × 200([− 4,4] x [− 4,4]) grid points. The corresponding
dataset. hardness shown in Fig. 4(a) and (b) are calculated by jointly trained
Integrated gradients method [84] is adopted to analyze the attribu­ predictor and independent predictors, respectively. It is obvious that the
tion of each element concentration to hardness and phases. The attri­ property landscapes obtained by both predictors are continuous. The
bution is calculated for data points with respect to a baseline point, so hardness distribution of the pseudo compositions is approximately in the
the result is a relative relationship according to a specific model. The range of 100–700Hv for both models. Note that the independent pre­
arguments mentioned above convey two highlights: (1) the attribution dictor trained with more dataset is employed to improve the perdition
analysis is a relative measurement estimating the effects caused by the accuracy. Supplementary Fig. 2 (a) and (b) compares the relationship
change between inputs. (2) The attribution analysis can work if and only between ML prediction and true values from jointly trained predictor
if the predictive model catches the problem behavior correctly. and independent predictor, respectively. Compared with the jointly
Furthermore, the condition of completeness for attribution analysis is trained predictor, the independently predictor has higher R2, similar
possessed by the integrated gradients method which guarantees the sum Pearson correlation scores, and lower MAE value while the jointly
of all the attributions would be equal to the difference of the function trained predictor tends to underestimate especially in high-hardness
values of any two points to be analyzed. The formula of integrated region. Hence, we choose the independent predictor for further appli­
gradients method is as following: cation in this work. Since the latent space of VAE follows gaussian dis­
∫1 ⃒ tribution while training, the training data/compositions are encoded
∂fc (̃x) ⃒⃒ and appear at around the center of latent space as shown by circle
Rci (x) = (xi − x) • dα (3)
α=0 ∂x ̃i ⃒̃x=x+α(x− x) symbols in the latent space. The colors of the circles indicate their real
hardness values. In the other words, the closer to the center of latent
and the definition of completeness is: space is, the more reliable the latent vector will be, which can be
∑N observed in Fig. 4(c). Consequently, the center region of latent space
Rc (x) = fc (x) − fc (x)
i=1 i
(4) shows a lower uncertainty as shown in Fig. 4(c). Likewise, the prediction
models for compressive strength and elongation can also be trained and
where xi is input composition, x is the referenced composition (baseline visualized on the latent space as shown in Supplementary Fig. 3 and
point), fc is the outputs, Rci is the attribution, c indicates dimension of Fig. 4, respectively.
outputs, and i indicates the dimension of inputs. Fig. 5(a) shows the atomic fractions for different HEAs components
Sensitivity analysis [84] provides an effective amount of the varia­ generated by the VVAE. As a matter of fact, the corresponding Cr% and
tion magnitude of the target output at a perturbation of input variables, Al% exceed the standard range of the conventional definition of HEA­
which was conducted to evaluate the rate of change in property pre­ —each element ranging from 5% to 35%, as shown in the upper region
diction at different fraction of HEAs components. Taking the absolute for Cr and the right-bottom regions for Al, respectively. Given only two
value of the partial derivative of the target output fc (hardness and phase data in the whole training data whose Al concentration exceeds the
in this work) with respect to the inputs xi (fraction of components in this conventional definition of HEA, the hardness distribution in high Al or
work), the formula is expressed as Cr regions may have higher uncertainty related to the others, which are
⃒ ⃒
⃒∂fc (x) ⃒ corresponding to the results of Fig. 4(c), where the errors in the edge
Sci (x) = ⃒⃒ ⃒ (5) region are much higher than that in the central region. In addition, the
∂xi ⃒
magnitude of the change in elemental concentration are significant

5
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

Fig. 4. Contour map of hardness for AlCoCrFeMnNi HEAs according to (a) jointly-trained predictor and (b) independent predictor (c) shows the corresponding
uncertainty of contour (b). The circles shown in contour (a) present the decoded data from training data, where the colors are their real value of hardness.

Fig. 5. (a) Elemental concentration (at%) (b) contribution distribution, and (c) sensitivity distribution of each element on the latent space of AlCoCrFeMnNi system.
The cross symbol at (0,0) on (b) and (c) indicates the baseline for attribution analysis.

different, i.e. Al and Cr have variations higher than 55%, whereas the Al% on the hardness becomes weaker. The variation leads the single FCC
variation of Mn is only 17%. The reason can be attributed to the dis­ phase transform to a BCC+B2 phase, which will be discussed in next
tribution ranges of components in the training dataset, which directly section. Together with hardness contour as shown in Fig. 4(b), Supple­
affects the result of concentration distributions in the latent space. mentary Fig. 5 (a) shows the traces for observing the changes in material
Through attribution analysis, the contribution of elements to hard­ composition, hardness, and phase in the latent space. It is found that the
ness can be calculated and subsequently visualized in the latent space hardness increases with Al% up to 0.2, while it gradually saturates asAl
via contour map as shown in Fig. 5(b). Comparing the results in Fig. 5(a) % higher than 0.2. The transition behavior can be apparently observed
and (b), it can be found that concentrations of Al and Cr are proportional by the calculation of sensitivity, which is obtained by calculating the
to hardness while the rest elements (Co, Fe, Mn and Ni) have negative gradients of hardness attribution with respect to each element as shown
correlation to hardness in this latent space. The Pearson correlation in Fig. 5 (c). Compare to other elements, both Al% and Ni% are much
coefficient are calculated according to the distributions of fraction and sensitive for hardness in specific ranges. Especially for Al, the gradient
the hardness contribution of element. The corresponding results are increases drastically at Al% between 0.1 and 0.2 as shown in around the
− 0.716, 0.861, − 0.898, − 0.956, − 0.637, and 0.777 for Co, Cr, Fe, Ni, red region of Fig. 5.(c). Ni’s sensitivity drastically changes in the partial
Mn, and Al, respectively. It shows that the hardness contribution is region of Ni% > 0.2. Obviously, the combination of the contribution of
highly associated with the variation of Cr, Fe, and Ni contents, while Mn Ni and Al creates a transition region that separates the hardness contour
and Co are relatively weakly associated. To be more specific, the into a strong and a weak hardness region at the latent space in Fig. 4(b).
contribution of Ni is strongly and inversely influenced by content of Ni Note that the change is caused by the phase transformation as it will be
as it is higher than 0.2. The hardness contribution of Al is significant at discussed in the following Discussion section.
the content region around 0.1–0.2, leading a high correlation between Using MLP classification model with dataset 2 to train the phase-
Al% and hardness locally. However, apart from this region, the effect of prediction model, the accuracies for each phase achieves 0.98, 0.93,

6
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

and 0.98 for FCC, BCC and B2 phases as can be seen in Fig. S1, respec­ respectively, while the BCC and B2 structures are inversely proportional
tively. The probability of each phase of the AlCoCrFeMnNi systems are to Ni% but proportional to Al%. For the rest elements compared with Ni
predicted and shown in the latent space of Fig. 6(a). Thereby, a phase of and Al, the concentration differences have no obvious effect on the
the system is considered as existing if its probability is > 0.5. Otherwise, phase configuration in AlCoCrFeMnNi system.
it is assumed that this phase is absent. As the three phase distribution
maps overlap, five phase combinations can be visualized in different 4. Discussion
colors on latent space as shown in Fig. 6(b). The phases included in the
distribution map are FCC, FCC+BCC, FCC+B2, FCC+BCC+B2, and Many previous literatures have successfully explored the novel alloy
BCC+B2 phases, the order of which corresponds to the variation of with desired properties via optimization methods, e.g., annealing pro­
hardness from low to high. The corresponding composition-hardness- cess [74], genetic algorithm [39], [55,56,85,86] and ML surrogate
phase relationship can be observed in Fig. 4(S). As the hardness model with experimental algorithms [87]. Compared with the schemes
variant from a low hardness to a high hardness, the phase structure that considerably screen the composition space using the optimization
could transfer from FCC to FCC+BCC to FCC+BCC+B2 to BCC+B2 as methods mentioned above, VVAE, taking advantage of visualization of
can be seen in paths p1 to p0 and p1 to p2 in Fig. 4(S)(a) and from FCC to the material properties and composition on a continuous latent space,
FCC+B2 to FCC+BCC+B2 to BCC+B2 in path p1 to p3 in Fig. 4(S)(b). allows us to intuitively explore multiple relationships among composi­
Analogous to the discussion above, BCC based structures dominates tions, properties, and phase configuration of the complicated HEAs. In
when either Al% or Cr% is high, while FCC are stabilized when content addition, the latent space is capable of performing in a single-shot
of Co, Fe and Ni increases. We note that there are no compositions fashion to generate the candidate materials, which leads the way for
classified into single BCC group, which is confirmed to be rare in novel material and application design. These new candidate materials
AlCoCrFeMnNi system. In addition, FCC+ BCC only occupied a small can be made and measured by experiments for validation, and then
region in the latent space. According to phase diagram reported [24], augment the training database for the next iteration for improving the
CoCrFeMnNiAlx system has small region of FCC+BCC and BCC struc­ reliability of model and enhancing the performance of materials. In this
ture. Thereby, the predictor is considered fairly reasonable and work, we propose the scheme to study the multi-correlation of a specific
acceptable based on the combinations of phases. The phases predicted HEAs (AlCoCrFeMnNi) system. However, excessive input dimension
by the phase classifier are compared with that by thermodynamic model (elements) can be detrimental to VAE model as it is likely to become too
CALPHAD (Thermo-Calc 2020b, TCHEA4). Several unseen compositions complicated. Supplementary Fig. 6 illustrates the composition distri­
were decoded from the latent space and fed into the CALPHAD for butions of AlCoCrCuFeMnMoNi system in the latent space, showing the
comparison. The results of CALPHAD are also plotted in Fig. 6(b). It can complex variations in compositional space. Instead, the VVAE trained
be seen that a considerable discrepancy is shown between CALPHAD and with fewer datasets qualitatively creates a latent space that mainly focus
machine learning. In Discussion section, the ML results are validated and on a specific HEA system. This method is particularly advantageous in
compared with the CALPHAD results using 18 HEA experimental data the situation of lacking data, which is a very common issue in materials
from the latest literatures. science. Independent predictors trained with more datasets and ele­
Fig. 7(a)-(d) shows the element concentrations and their contribu­ ments can be used to quantitively predict properties and analyze the
tion to the FCC, BCC and B2 phases, respectively. Since the base point on results. Combining the advantages of both VAE model and MLP pre­
the latent space is a BCC+B2 structure, the values shown in Fig. 7(b)-(d) dictor, the proposed scheme makes it easier to study the physical cor­
represent the relative contribution from the base point. Zero values relation of HEAs in a latent space generated by sparse experimental data.
shown in Fig. 7(b)-(d) indicate that the phase of FCC, BCC, and B2 has Analogous to the work reported by Lee et al. [39], which augments data
equal contribution to the base point. In Fig. 7, it is clearly found that volume through conditional GAN to improve the accuracy of phase
variation of Al and Ni concentration play the major roles to influence the perdition and explain the design parameters involved in their deep
phase transformation of AlCoCrFeMnNi system, where the FCC phase learning network, we utilize VAE to reproduce data and to interpret the
gradually increases with increasing Ni% and decreasing Al%, physical information conveyed by the latent space. In this study, a more

Fig. 6. (a) The upper subplots are the probability of existence, and the lower subplots are classifications, existence if probability > 0.5; absence if probability ≤ 0.5.
(b) Overall phase configuration with five different combinations predicted by independent MLP phase predictor. Several compositions are selected to predict their
phases via CALPHAD, and the results of which are labeled with different marks. Red hexagon means FCC, blue cross means BCC and green circle means B2.
Overlapping marks indicate the presence of multiple phases.

7
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

Fig. 7. Subplots in (a) are elemental concentration(at%) of each element. (b), (c) and(d) are attribution distribution in latent space for FCC, BCC and B2 respectively.

rapid but comprehensive analysis can be given out through the sche­ elevated its capability in multicomponent system prediction, it still does
matic framework presented and close the loop of alloy design [88,89]. not ensure the validity of its predictions for complex structure of HEAs
In order to validate the accuracy of our model and understand the [5], due to the lack of available thermodynamic and kinetic data for
difference between experiment and theoretical prediction, Table 1. ternary and more complex systems in the database. Therefore, the reli­
compares the results of 18 testing data among literature, ML prediction ability of HEAs prediction from CALPHAD should be further studied.
(Model 1 and Model 2), and CALPHAD calculation. It is clear that Model Therefore, we suggest that ML-based predictors are more suitable for
1 has a relatively higher score of recall, which means the true labels are non-equilibrium HEAs as the training data are basically collected from
all predicted true correctly; Model 2 has higher score of precision as the experiments.
true predictions are all actually true. In summary, both Model 1 and Regarding the combinations of phases predicted in Fig. 6(b), it is
Model 2 have higher accuracy (0.88) than CALPHAD (0.55). The reason observed that the transforms from FCC to FCC+BCC to FCC+BCC+B2 to
can attribute to the fact that the calculated results of CALPHAD are based BCC+B2 and from FCC to FCC+B2 to FCC+BCC+B2 to BCC+B2 in
on the thermodynamic equilibrium state, while the experimental data different path. Pure disordered BCC (BCC) and ordered BCC (B2) do not
are often in a non-equilibrium as-cast condition. Although CALPHAD has exist. It is reasonable since Al atom has the largest radius than the other

Table 1
The comparison of phases predictions between real phases discovered from literature and 1. Model 1 (trained with 76 data) 2. Model2 (trained with 94 data) 3.
CALPHAD. The composition chosen to be tested here are more lately and under precise characterization. Compositions highlighted in bold font and denoted with * are
incorrect phase prediction.
Phases (Literature) Phases (Model1) Phases (Model2) Phases (CALPHAD)

Al0.7CoCrFeNi[92] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 *FCCþB2

Al0.9CoCrFeNi[92] BCC+B2 BCC+B2 BCC+B2 *FCCþBCCþB2
Al1.2CoCrFeNi[92] BCC+B2 BCC+B2 BCC+B2 BCC+B2
Al1.8CoCrFeNi[92] BCC+B2 BCC+B2 BCC+B2 BCC+B2
Al2.75CoCrFeNi[93] BCC+B2 BCC+B2 BCC+B2 BCC+B2
AlCoCr2FeNi[94] BCC+B2 BCC+B2 BCC+B2 BCC+B2
Al0.7CoCr2FeNi[95] FCC+B2 *FCCþBCCþB2 *BCCþB2 *FCCþBCCþB2
AlCoCrFeNi2.1[11] FCC+BCC+B2 FCC+BCC+B2 *FCCþB2 *FCCþB2
AlCoCr1.2FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 *FCCþB2
AlCoCr1.4FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 *FCCþB2
AlCoCr1.6FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 *FCCþB2
AlCoCr1.8FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2
AlCoCr2FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2
AlCoCr2.2FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2
AlCoCr2.4FeNi2[96] FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2 FCC+BCC+B2
Al1.3CoCrCuFeNi2 FCC+B2 *FCCþBCCþB2 FCC+B2 *FCCþBCCþB2
AlCoCrFe1.5Ni[97] BCC+B2 BCC+B2 BCC+B2 BCC+B2
AlCoCrFe2.5Ni[97] BCC+B2 BCC+B2 BCC+B2 BCC+B2

8
 C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

transition metals in the AlCoCrFeMnNi system. Therefore, a higher correlation with hardness of this alloy system as shown in Fig. 5. Since
content of Al dissolved into FeCoNiCrMn is expected to increase the the smaller atomic size of Fe reduces the atomic size differences and
lattice strain and distort the lattice structure, leading to solid solution results less lattice distortion, hardness declines as Fe’s concentration
strengthening. In order to relax the lattice distortion energy, as Al% increases. In AlCoCrFexNi system, Chen et al. [104] revealed that Fe can
increases over a critical value, the unstable close-packed structure of reduces Cr and Ni segregation. Further, it restrains the formation of
FCC transforms into loose-packed BCC structure gradually. This obser­ Cr3Ni2 compounds, resulting a lower hardness. From attribution anal­
vation has been in fact validated by He et al. [90]. They find that the ysis, concentrations of Co and Mn are less correlated to hardness and
duplex phases of the BCC+ FCC region is located at regions, where Al% phases, which is analogous to previous studies. Co was proved to
is between 8~16% in the (FeCoNiCrMn)100-xAlx system [75,91]. The decrease strength and stabilize FCC attributed to higher VEC [105]. Mn
detailed microstructure of B2 in BCC phase increase with increasing Al addition was seen to hardly make impact on phase transformation and
%, revealing the ordering degree is further increased. FCC is limitedly hardness [22] in this work. It is noted that since the construction of
found only on the boundaries of the BCC grains at 16% addition of Al. In latent space is based on the training data, which the concentration dis­
addition, the hardness is saturated to a constant ≈ 538HV as the Al% tribution of the latent space could be limited. However, we can take
> 16% in a (FeCoNiCrMn)100− xAlx alloys system [75]. Our previous ML advantage of the latent space to reveal the complex relationship in this
study validated by experiments also found that a high hardness of range.
AlCoCrFeMnNi system is strongly related to both high Al%(35%) and Cr The reason why hardness distribution in our latent space is variable
%(35%) [74]. As the variation of the content of Co and Mn can be used for HEAs of BCC+B2 phase configuration could be attributed to the
to adjust the fraction of B2 and BCC structures, obtaining three difference in microstructures. Noted that alloy design with similar
AlCoCrFeMnNi HEAs with hardness higher than 600HV. hardness can result from different weights of solid solution strength­
As opposed to conventional one or two-principal-element alloys, ening and precipitation strengthening or even fine-grain strengthening
HEAs can be altered significantly in mechanical properties and micro­ [106]. To elaborate, Al has a larger atomic radius than other transitional
structures due to the addition of certain elements, even very subtle metals, so high concentration of Al can distort the lattice making solid
changes in elemental fractions. The correlations between compositions solution strengthening. HEAs of B2 matrix with cuboidal BCC structures
and microstructures observed in our study are similar to reported lit­ are less hard while BCC matrix with B2 spinodal decomposition due to
eratures [24,74,75,91,98–100] These reports have shed light on Al precipitation hardening show higher hardness [74]. Bai et al. [107]
addition to 3d transitional HEAs because of three reasons: 1. stabiliza­ reported a study on Alx (Fe40Mn40Ni10Cr10)100− x HEAs, it was found that
tion of BCC structures due to its low valence electron concentration BCC/B2 lattice misfit increases sharply along with Al addition, and
(VEC); 2. large atomic radius makes lattice distortion 3; affinity of transformation of B2 morphologies from spherical to cuboidal and
forming ordered intermetallic, such as L12 and B2. In AlxCrFeMnNi finally to rod-like was observed. The BCC structures containing fine
system, when Al% is 0–6%, dual phases of FCC (Mn, Ni) and BCCs is mixtures of BCC and B2 which have similar lattice parameters and often
observed; BCC (Fe, Cr) matrix +B2 precipitate (Al, Ni) dominates when coherent; high density of interfaces between them may also contribute
Al% exceeds 10%. Another XRD pattern indicates that the phase in to high hardness, shown from transmission electron microscopy of Alx
Alx(CoCrFeMnNi)100− x exhibit FCC as Al% is less than 4%, duplex alloys CoCrCuFeNi HEAs [108]. The morphologies and the volume fractions of
with Al content is between 8% and 16%, and simple BCC alloys with each phase are still not be studied by currently ML models. These
more than 16% of Al [22]. To extensively study elemental contribution important issues should be paid attention in the further.
of eutectic Al-Co-Cr-Fe-Ni HEAs, Wu et al. [21] found that Al is critical
for phase formation as eutectic point is located around 18% of Al. Cr is 5. Conclusion
strongly associated with Al content since Co, Fe and Ni can all form
B2-intermetallic compounds with Al or their subsets can form FCC solid This study proposes a Deep Interpretable Scheme for Materials
solutions while Cr cannot. Al is confirmed again as the critical element Attribution Analysis (DISMAA) that integrates VAE model, MLP pre­
for phase transformation. Only few studies have addressed the contri­ dictors, attribution analysis, and sensitivity analysis. With the visuali­
butions of element other than Al, with most of them focusing on com­ zation of latent space, DISMAA provides a new perspective and intuitive
positions with an equiatomic composition ratio, involving single guideline for scientists to discover new materials and systematically
element substitutions within a narrow concentration range. extract the composition-structure-property relationships. The analysis
With respect to the compositions in this latent space, from our results show that the concentrations of Al and Cr are positively corre­
observation, our results are identical to Alx(CoCrFeMnNi)100− x system lated with hardness in AlCoCrFeMnNi system, while others are nega­
which shows a drastic increase in hardness when Al% between 8% and tively related to the variation of hardness. The order of correlation
16% [21]. Alloys with more than 16% of Al, hardness becomes saturated magnitude to hardness are |Ni|> |Fe| >|Cr|>|Al|> |Co|>|Mn| in
and remains almost constant essentially. Cr is again proved to be highly AlCoCrFeMnNi system. Al was found to be the most important element
associated with hardness, particularly for high-hardness HEAs while it is in AlCoCrFeMnNi system in terms of phase transformation, followed by
not determinant in phase formation. Shun et al. [101] found that Cr Ni. The variation of both Al and Ni are very sensitive to the hardness
addition increases the hardness of AlCoCrxFeNi alloys because Cr is a variation at local content regions. The variation of Al content within the
BCC former in AlCoCrxFeNi alloys. It can be explained that solute range of 10–20% significantly contributes to the medium hardness
strengthening effect and Cr’s high melting point provide great slip (250–400Hv) of the AlCoCrFeMnNi system, while the increase of Cr
resistance. We can therefore conclude that Cr is relatively critical in plays the role in enhancing the hardness, particularly in high hardness
BCC-based structures. Ni concentration is seen to be inversely propor­ (>500 Hv). In addition, Al was found to be the most important element
tional to hardness, and especially sensitive for HEAs of duplex phases. in AlCoCrFeMnNi system with respect to control phase transformation,
The effect of Ni in AlCoCrFeNix compositions was studied in López Ríos followed by Ni. The increase of Cr plays the role to enhance the hard­
et al. [102];they found that hardness values decrease as the percentage ness, especially in high hardness region. The Co and Mn are relatively
of Ni inclines due to dissolution of precipitation in Ni-rich matrix not significant for hardness variation and phase transformation. To
resulting solid solutions. Besides, the grain size decreases considerably further enhance the capability of phase prediction, identifying micro­
as Ni is added. In addition, the addition of Ni is found to led the for­ structures configuration and decomposition quantitatively are crucial
mation of plastic FCC phases, which causes a lower strength and higher for understanding HEA. To further quantify the relationship between
plasticity [103]. Hence, Ni plays an important role since it affects both properties and phases, it is crucial to consider the influence of micro­
hardness and phase formation, which is consistent with our results as structure configurations, precipitates, and their proportions in ML
shown in Fig. 5 and Fig. 7. Fe content also has highly negative model. However, there hasn’t been an open or commercial database

9
­
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

collecting such comprehensive information in HEAs. It is expected that [5] E.-W. Huang, et al., Machine-learning and high-throughput studies for high-
entropy materials, Mater. Sci. Eng. R. Rep. 147 (2022), 100645.
machine learning (ML) models will achieve even greater success when
[6] C.-J. Tong, et al., Microstructure characterization of AlxCoCrCuFeNi high-entropy
more comprehensive experimental databases are available for training alloy system with multiprincipal elements, Metall. Mater. Trans. A 36 (2005)
in the future. 881–893.
[7] X.F. Wang, Y. Zhang, Y. Qiao, G.L. Chen, Novel microstructure and properties of
multicomponent CoCrCuFeNiTix alloys, Intermetallics 15 (2007) 357–362.
CRediT authorship contribution statement [8] S. Gorsse, D.B. Miracle, O.N. Senkov, Mapping the world of complex concentrated
alloys, Acta Mater. (2017), https://doi.org/10.1016/j.actamat.2017.06.027.
[9] C.-Y. Hsu, J.-W. Yeh, S.-K. Chen, T.-T. Shun, Wear resistance and high-
Chih-Yu Lee and Chia-Yung Jui conducted the research under the
temperature compression strength of Fcc CuCoNiCrAl0.5Fe alloy with boron
guidance of Wen-Jay Lee, with equal contributions from both. Chih-Yu addition, Metall. Mater. Trans. A 35 (2004) 1465–1469.
Lee, Chia-Yung Jui and Wen-Jay Lee collaborated on writing the [10] Z. Li, K.G. Pradeep, Y. Deng, D. Raabe, C.C. Tasan, Metastable high-entropy dual-
manuscript. An-Chou Yeh and Yao-Jen Chang contributed to the anal­ phase alloys overcome the strength–ductility trade-off, Nature 534 (2016)
227–230.
ysis of CALPHAD simulation results and the data collection for ML [11] Y. Lu, et al., A promising new class of high-temperature alloys: eutectic high-
training. All authors participated in the revision process and approved entropy alloys, Sci. Rep. 4 (2014) 6200.
the final manuscript. [12] C. Li, J.C. Li, M. Zhao, Q. Jiang, Effect of aluminum contents on microstructure
and properties of AlxCoCrFeNi, Alloy. J. Alloy. Compd. 504 (2010) S515–S518.
[13] C. Li, J.C. Li, M. Zhao, L. Zhang, Q. Jiang, Microstructure and properties of
AlTiNiMnB x high entropy alloys, Mater. Sci. Technol. 24 (2008) 376–378.
Declaration of Competing Interest
[14] X. Qiu, Microstructure, hardness and corrosion resistance of Al2CoCrCuFeNiTix
high-entropy alloy coatings prepared by rapid solidification, J. Alloy. Compd. 735
The authors declare the following financial interests/personal re­ (2018) 359–364.
[15] S. Varalakshmi, M. Kamaraj, B.S. Murty, Synthesis and characterization of
lationships which may be considered as potential competing interests:
nanocrystalline AlFeTiCrZnCu, High. entropy Solid Solut. Mech. Alloy. J. Alloy.
Wen-Jay Lee reports financial support was provided by National Center Compd. 460 (2008) 253–257.
for High-Performance Computing, Taiwan (R.O.C.). Wen-Jay Lee reports [16] K.B. Zhang, et al., Characterization of nanocrystalline CoCrFeNiTiAl high-entropy
a relationship with National Center for High-Performance Computing, solid solution processed by mechanical alloying, J. Alloy. Compd. 495 (2010)
33–38.
Taiwan (R.O.C.) that includes: employment. The first author Chih-Yu [17] K. Zhang, Z. Fu, Effects of annealing treatment on phase composition and
Lee now is pursuing PhD degree in University of Maryland. But this microstructure of CoCrFeNiTiAlx high-entropy alloys, Intermetallics 22 (2012)
work was done before she joined the current institute. 24–32.
[18] D.B. Miracle, Critical Assessment 14: high entropy alloys and their development
as structural materials, Mater. Sci. Technol. 31 (2015) 1142–1147.
Data Availability [19] B.S. Li, Y.P. Wang, M.X. Ren, C. Yang, H.Z. Fu, Effects of Mn, Ti and V on the
microstructure and properties of AlCrFeCoNiCu high entropy alloy, Mater. Sci.
Eng. A 498 (2008) 482–486.
I have shared the link to my data/code at the Attach File step. [20] C.-C. Tung, et al., On the elemental effect of AlCoCrCuFeNi high-entropy alloy
DISMA_Research Dataset (High-entropy alloys) (Original data) system, Mater. Lett. 61 (2007) 1–5.
(Mendeley Data) [21] Q. Wu, et al., Uncovering the eutectics design by machine learning in the
Al–Co–Cr–Fe–Ni high entropy system, Acta Mater. 182 (2020) 278–286.
[22] J.Y. He, et al., Effects of Al addition on structural evolution and tensile properties
of the FeCoNiCrMn high-entropy alloy system, Acta Mater. 62 (2014) 105–113.
Acknowledgement [23] C. Zhang, F. Zhang, S. Chen, W. Cao, Computational thermodynamics aided high-
entropy alloy design, JOM 64 (2012) 839–845.
[24] N.D. Stepanov, D.G. Shaysultanov, R.S. Chernichenko, M.A. Tikhonovsky, S.
The authors would like to thank the National Science and Technol­ V. Zherebtsov, Effect of Al on structure and mechanical properties of Fe-Mn-Cr-
ogy Council for financially supporting this research under Contract Ni-Al non-equiatomic high entropy alloys with high Fe, Content J. Alloy. Compd.
770 (2019) 194–203.
NSTC 111–2221-E-492-012, NSTC 112–2927-I-007-504, NSTC
[25] X. Yang, Y. Zhang, Prediction of high-entropy stabilized solid-solution in multi-
110–2221-E-007–020-MY3, and NSTC 112-2224-E-007-003. Support of component alloys, Mater. Chem. Phys. 132 (2012) 233–238.
significant computing resources of TWCC from the NCHC in Taiwan is [26] B.S. Murty, S. Ranganathan, J.W. Yeh, P.P. Bhattacharjee, High-entropy alloys.
high-entropy alloys, Elsevier,, 2019, https://doi.org/10.1016/C2017-0-03317-7.
also acknowledged. In addition, this work was also supported by the
[27] D.J.M. King, S.C. Middleburgh, A.G. McGregor, M.B. Cortie, Predicting the
“High Entropy Materials Center” from The Featured Areas Research formation and stability of single phase high-entropy alloys, Acta Mater. 104
Center Program within the framework of the Higher Education Sprout (2016) 172–179.
Project by the Ministry of Education (MOE) in Taiwan. [28] Q. Hu, et al., Parametric study of amorphous high-entropy alloys formation from
two new perspectives: atomic radius modification and crystalline structure of
alloying elements, Sci. Rep. 7 (2017) 39917.
[29] O.N. Senkov, J.D. Miller, D.B. Miracle, C. Woodward, Accelerated exploration of
Code availability multi-principal element alloys with solid solution phases, Nat. Commun. 6 (2015)
6529.
The authors will make available, upon request, the code used in the [30] R. Feng, P.K. Liaw, M.C. Gao, M. Widom, First-principles prediction of high-
entropy-alloy stability, Npj Comput. Mater. 3 (2017) 50.
applications described in this work. It is understood that the code will
[31] B.S. Murty, J.W. Yeh, S. Ranganathan, Phase Selection in High-Entropy Alloys. in
not be for commercial use. High Entropy Alloys, Elsevier,, 2014, pp. 37–56, https://doi.org/10.1016/B978-
0-12-800251-3.00003-1.
[32] C. Niu, A.J. Zaddach, C.C. Koch, D.L. Irving, First principles exploration of near-
Appendix A. Supporting information equiatomic NiFeCrCo, High. entropy Alloy. J. Alloy. Compd. 672 (2016) 510–520.
[33] B. Yin, W.A. Curtin, First-principles-based prediction of yield strength in the
Supplementary data associated with this article can be found in the RhIrPdPtNiCu high-entropy alloy, Npj Comput. Mater. 5 (2019) 14.
[34] N. Qu, Y. Chen, Z. Lai, Y. Liu, J. Zhu, The phase selection via machine learning in
online version at doi:10.1016/j.jallcom.2023.173144.
high entropy alloys, Procedia Manuf. 37 (2019) 299–305.
[35] W. Huang, P. Martin, H.L. Zhuang, Machine-learning phase prediction of high-
References entropy alloys, Acta Mater. 169 (2019) 225–236.
[36] N. Islam, W. Huang, H.L. Zhuang, Machine learning for phase selection in multi-
principal element alloys, Comput. Mater. Sci. 150 (2018) 230–235.
[1] D.B. Miracle, O.N. Senkov, A critical review of high entropy alloys and related
[37] Z. Zhou, et al., Machine learning guided appraisal and exploration of phase design
concepts, Acta Mater. 122 (2017) 448–511.
for high entropy alloys, Npj Comput. Mater. 5 (2019) 128.
[2] J.-W. Yeh, et al., Nanostructured high-entropy alloys with multiple principal
[38] S. Dixit, V. Singhal, A. Agarwal, A.K. Prasada Rao, Multi-label phase-prediction in
elements: novel alloy design concepts and outcomes, Adv. Eng. Mater. 6 (2004)
high-entropy-alloys using Artificial-Neural-Network, Mater. Lett. 268 (2020),
299–303.
127606.
[3] Y.F. Ye, Q. Wang, J. Lu, C.T. Liu, Y. Yang, High-entropy alloy: challenges and
[39] S.Y. Lee, S. Byeon, H.S. Kim, H. Jin, S. Lee, Deep learning-based phase prediction
prospects, Mater. Today 19 (2016) 349–362.
of high-entropy alloys: Optimization, generation, and explanation, Mater. Des.
[4] L. Qiao, Y. Liu, J. Zhu, A focused review on machine learning aided high-
197 (2021), 109260.
throughput methods in high entropy alloy, J. Alloy. Compd. 877 (2021), 160295.

10
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

[40] D. Dai, et al., Using machine learning and feature engineering to characterize [76] S. Gorsse, M.H. Nguyen, O.N. Senkov, D.B. Miracle, Database on the mechanical
limited material datasets of high-entropy alloys, Comput. Mater. Sci. 175 (2020), properties of high entropy alloys and complex concentrated alloys, Data Brief. 21
109618. (2018) 2664–2678.
[41] K. Kaufmann, K.S. Vecchio, Searching for high entropy alloys: a machine learning [77] J.M. Zhu, et al., Microstructures and compressive properties of multicomponent
approach, Acta Mater. 198 (2020) 178–222. AlCoCrFeNiMox alloys, Mater. Sci. Eng. A (2010), https://doi.org/10.1016/j.
[42] A. Roy, T. Babuska, B. Krick, G. Balasubramanian, Machine learned feature msea.2010.07.028.
identification for predicting phase and Young’s modulus of low-, medium- and [78] Y.F. Kao, T.J. Chen, S.K. Chen, J.W. Yeh, Microstructure and mechanical property
high-entropy alloys, Scr. Mater. 185 (2020) 152–158. of as-cast, -homogenized, and -deformed AlxCoCrFeNi (0 ≤ x ≤ 2) high-entropy
[43] Xiong, J., Shi, S.Q. & Zhang, T.Y. Analysis of phase formations and mechanical alloys, J. Alloy. Compd. (2009), https://doi.org/10.1016/j.jallcom.2009.08.090.
properties in complex concentrated alloys by machine learning approach. arXiv [79] B. Gwalani, et al., Dynamic shear deformation of a precipitation hardened
Preprint at (2020). Al0.7CoCrFeNi eutectic high-entropy alloy using hat-shaped specimen geometry,
[44] F.Z. Dai, B. Wen, Y. Sun, H. Xiang, Y. Zhou, Theoretical prediction on thermal and Entropy 22 (2020) 431.
mechanical properties of high entropy (Zr0.2Hf0.2Ti0.2Nb0.2Ta0.2)C by deep [80] A. Shi, R. Li, Y. Zhang, Z. Wang, Z. Guo, Effect of enhanced gravity on the
learning potential, J. Mater. Sci. Technol 43 (2020) 168–174. microstructure and mechanical properties of al0.9 cocrfeni high-entropy alloy,
[45] Y. Zhang, et al., Microstructures and properties of high-entropy alloys, Prog. Entropy 22 (1–12) (2020).
Mater. Sci. 61 (2014) 1–93. [81] S. Gorsse, M.H. Nguyen, O.N. Senkov, & Miracle, D. B. Database on the
[46] D.N. Klimenko, N.Y. Yurchenko, N.D. Stepanov, S.V. Zherebtsov, Prediction of mechanical properties of high entropy alloys and complex concentrated alloys,
strength characteristics of high-entropy alloys Al-Cr-Nb-Ti-V-Zr systems, Mater. Data Brief. 21 (2018) 2664–2678.
Today Proc. 38 (2021) 1535–1540. [82] Y. Ma, et al., Temperature-affected microstructural stability of coherent cuboidal
[47] N. Filipoiu, G.A. Nemnes, Prediction of equilibrium phase, stability and stress- B2 particles in precipitation-strengthened body-centered-cubic Al0.7CoCr2FeNi
strain properties in co-cr-fe-ni-al high entropy alloys using artificial neural high-entropy alloy, J. Mater. Sci. 54 (8696–8710) (2019).
networks, Metals 10 (2020) 1–10. [83] X. Gao, et al., Microstructural origins of high strength and high ductility in an
[48] A. Choudhury, T. Konnur, P.P. Chattopadhyay, S. Pal, Structure prediction of AlCoCrFeNi2.1 eutectic high-entropy alloy, Acta Mater. 141 (2017) 59–66.
multi-principal element alloys using ensemble learning, Eng. Comput. 37 (2019) [84] Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. in
1003–1022. 34th International Conference on Machine Learning, ICML 2017 (2017).
[49] L. Zhang, et al., Machine learning reveals the importance of the formation [85] J.M. Rickman, et al., Materials informatics for the screening of multi-principal
enthalpy and atom-size difference in forming phases of high entropy alloys, elements and high-entropy alloys, Nat. Commun. 10 (2019) 1–10.
Mater. Des. (2020), https://doi.org/10.1016/j.matdes.2020.108835. [86] J.M. Rickman, G. Balasubramanian, C.J. Marvel, H.M. Chan, M.-T. Burton, Mach.
[50] Z. Pei, J. Yin, J.A. Hawk, D.E. Alman, M.C. Gao, Machine-learning informed Learn. Strateg. High. -entropy Alloy. J. Appl. Phys. 128 (2020), 221101.
prediction of high-entropy solid solution formation: Beyond the Hume-Rothery [87] C. Wen, et al., Machine learning assisted design of high entropy alloys with
rules, Npj Comput. Mater. 6 (2020) 50. desired property, Acta Mater. 170 (2019) 109–117.
[51] Y. Zhang, et al., Phase prediction in high entropy alloys with a rational selection [88] B. Sanchez-Lengeling, A. Aspuru-Guzik, Inverse molecular design using machine
of materials descriptors and machine learning models, Acta Mater. 185 (2020) learning:generative models for matter engineering, Science 361 (2018) 360–365.
528–539. [89] A.G. Kusne, et al., On-the-fly closed-loop materials discovery via Bayesian active
[52] R. Machaka, Machine learning-based prediction of phases in high-entropy alloys, learning, Nat. Commun. 11 (2020) 5966.
Comput. Mater. Sci. 188 (2021), 110244. [90] J.Y. He, et al., Effects of Al addition on structural evolution and tensile properties
[53] A. Roy, G. Balasubramanian, Predictive descriptors in machine learning and data- of the FeCoNiCrMn high-entropy alloy system, Acta Mater. 62 (2014) 105–113.
enabled explorations of high-entropy alloys, Comput. Mater. Sci. 193 (2021), [91] W.-R. Wang, et al., Effects of Al addition on the microstructure and mechanical
110381. property of AlxCoCrFeNi high-entropy alloys, Intermetallics 26 (2012) 44–51.
[54] J. Xiong, S.Q. Shi, T.Y. Zhang, Mach. Learn. phases Mech. Prop. Complex Conc. [92] W.-R. Wang, W.-L. Wang, J.-W. Yeh, Phases, microstructure and mechanical
Alloy. J. Mater. Sci. Technol. 87 (2021) 133–142. properties of AlxCoCrFeNi, High. -entropy Alloy. Elev. Temp. J. Alloy. Compd. 589
[55] E. Menou, et al., Evolutionary design of strong and stable high entropy alloys (2014) 143–152.
using multi-objective optimisation based on physical models, statistics and [93] M. Aizenshtein, E. Strumza, E. Brosh, S. Hayun, Precipitation kinetics,
thermodynamics, Mater. Des. 143 (2018) 185–195. microstructure, and equilibrium state of A2 and B2 phases in multicomponent
[56] B. Debnath, A. Vinoth, M. Mukherjee, S. Datta, Designing Fe-based high entropy Al2.75CoCrFeNi alloy, J. Mater. Sci. 55 (7016–7028) (2020).
alloy-a machine learning approach, IOP Conf. Ser. Mater. Sci. Eng. (912) (2020). [94] Y. Lu, et al., Preparing bulk ultrafine-microstructure high-entropy alloys via
[57] Buranich, V., Rogoz, V., Postolnyi, B. & Pogrebnjak, A. Predicting the Properties direct solidification, Nanoscale 10 (2018) 1912–1919.
of the Refractory High-Entropy Alloys for Additive Manufacturing-Based [95] Y. Ma, et al., Controlled formation of coherent cuboidal nanoprecipitates in body-
Fabrication and Mechatronic Applications. Proc. 2020 IEEE 10th Int. Conf. centered cubic high-entropy alloys based on Al2(Ni,Co,Fe,Cr)14 compositions,
Nanomater. Appl. Prop. NAP 2020 9–13 (2020) doi:10.1109/NAP51477.202 Acta Mater. 147 (2018) 213–225.
0.9309720. [96] X. Jin, Y. Liang, J. Bi, B. Li, Non-monotonic variation of structural and tensile
[58] T. Zheng, et al., Tailoring nanoprecipitates Ultra-Strong High. -entropy Alloy. via properties with Cr content in AlCoCrxFeNi2, High. entropy Alloy. J. Alloy. Compd.
Mach. Learn. prestrain Aging J. Mater. Sci. Technol. 69 (2021) 156–167. 798 (2019) 243–248.
[59] J.M. Rickman, Data analytics and parallel-coordinate materials property charts, [97] G.J. Zhang, et al., Effect of Fe on microstructure and properties of AlCoCrFexNi
Npj Comput. Mater. 4 (1) (2018) 8. (x=1.5, 2.5) high entropy alloy coatings prepared by laser cladding,
[60] J.M. Rickman, et al., Materials informatics for the screening of multi-principal Intermetallics 119 (2020), 106722.
elements and high-entropy alloys, Nat. Commun. 10 (2019) 2618. [98] C.-J. Tong, et al., Mechanical performance of the Al x CoCrCuFeNi high-entropy
[61] Q. Wu, et al., Uncovering the eutectics design by machine learning in the alloy system with multiprincipal elements, Metall. Mater. Trans. A 36 (2005)
Al–Co–Cr–Fe–Ni high entropy system, Acta Mater. 182 (2020) 278–286. 1263–1271.
[62] S. Lapuschkin, et al., Unmasking Clever Hans predictors and assessing what [99] Z. Tang, et al., Aluminum alloying effects on lattice types, microstructures, and
machines really learn, Nat. Commun. 10 (2019) 1–8. mechanical behavior of high-entropy alloys systems, JOM (2013), https://doi.
[63] J. Noh, et al., Inverse design of solid-state materials via a continuous org/10.1007/s11837-013-0776-z.
representation, Matter 1 (2019) 1370–1384. [100] N.D. Stepanov, et al., Effect of thermomechanical processing on microstructure
[64] R. Gómez-Bombarelli, et al., Automatic chemical design using a data-driven and mechanical properties of the carbon-containing CoCrFeNiMn high entropy
continuous representation of molecules, ACS Cent. Sci. 4 (2018) 268–276. alloy, J. Alloy. Compd. (2017), https://doi.org/10.1016/j.jallcom.2016.09.208.
[65] De Cao, N., Kipf, T. MolGAN: An implicit generative model for small molecular [101] T.-T. Shun, W.-J. Hung, Effects of Cr content on microstructure and mechanical
graphs. arXiv (2018). properties of AlCoCrxFeNi high-entropy alloy, e5826467, Adv. Mater. Sci. Eng.
[66] Nouira, A., Sokolovska, N., Crivello, J.-C. CrystalGAN: Learning to Discover 2018 (2018). e5826467.
Crystallographic Structures with Generative Adversarial Networks. arXiv (2018). [102] M. López Ríos, et al., Effects of nickel content on the microstructure,
[67] Hatakeyama-Sato, K., Oyaizu, K. A Generative Model for Extrapolation Prediction microhardness and corrosion behavior of high-entropy AlCoCrFeNix alloys, Sci.
in Materials Informatics. arXiv (2021). Rep. 10 (2020) 1–11.
[68] A. Debnath, et al., Generative deep learning as a tool for inverse design of high [103] Ł. Rogal, et al., Microstructure and mechanical properties of Al–Co–Cr–Fe–Ni
entropy refractory alloys. J. Mater, Inform 1 (2021) 3. base high entropy alloys obtained using powder metallurgy, Met. Mater. Int. 25
[69] Kingma, D.P., Welling, M. Auto-Encoding Variational Bayes, CoRR, abs/ (2019) 930–945.
1312.6114. , 2013. [104] Q. Chen, K. Zhou, L. Jiang, Y. Lu, T. Li, Effects of Fe content on microstructures
[70] Chollet, F. & others. Keras. 〈https://github.com/fchollet/keras〉 , 2015. and properties of AlCoCrFexNi high-entropy alloys, Arab. J. Sci. Eng. 40 (2015)
[71] F. Pedregosa, et al., Scikit-Learn.: Mach. Learn. Python 12 (2011) 2825–2830. 3657–3663.
[72] Shrikumar, A., Greenside, P. & Kundaje, A. Learning important features through [105] G. Qin, et al., Effect of Co content on phase formation and mechanical properties
propagating activation differences. in 34th International Conference on Machine of (AlCoCrFeNi)100-xCox high-entropy alloys, Mater. Sci. Eng. A 710 (2018)
Learning, ICML 2017 (2017). 200–205.
[73] R. Gómez-Bombarelli, et al., Automatic chemical design using a data-driven
continuous representation of molecules, ACS Cent. Sci. 4 (2018) 268–276.
[74] Y.-J. Chang, C.-Y. Jui, W.-J. Lee, A.-C. Yeh, Prediction of the composition and
hardness of high-entropy alloys by machine learning, JOM 71 (2019) 3433–3442.
[75] J.Y. He, et al., Effects of Al addition on structural evolution and tensile properties
of the FeCoNiCrMn high-entropy alloy system, Acta Mater. 62 (2014) 105–113.

11
C.-Y. Lee et al. Journal of Alloys and Compounds 976 (2024) 173144

[106] X. Chen, et al., Effects of aluminum on microstructure and compressive properties [107] L. Bai, et al., Effects of Al addition on microstructure and mechanical properties of
of Al-Cr-Fe-Ni eutectic multi-component alloys, Mater. Sci. Eng. A 681 (2017) Co-free (Fe40Mn40Ni10Cr10)100− xAlx high-entropy alloys, J. Alloy. Compd.
25–31. 879 (2021), 160342.
[108] B.A. Welk, et al., Nature of the interfaces between the constituent phases in the
high entropy alloy CoCrCuFeNiAl, Ultramicroscopy 134 (2013) 193–199.

12