Review

www.afm-journal.de

Innovative Materials Science via Machine Learning

Chaochao Gao, Xin Min,* Minghao Fang, Tianyi Tao, Xiaohong Zheng, Yangai Liu,
Xiaowen Wu, and Zhaohui Huang*

statistical machine learning, such as deep

Nowadays, the research on materials science is rapidly entering a phase of learning,[9] reinforcement learning,[10]
data-driven age. Machine learning, one of the most powerful data-driven adversarial learning,[11] etc. which could
methods, have been being applied to materials discovery and performances double improve the research efficiency
prediction with undoubtedly tremendous application foreground. Herein, in the fields of information data mining,
data classification, and new data predic-
the challenges and current progress of machine learning are summarized in
tion. In present, AI based on this machine
materials science, the design strategies are classified and highlighted, and learning process has become the mostly
possible perspectives are proposed for the future development. It is hoped important driving force for science and
this review can provide important scientific guidance for innovating materials industrial revolution.
science and technology via machine learning in the future. Material innovation always plays a great
role in the process of industrial revolu-
tion and human social development.[12]
For example, the advanced alloys, semi-
1. Introduction conductor materials, polymer materials, composite materials,
superconducting materials and biocompatible materials have
At present stage of human society, artificial intelligence (AI) been promoting the technology revolution of new energy,
technology is becoming a tendency of developing all over the microelectronics, bioengineering technology and space tech-
world. Especially, the successful application in vision recogni- nology, which opened the door to the unprecedented infor-
tion,[1] speech recognition,[2] natural language understanding,[3] mation society.[13,14] In the nearly future, the develop of novel
and man-machine game[4] also accelerates the integration of material science and technology is accelerating to the intercon-
different knowledge and the cross of multi-disciplinary. For nected and intelligent direction, which will further promote the
example, the so-called AlphaGo Zero achieved a score of 100:0 fourth industrial revolution – green intelligent industry.[15,16]
in the game with the previous champion AlphaGo Lee in 2017;[5] In comparison with the traditional experiment- and computa-
A Nature Biotechnology paper published in 2019 reported six tion-driven research methods for materials science, the data-
molecules based on a generative model, four of which have driven mode, which integrates high-throughput experiments,
measurable biochemical activity;[6] One Nature cover article high-throughput computing and material data based on data
reported a tireless mobile robotic chemist in 2020, which was mining and AI, would be more revolutionary in the future
able to operate autonomously over eight days, performing 688 development.[17–20]
experiments within a ten-variable experimental space, and As one of the most essential AI methods, machine learning
finally selected a novel and efficient photocatalyst.[7] is becoming an excellent tool for material innovation due to its
Generally, the implementation process of AI would be low computing cost, short development cycle, strong data anal-
divided into three stages: handwritten knowledge, statistical ysis and prediction ability (Figure 1). Nowadays, the machine
learning, and context adaptation,[8] which is currently in the learning has been used and shown great potential in materials
statistical learning level. The key core technology depends on prediction for magnetocaloric effect,[21] bandgap,[22] dielec-
tric constant,[23] quantum chemistry,[24] thermal properties,[25]
new materials design, and discovery.[26] Summing up the pre-
C. Gao, X. Min, M. Fang, X. Zheng, Y. Liu, X. Wu, Z. Huang
Beijing Key Laboratory of Materials Utilization of Nonmetallic Minerals sent development of data-driven machine learning, it is still at
and Solid Wastes primary stage in material innovation. As far as we know, few
National Laboratory of Mineral Materials comprehensive review papers on machine learning and mate-
School of Materials Science and Technology rial applications have attracted close attention in recent years,
China University of Geosciences (Beijing)
Beijing 100083, China
focusing on machine learning in chemical discovery, energy
E-mail: [email protected]; [email protected] materials, solid-state materials science and so on.[17–20] Recently,
T. Tao, X. Zheng there has been a sharp increase in the number of articles on
Division of Environment Technology and Engineering machine learning and materials science, which suggests that
Institute of Process Engineering the research direction and content have been widely extended.
Chinese Academy of Sciences Thus, it is significantly important to update the summarize the
Beijing 100190, China
prospective in this field. Herein, this review attempts to high-
The ORCID identification number(s) for the author(s) of this article light the main challenges of machine learning in materials
can be found under https://doi.org/10.1002/adfm.202108044.
science, focus on the research progress of existing strategies,
DOI: 10.1002/adfm.202108044 and possible directions and perspective of machine learning in

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (1 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Figure 1. The introduction of AI and machine learning in materials science.

materials innovation. Our main objective is to illustrate main understanding of thermal properties of amorphous solids,
strategies for machine learning in materials, provide possible identification of high temperature thermal conductors, clas-
directions for next-generation machine learning, and sug- sification of crystal structure, magnetocaloric effect, etc.[27–54]
gest methodologies for the future development of materials As well as the promotion of the material Genome Project,[55]
sciences. machine learning has been combined with density functional
theory (DFT) calculation to establish a development model
for material research, such as prediction of thermodynamic
2. Challenges of Machine Learning in Materials stability, bandgaps, AB2C2 compounds,[44,56–58] and graphene-
based bimetallic catalysts.[59,60] Furthermore, machine learning
Science
could even replace DFT to achieve the desired goals, including
In the development process of materials science, the traditional the prediction of crystal structure, adsorption energy on metal
methods slow down the pace of materials research because of alloys and so on.[40,43,61] Besides, the failed experimental data
the long research cycle, potential data information loss and about crystallization of templated vanadium selenite was also
low efficiency. The up-to-date machine learning expectations been used for information mining by machine learning.[62]
based on material data start from three components, including Although the cross-research of materials science associated
data, algorithms, and models. Firstly, the material data should with machine learning has shown special advantages, there
be collected and obtained, then preprocess these data. For dif- are still several challenges (Figure 2) which need to be properly
ferent material data, the operations, such as encoding conver- address and deeply understood as shown below.
sion, feature crossover, calculation optimization, filtering and
deleting, may be performed, and the optimized data would be
divided into training set and test set. After that, the appropriate 2.1. Insufficient Data
algorithm model for machine learning would be selected and
trained based on the training set. The test set is used for eval- Data is not only the basic premise of the fourth paradigm of
uating and adjusting the algorithm model to achieve the best materials science, but also the first challenge to be solved in
evaluation. Finally, the optimized algorithm model would be application of materials.[63] Unlike most disciplines, materials
effective for further data prediction to obtain the desired target. science and technology owns many different data categories.
Therefore, combined with the machine learning technology, the More severely, the data output of each category is less, and the
efficiency of materials research could be significantly improved. feature dimension is lower. For example, even a simple experi-
For example, machine learning can give some guiding mental data always depends on different controllable factors,
conditions in many fields for predicting crystal structure, such as raw materials, contents, temperatures, times, humidity,

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (2 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Figure 2. Summary of challenges for the machine learning in materials science.

etc. In most studies, the research data are obtained via limited 3) Difficulty in feature classification. Since the relationship
experimental conditions, which takes a long time in the mean- among the features remains unclear, their influences on
time. Thus, these measured results are just a small sampling, materials science should be systematically discussed. When
which also lacks comparability in different published studies. combining with the machine learning technology, it is still
It could be concluded that sufficient data can not be obtained difficult to scientifically clarify these features as needed.
quickly to meet the operation of machine learning, which limit 4) Difficulty in feature recognition. The initially collected data
their further development in materials science. often cannot be directly identified for machine learning.
Many modifying strategies are needed to convert them into
data features to be well recognized. However, each conver-
2.2. Difficult to Identify, Classify, and Correlate the Features sion is different, which also makes the feature recognition
become a huge limitation of data-driven development in ma-
In the process of material research, the key point is to explore terials science.
the relationship among the four elements of materials,
which could be controlled by the features of experimental
parameters.[64–66] The quantity, quality, form, and relationship 2.3. The Specific Algorithm with Poor Universality
of these features are essential for machine learning. In general,
the initial data features can not fit the ideal service conditions Machine learning shows different abilities based on various
for machine learning, thus the feature engineering operations, algorithms, for example, different results would be obtained by
such as adding, deleting, and modifying, are necessary. Herein, the same algorithm to deal with different objects or different
four challenges related to features are shown as below: algorithms to deal with the same object. In materials science,
the machine learning algorithms needs to be especially con-
1) Processing difficulties because of the large number of fea- sidered in different directions because of the rich research
tures. With the promotion of high-throughput computing[67] fields, great differences, and small relevance. Moreover, the
and Materials Genome Initiative (MGI),[68] the number of performance optimization, structure regulation, and some other
features, as well as the noise and redundancy of the origi- directions also should be considered during the new materials
nal data, continue to increase. Finally, the computational and development, so that the selection and optimization of algo-
storage became more and more complex, so that the tradi- rithms is an issue of concern. In the field of materials science,
tional learning methods are no longer suitable to gain satis- the relationship and interaction mechanism between the four
factory results. elements “component-structure-performance-application” usu-
2) Challenges to explore the relationship among various fea- ally needs to be explored firstly, which could provide theoretical
tures. It is mentioned before that researcher can get many support for the materials research.[69–71] However, that raised
data with a variety of features via experiments. When study- much higher challenges to the targeted selection and optimi-
ing the performances, the control variable method is used to zation of machine learning algorithms. Therefore, different
explore the optimized parameters. However, it is complicated objects and different situations demand different algorithms
to establish the relationship in multiple features. without universality.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (3 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

2.4. The Abstract Model with Poor Interpretability and successfully synthesized, which exhibits a high measured
density, high thermal decomposition temperature, high deto-
The deep neural network, which belongs to the representative nation velocity, and extremely low sensitivities. While Pablo
machine learning model, has developed rapidly and shown pow- and the co-workers put forward the application of information
erful ability in intelligent products. For example: self-driving materials, functional materials, energy and catalytic materials
cars,[72] image recognition,[73] dialogue systems.[74] However, the according to MGI in ref. [68]. Through the MIG strategy, both
internal working mechanism is complicated to understand the the data generation and the ability to identify material attrib-
decision-making model in human inborn thought processes. utes are strengthened in cooperation with machine learning.
In the field of materials science, machine learning is always Besides, several strategies, including high-throughput cal-
regard as a “black box” and the principles behind data-driven culation,[67] DFT calculation[81] and traditional experimental
are ignored in most studies. For strict research for materials, data accumulation in published papers,[27] also provide a large
the prediction of properties might not be well trusted in this number of valuable data for machine learning in materials
situation. Thus, in addition to final decision, both the knowl- science. High-throughput and DFT calculation can provide
edge that learns from the data and the features that attributes abundant data without the cost of human resources in experi-
most to the results should be presented in detail. Therefore, the ment, so that more theoretical data could be directly produced
interpretability of the machine learning model is also the focus in a limited time.[57,82] At the same time, there are many open
of attention in AI for materials science. high-throughput material databases nowadays, such as the
open quantum material database,[83] open inorganic material
database,[84] crystallographic open database,[85] thermoelectric
2.5. The Highly Specialized Research in Materials Science open data resource,[86] two-dimensional material database,[87]
novel material discovery database,[88] high-throughput com-
With the trend of AI technology, many researchers begin to bination database of electronic band structure for inorganic
combine the materials science with machine learning intelli- scintillator materials,[89] inorganic amorphous database,[90] and
gent technology, relying on the data-driven strategy to predict so on. All these open databases could provide much more con-
and analyze material properties.[75–77] As known to all, the code veniences for researchers in machine learning. Moreover, Dong
programs should be firstly designed to achieve the goals for et al. develop a deep neural network (DNN) to predict mate-
machine learning, which includes both the feature engineering rial defects based on small data, which also achieve very good
and model building after data collection. Many programming results.[91] By above strategies, the open-source of database for
languages can be used for machine learning, such as the most materials science has gradually increased for machine learning
popular one “python” with simple but rich expansion package. in recent years, as well as the types of databases. This could
However, due to the lack of deep mathematical theory and pro- provide more choices for machine learning research in mate-
gramming ability for researchers in materials science, the pro- rials and promote more researchers in different fields to partici-
gress of machine learning in materials still going slowly. pate. However, the collection efficiency for databases still needs
to be further improved, and more strategies should be devel-
oped. Meanwhile, optimizing the algorithm model for large
3. Strategies for Machine Learning Applied sample data to promote the improvement of machine learning
capabilities and making more breakthroughs for small sample
in Materials Science
data to adapt machine learning are still the main challenges.
3.1. Increasing the Amount of Data

Since the foundation in 2011, the Materials Genome Initiative 3.2. Dealing with the Perplexity of Research in Features
(MGI) continually accelerates the pace of material discovery, by Machine Learning
design and deployment by training to collaborate experi-
ment, theory and computing for data generation, analysis As shown in Figure 3, the way to deal with the perplexity in
and sharing.[78] Zhang et al. combined the MGI strategy with features could be concluded in feature selection and dimen-
high-throughput measurement and CALPHAD software to sionality reduction with a subsequent machine learning. In
accelerate the development and design of new biological tita- addition, the feature contact information mining, classification,
nium alloys.[79] MGI strategy is used to enhance the mining of and reconstruction based on machine learning are also sum-
insensitive high energy density materials by Zhang et al.[80] The marized herein.
authors discussed how a materials genome approach could be
used to accelerate the discovery of promising insensitive high
explosive (IHE) molecules. By rationalizing the relationship 3.2.1. Strengthening the Analysis and Optimization on Features
between structure and properties, the “genetic” features are
firstly identified and extracted. Then, the computation-guided The methods by selecting and transforming multi-dimensional
molecular designing and rapid screening are carried out, features into low-dimensional features have been proposed to
which includes library construction of molecular fragments improve the quality of features, reduce computational com-
and structural selection of candidate molecules based on plexity, and improve recognition accuracy.[92–98] In present
filter conditions of “genetic” features. Finally, the ideal target machine learning feature engineering, these two powerful tech-
molecule, 2,4,6-triamino-5-nitropyrimidine-1,3-dioxide, is obtained nologies have shown great potential in various applications,

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (4 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Figure 3. The strategies to deal with perplexity of research in features.

such as the text data,[99] planetary system metrics,[100] clinical combination of paired radial, nearest neighbor, bond angle,
medicine,[101] human hand motion classification,[102] human dihedral angle and core charge distribution plays a vital role in
activity recognition,[103] etc. In materials science, the trans- predicting formation energy, bandgap, static refractive index,
formation in various features is also important, which might and magnetic properties via this similar method.[108] In mate-
directly affect the analytical accuracy in structure and proper- rial research, the irrelevant and redundant features are also
ties. The familiar machine learning algorithms, e.g. support eliminated by feature selection in most cases, to improve the
vector machine and clustering,[104,105] have exhibited some solu- accuracy of the model and reduce the running time. Mean-
tions for feature selection. For example, Brgoch et al. selects while, the real relevant feature simplification model could be
and classifies the features of structure and polymorphism of selected to help understanding the process of data generation.
equiatomic ternary ABC phase based on these two methods.[105] In the publication reported by Adibi, the dimensionality reduc-
In predicting the performances of zeolites, Evans’s group con- tion technique has been used to reduce the dimension of the
structs a machine learning method to automatically predict response space while keep the number of identical points in
the mechanical properties without any calculation of chemical the response space for electromagnetic nanostructures design
properties, just based on their local geometry, structure, and and optimization.[109] Thus the dimension of these features are
porosity features.[106] According to the exceptional insight into reduced based on the correlation between them, which could
the mechanics of zeolitic frameworks, the development of this reduce the computational complexity and avoid more mistakes.
model has highlighted the correlations between characteristic
features and elastic properties, which is further trained with
a DFT data set. This methodology is employed to predict the 3.2.2. Exploring the Relationship among Features
elastic response of 590 448 hypothetical zeolites, which are
more accurate than those calculated by the force field method. In present research of machine learning in materials science,
Besides, some methods can not only find the minimum of the built-in methods are often used to explore the relationship
the surface free energy, but also screen out the special struc- among different features. Artrith and co-workers automati-
ture to obtain these properties, develop novel feature classifica- cally construct atomic interaction potentials by a Behler-Par-
tion methods, and propose complete structure descriptors,[107] rinello approach, which is based on artificial neural networks
which could overcome the problems that can not be realized (ANNs).[110] They try to establish a free and open-source atomic
by common descriptors. Choudhary et al have proved that the energy network package for modeling electrocatalytic CuAu and

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (5 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Cu-doped ceria nanoparticles in water. Taking the well-studied a single given target, which firstly encodes the initial micro-
TiO2 as the example, the construction of realistic structure structure, and then pre-train through the convolution neural
models is optimized in consideration of defected crystals and network to further obtain the optimized microstructure. After
nano structures for catalytic activity. Thus, the lattice parame- unsupervised learning, the required reconstructed microstruc-
ters, energies, and bulk moduli are accurately predicted by TiO2 ture is finally obtained.
ANN potential. The researchers also propose a generalized Looking at the above strategies at this stage, various feature
crystal graph convolution neural network framework to present engineering methods are usually used to deal with the fea-
the periodic crystal system at the example of perovskite.[42] By ture challenges. Automatic selection of features based on deep
predicting the properties according to the relationship between learning and automatic coding could make the feature pro-
atoms in crystal, a theoretically optimal crystal material could cessing much easier. However, these are still in the static space
be obtained. The neural network potential and Gaussian stage, and the relationship among the material phase, composi-
approximate potential have also been used to simulate the tion, morphology, and performances still cannot be well studied
interatomic potential.[29] After unifying the different parameter in parallel at the same time. Due to the particularity, variability,
factors into a free energy by the graph neural network model, polymorphism, and uncertain factors of materials science,
the relationship between the features could be automatically more attention should be paid to obtain the real relationships
linked together associating with the information mining pro- in dynamic space in the future. For data processing methods,
cess, which finally help to predict performances.[111] Further- the internal response of data preprocessing to material proper-
more, machine learning has also shown significant value in ties and their intersection should also be considered with more
constructing the correlation between physical features in phase influencing factors.
change material system.[112]

3.3. Developing Professional Algorithms for Materials Science

3.2.3. Classifying Features into Different Categories
In order to fully solve the challenges that machine learning algo-
Machine learning has also been used for data classification to rithms are not universal to materials science, many researchers
provide more convenience and improve the learning efficiency. draw lessons from the successful cases or empirical analysis in
In ref. [31], Sohn et al. reported a deep machine learning tech- similar fields and choose the appropriate algorithm model to
nique based on convolution neural network, which was used efficiently optimize the material object. As shown in Figure 4,
to classify powder X-ray diffraction (XRD) patterns according to the currently popular machine learning algorithms are sum-
crystal system, extinction group, and spatial group. The XRD marized, which could be divided into four major directions:
data are gained from the inorganic crystal structure database. classical learning,[116] reinforcement learning,[117–119] ensemble
Meanwhile, convolution neural network is also used to classify learning,[120,121] neural network and deep learning.[122]
the crystal structure automatically by Ziletti et al.[32] The crystal These four kinds of machine learning algorithms have
structures are firstly transformed into diffraction patterns, then been widely concerned by researchers in materials science. As
small amounts of data are extracted as the training set to gen- shown in Table 1, ten classical algorithms are applied for mate-
erate a final classification model. Besides, visualization is also rials information mining with huge application potential in
developed in the neural network’s internal operation to ensure machine learning.[116] Among them, the decision tree and SVM
the classification decision. Finally, through the deployment of model are the main classification algorithms. The advantage of
crystal structure classification model, the highly efficient clas- decision tree model is that no domain knowledge or parameter
sification is realized without any additional model optimiza- setting is needed in the construction process, so that the deci-
tion. These studies would pave the way to optimize the noise sion tree is more suitable for knowledge discovery in mate-
and incomplete three-dimensional structure data for machine rial research. While the SVM could avoid over-fitting to some
learning in materials science. extent, since does it only depends on support vectors. Thus,
SVM is still effective in medium to small samples. Riley et al.
reported a Molecule Deep Q-Networks model for molecular
3.2.4. Recognizing the Attributes of Material Features optimization by combining chemical knowledge and reinforce-
ment learning,[123] which exhibit advantages including higher
Because of the particularity of material data and features, the sampling efficiency and better molecules exploitation. Sparks
data was always encoded and transformed to be easily identi- et al. simulate the experimental data effectively by an ensemble
fied by machine learning.[34,113] For example, Dimiduk et al. learning method by merging different data sources into the
simplified the material topics into digital fingerprint vectors, modeling of sparsely represented experimental data. In the case
and then developed appropriate measure of chemical (dis)simi- of bandgap prediction, the root mean square error could be
larity or chemical distance in the learning scheme to map the reduced by over 9%.[57] Yu et al. reported a decision tree-based
features’ distance.[114] The principle is to encode and map the ensemble learning model to accurately predict the material
attributes to form the basis of further operation. Besides, Chen removal rate for chemical mechanical planarization.[124]
and co-workers also propose a training-saving method based Apart from the three categories mentioned above, the neural
on transfer learning, combining the encoder-decoder process network and deep learning are currently the focus of most
via deep convolution network and feature matching optimi- attention. In 2006, Hinton et al. alleviated the local optimal
zation.[115] This method mainly aims at the microstructure of solution problem by using a pre-training method, which pushed

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (6 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Figure 4. The classification for machine learning algorithms.

the hidden layer to 7 layers.[149] After that, the neural network established a highly accurate materials property prediction
become the real depth learning, such as the subsequent Deep model by leveraging the large DFT-computational data sets,
Belief Nets,[150] Convolutional Neural Network,[151–153] and smaller DFT-computed data sets, and available experimental
Recurrent neural network.[154–156] observations. Besides, the neural networks and deep learning
The rapid development of deep learning has also solved methods have also demonstrated their strong ability in many
many thorny problems in materials science. For example, Rao other material research.
and Liu reported a three-dimensional depth convolution neural Although the diversity of machine learning algorithms could
network to predict the anisotropic effective material proper- provide much selectivity, the algorithms applicable to materials
ties for representative volume elements (RVEs) with random science are still limited since it is obviously different from the
inclusions.[157] A radial basis function artificial neural network time series and context in the text. Moreover, it also contains
model is developed by Ince et al. to predict the propagation and many indeterminate factors, such as ion valence and condi-
growth behavior of fatigue crack.[158] In ref. [48], Agrawal et al. tional chemical reaction, which need more hyperparameters

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (7 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Table 1. Ten common algorithms of machine learning for application in materials science.

Algorithm type Algorithm Model Examples in Materials Science

[125] Analysis of the causes of Coffee defects by decision Tree[126]
Classification algorithm C4.5
[127]
Naive Bayes Classification of metal binders[128]
SVM[129] Material monitoring and defect diagnosis[130]
Prediction of rock brittleness[131]
KNN[132] Prediction of process parameters of reinforced metal casting[133]
Analysis of welding modeling of different materials[134]
Adaboost[135] Temperature compensation of Silicon Piezoresistive pressure Sensor[136]
Cart[137] Differential diagnosis of mucosanase[138]
Clustering algorithm K-Means[139] Structural texture similarity recognition of materials[140]
Establishment of parametric homogenized crystal plasticity model of single crystal Ni-base superalloy[141]
EM[142] Estimation of dose distribution from positron emitter distribution combined with filtering[143]
Correlation distribution Apriori[144] Identify the frequency trajectory of material transportation[145]
[146]
Connection analysis PageRank Measurement of hyperelastic materials[147]
Remote protein homology detection[148]

to be set for algorithm model. At present, most materials also a problem, which may lead to the omission of the internal
researchers rely on the existing algorithms, traditional machine influence factors of deep learning on materials research in the
learning models, and simple neural networks. In the future, future.
more suitable algorithm models should be developed in actual
learning process for materials science.
3.5. Developing Open-Source Material Packages and Machine
Learning Frameworks
3.4. Optimizing and Evaluating the Learning Models
To promote the application of machine learning in materials
The most straightforward interpretability for machine learning science, many open-source machine learning tools have been
is using the interpretable models, such as logical regression, provided. For example, SchNet is a deep learning framework
linear model, and decision tree.[159] Ribeiro et al. reported the for molecular materials, which is specifically designed to
LIME explanation technique to explain the predictions of any model atomic systems by using continuous filter convolution
classifier in an interpretable and faithful manner by learning layers.[167] DScribe is a software package for machine learning
an interpretable model.[160] Besides, by transforming the task that could provide widespread feature transformation for
framework into the sub-module optimization, it could help to atomic material simulations.[168] Matminer is an open-source
decide if one should trust a prediction, choose between dif- software based on Python, which can analyze and predict
ferent models, improve an untrustworthy classifier, and iden- material properties by the data-driven method.[169] Pymatgen
tify why a classifier should not be trusted. Murdoch et al. tried is a robust open-source Python Materials Genomics library for
to address these concerns by defining interpretability in the material analysis,[170] which could the initial setup and original
context of machine learning, and the Predictive, Descriptive, calculated data for high-throughput computing materials sci-
Relevant framework are also introduced for interpretations.[161] ence. COMBO designs an effective Bayesian optimization
Yuma Iwasaki and co-workers predicted the spin-driven ther- scheme, which is also an as an open-source python library by
moelectric materials with anomalous Nernst effect by an inter- combining Thompson sampling, random feature graph, first-
pretable machine learning method called factorized asymptotic order Cholesky update and automatic hyperparameter adjust-
Bayesian inference hierarchical mixture of experts (FAB/ ment.[171] COMBO is available on the website (https://github.
HMEs), according to the prior knowledge of material science com/tsudalab/combo). AFLOW-ML provides an open REST-
and physics. After practical material synthesis, a new type of fulAPI with direct access to continuously updated algorithms,
spin-driven thermoelectric material is finally demonstrated.[162] which can be transparently integrated into any workflow, such
Moreover, Kailkhura et al. proposed a new evaluation metric as electronic retrieval and thermal or mechanical performances
and a trust score to better quantify the confidence in the pre- prediction.[172] In addition to the open-source packages men-
dictions, associating with a rationale generator component to tioned above, the open-source framework including Tensor-
provide both model-level and decision-level explanations. As a Flow,[173] Pytorch,[174] Keras,[175] Scikit-Learn[176] can also reduce
result, the properties of crystalline compounds and potentially the time of building models from the beginning and improve
stable solar cell materials could be primely predicted.[163] And work efficiency.[177] For example, the python library-based Scikit-
other studies on interpretable machine learning[164–166] can also Learn is used to predict thermal stability of perovskite materials
provide reference to machine learning for principal research by machine learning, feature selection, and model evaluation.[44]
in materials. However, the research of machine learning in We believe that these types can further accelerate the adoption
materials science is still insufficient, and the interpretability is of machine learning methods in material development, if the

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (8 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

open-source material packages and machine learning frame- scientific research progress in materials science. Thus, it is
works could be effectively connected by the cloud-based inter- believed that notable advances still need to be developed to
connected applications. At present, in the research of machine meet the requirements of practical applications. Herein, as
learning in materials science, the materials-related open-source shown in Figure 5, we outline several possible directions for
toolkits and programming language frameworks have been machine learning and hope that these perspectives might be
well designed by programming tools, which can provide great useful for researchers in materials science.
convenience for non-professional programming researchers,
such as materials researchers. However, more development in
open-source toolkits are still needed in all directions of mate- 4.1. Machine Learning with Small Sample Size
rials science to promote the further popularization of feature
engineering and machine learning. In the process of material research, it will be a huge step if
the studies could be impelled by a small amount of data. For
example, the DNN regression has been used to predict defects
4. Summary and Perspectives and solidification cracking susceptibility of stainless steels based
on 487 data points.[91] In future study, DNN could also be trained
After the above discussion, we fully believe that machine and tuned to convert the scattered small dataset into high-pre-
learning exhibits huge ability to mine new materials in the cision maps in high-dimensional chemical, processing para-
data driven era. Based on the challenges of machine learning meter, and performances space. In addition, in order to learn
in materials, the current solutions and research progress have from the supervised information of limited samples, zero shot
been summarized and discussed. Although these pioneering learning (ZSL),[178] one shot learning (OSL),[179] and few shot
studies have been conducted to promote the machine learning learning (FSL)[180] would be mostly potential research objects
in materials science, novel algorithm model technologies, effi- for small sample learning because of the learning based on a
cient data preprocessing methods, information mining, mate- small sample size. The typical example in character generation
rial structure and performance prediction, and new function indicates that a computer program is required to break the char-
prediction, as well as the interaction among these features, acters into smaller parts, which can be transferred among each
should be further prospected to improve work efficiency and character, and then aggregate these smaller components into

Figure 5. The perspective of machine learning in materials science.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (9 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

new characters.[181] We think this effective strategy could also preparation parameters, etc. The limited correlation between
be similar for potential application in material detection, defect different samples and the lack of unified standard contribute
prediction, classification and so on. Accordingly, the major prob- to a poor comparability. Thus, the exploration on the relevance
lems of machine learning in materials science will be solved, and similarity between features is still very important in mate-
such as learning like human beings, learning rare situations, rials science. We can try to establish the relationship between
reducing the sample collection and computational costs. nodes and edges by using graph neural network to explore the
relationship between the child nodes and some others.[184] The
materials can be divided into different categories, and each cat-
4.2. Constructing New Features by Combining Multiple Features egory has many different research directions, which could have
many similarities such as time, space, structure, and perfor-
Resembling the deep learning of graphics and audio synchro- mance. Furthermore, the transfer learning to transfer knowl-
nous recognition in the time dimension, the multi-feature edge from the source domain rich in training data to the target
combination effects among phase, structure, performance, and domain lacking in training data or clustering in graph embed-
application based on time and space dimensions should also ding might also be effective for seeking similarity.
be developed during the whole life cycle of material prediction,
material synthesis, material service, recycling, and reusing.[182]
If the number of size parameters is too large, the common 4.5. Presetting Experimental Variable Parameters
matrix decomposition can be used as fair idea for reference.
The combination of multi-features can not only expand the There are always many uncertain parameters in material
amount of sample data, but also improve the fitting ability of research, such as temperature, humidity, time and so on, which
the complex relationship of experimental features. Here are would affect the phase composition and morphology of the final
our recommendations for multi-features combination. Firstly, materials. Traditional machine learning also has limitations in
the initial features should be linearly combined to form new dealing with the related features. Therefore, it is necessary to
features, and the influence of combined features on the whole provide pre-parameters for these variable parameters, so that a
research process should be explored. In addition, the initial fea- conditional database should be established to meet the require-
tures could also be disassembled into new features, and then ments. In addition, we believe that we can consider strength-
add them into other features to form new features. ening the generative adversarial network,[185] transformer[186] or
other ideas in the process of deep learning, which can be used
as references to deal with the variable parameter.
4.3. Controlling Data to Ensure Feature Balance

Due to the inevitable data bias of the material experiment and 4.6. Developing Novel Learning Models
the neglect of the collection method, the problem of data imbal-
ance always occurs in the process of data collection which would Based on the material prediction according to the theoretical
further lead to the situation of judgment confusion. Herein, we calculation, several new laws would be found by machine
try to provide some suggestions to solve the unbalanced data. learning. And novel models, algorithms or integrated algo-
The first one is data sampling, which could be divided into up- rithms for material computing could be developed based on the
sampling and down-sampling. Up-sampling is the process of paradigm named “theory-oriented, data-driven, and parameter
duplicating a small amount of data to equalize the proportion learning optimization”, which could help to mine unknown
of each category, in which adding random disturbance into the new theories, develop new mechanisms, and in turn promote
newly generated data is necessary to overcome the overfitting. the progress of machine learning in materials science. For
While the down-sampling is to select a small part from most cat- example, it is suggested to develop novel learning models based
egories to keep the data proportion of each category at a normal on the material genome initiative via machine learning, to
level, in which multiple random samples are required to ensure verify them in combination with the experimental results, and
the integrity of information. The second one is data combina- finally to optimize the genome initiative for materials science.
tion, which could generate more samples by using the similar In conclusion, we believe that machine learning is emerging
features of existing samples. The third one is weighting, which as a great strategy for materials science, which would speed up
refers to applying different weights to each samples according to the process of material exploration. The day, when machine
the corresponding errors, so that machine learning could focus learning is maturely applied to materials science and tech-
more on fewer and error-prone samples. And finally, one clas- nology, would be the time for opening a new chapter of human
sification, such as One-class SVM,[183] is also a very key strategy civilization. Therefore, summarizing all possible strategies and
for controlling data to ensure feature balance, when the propor- providing in-depth perspectives will be necessary to make the
tion of positive and negative samples is out of balance. machine learning viable in the future.

4.4. Discovering the Relevance and Similarity between Features

Acknowledgements
In materials science, many factors could affect the features of This work was supported by the National Key R&D Program of China
labeled training samples, such as different preparation methods, (No. 2018YFC1901500).

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (10 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

Conflict of Interest [23] C. Kim, G. Pilania, R. Ramprasad, J. Phys. Chem. C 2016, 120,
14575.
The authors declare no conflict of interest. [24] K. T. Schütt, F. Arbabzadah, S. Chmiela, K. R. Müller,
A. Tkatchenko, Nat. Commun. 2017, 8, 13890.
[25] S. A. Tawfik, O. Isayev, M. J. S. Spencer, D. A. Winkler, Adv. Theory
Simul. 2020, 3, 1900208.
Keywords [26] J. Schmidt, J. Shi, P. Borlido, L. Chen, S. Botti, M. A. L. Marques,
data-driven, machine learning, materials discovery, materials science, Chem. Mater. 2017, 29, 5090.
performances prediction, perspective [27] Y.-J. Wu, L. Fang, Y. Xu, npj Comput. Mater. 2019, 5, 56.
[28] P. V. Balachandran, B. Kowalski, A. Sehirlioglu, T. Lookman, Nat.
Received: August 12, 2021 Commun. 2018, 9, 1668.
Revised: September 15, 2021 [29] G. C. Sosso, V. L. Deringer, S. R. Elliott, G. Csányi, Mol. Simul.
Published online: October 4, 2021 2018, 44, 866.
[30] J.-P. Correa-Baena, K. Hippalgaonkar, J. Van Duren, S. Jaffer,
V. R. Chandrasekhar, V. Stevanovic, C. Wadia, S. Guha,
T. Buonassisi, Joule 2018, 2, 1410.
[1] A. Esteva, K. Chou, S. Yeung, N. Naik, A. Madani, A. Mottaghi, [31] W. B. Park, J. Chung, J. Jung, K. Sohn, S. P. Singh, M. Pyo, N. Shin,
Y. Liu, E. Topol, J. Dean, R. Socher, NPJ Dig. Med. 2021, 4, 5. K.-S. Sohn, IUCrJ 2017, 4, 486.
[2] O. Yakovenko, I. Bondarenko, Soc. Networks Texts 2021, 1357, 115. [32] A. Ziletti, D. Kumar, M. Scheffler, L. M. Ghiringhelli, Nat. Commun.
[3] S. X. Tang, R. Kriz, S. Cho, S. J. Park, J. Harowitz, R. E. Gur, 2018, 9, 2775.
M. T. Bhati, D. H. Wolf, J. Sedoc, M. Y. Liberman, npj Schizophr. [33] B. Meredig, E. Antono, C. Church, M. Hutchinson, J. Ling,
2021, 7, 25. S. Paradiso, B. Blaiszik, I. Foster, B. Gibbons, J. Hattrick-Simpers,
[4] J. W. Crandall, M. Oudah, Tennom, F. Ishowo-Oloko, S. Abdallah, A. Mehta, L. Ward, Mol. Syst. Des. Eng. 2018, 3, 819.
J.-F. Bonnefon, M. Cebrian, A. Shariff, M. A. Goodrich, I. Rahwan, [34] G. Pilania, X.-Y. Liu, J. Mater. Sci. 2018, 53, 6652.
Nat. Commun. 2018, 9, 233. [35] Y. Liu, B. Guo, X. Zou, Y. Li, S. Shi, Energy Storage Mater. 2020, 31,
[5] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, 434.
A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap, [36] T. Lookman, P. V. Balachandran, D. Xue, R. Yuan, npj Comput.
F. Hui, L. Sifre, G. Van Den Driessche, T. Graepel, D. Hassabis, Mater. 2019, 5, 21.
Nature 2017, 550, 354. [37] P V. Balachandran, Comput. Mater. Sci. 2019, 164, 82.
[6] A. Zhavoronkov, Y. A. Ivanenkov, A. Aliper, M. S. Veselov, [38] P. Avery, X. Wang, C. Oses, E. Gossett, D. M. Proserpio, C. Toher,
V. A. Aladinskiy, A. V. Aladinskaya, V. A. Terentiev, D. A. Polykovskiy, S. Curtarolo, E. Zurek, npj Comput. Mater. 2019, 5, 89.
M. D. Kuznetsov, A. Asadulaev, Y. Volkov, A. Zholus, [39] X. Zhai, M. Chen, W. Lu, Comput. Mater. Sci. 2018, 151, 41.
R. R. Shayakhmetov, A. Zhebrak, L. I. Minaeva, B. A. Zagribelnyy, [40] E. V. Podryabinkin, E. V. Tikhonov, A. V. Shapeev, A. R. Oganov,
L. H. Lee, R. Soll, D. Madge, L. Xing, T. Guo, A. Aspuru-Guzik, Phys. Rev. B 2019, 99, 064114.
Nat. Biotechnol. 2019, 37, 1038. [41] S. Kim, J. Noh, G. Ho Gu, A. Aspuru-Guzik, Y. Jung, ACS Cent. Sci.
[7] B. Burger, P. M. Maffettone, V. V. Gusev, C. M. Aitchison, Y. Bai, 2020, 6, 1412.
X. Wang, X. Li, B. M. Alston, B. Li, R. Clowes, N. Rankin, B. Harris, [42] T. Xie, J. C. Grossman, Phys. Rev. Lett. 2018, 120, 145301.
R. S. Sprick, A. I. Cooper, Nature 2020, 583, 237. [43] W. Tong, Q. Wei, H.-Y. Yan, M.-G. Zhang, X.-M. Zhu, Front. Phys.
[8] J. Launchbury, Retrieved November 2017, 11, 2019. 2020, 15, 63501.
[9] Y. Lecun, Y. Bengio, G. Hinton, Nature 2015, 521, 436. [44] W. Li, R. Jacobs, D. Morgan, Comput. Mater. Sci. 2018, 150,
[10] R. S. Sutton, A. G. Barto, Reinforcement Learning: An Introduction, 454.
2nd ed., The MIT Press, Almere 2018. [45] H. Wei, S. Zhao, Q. Rong, H. Bao, Int. J. Heat Mass Transfer 2018,
[11] S. Lloyd, C. Weedbrook, Phys. Rev. Lett. 2018, 121, 040502. 127, 908.
[12] S. L. Sass, The Substance of Civilization: Materials and Human His- [46] F. Musil, S. De, J. Yang, J. E. Campbell, G. M. Day, M. Ceriotti,
tory from the Stone Age to the Age of Silicon, Arcade Publishing, Chem. Sci. 2018, 9, 1289.
New York 1998. [47] D. Padula, J. D. Simpson, A. Troisi, Mater. Horiz. 2019, 6, 343.
[13] P. S. Anton, R. Silberglitt, J. Schneider, The Global Technology Revo- [48] D. Jha, K. Choudhary, F. Tavazza, W.-K. Liao, A. Choudhary,
lution: Bio/Nano/Materials Trends and Their Synergies with Informa- C. Campbell, A. Agrawal, Nat. Commun. 2019, 10, 5316.
tion Technology by 2015, Rand Corporation, Santa Monica 2001. [49] A. Mangal, E. A. Holm, Int. J. Plast. 2019, 114, 1.
[14] C. Freeman, F. Louçã, As Time Goes By: From the Industrial Revo- [50] S. K. Kauwe, J. Graser, A. Vazquez, T. D. Sparks, Integr. Mater.
lutions to the Information Revolution, Oxford University Press, Manuf. Innov. 2018, 7, 43.
New York 2001. [51] Y. Zhuo, A. Mansouri Tehrani, J. Brgoch, J. Phys. Chem. Lett. 2018,
[15] K. Schwab, The Fourth Industrial Revolution, Currency Books, New 9, 1668.
York 2017. [52] S. Lu, Q. Zhou, Y. Ouyang, Y. Guo, Q. Li, J. Wang, Nat. Commun.
[16] A. Agrawal, A. Choudhary, APL Mater. 2016, 4, 053208. 2018, 9, 3405.
[17] R. Ramprasad, R. Batra, G. Pilania, A. Mannodi-Kanakkithodi, [53] W. Ye, C. Chen, Z. Wang, l.-H. Chu, S. P. Ong, Nat. Commun. 2018,
C. Kim, npj Comput. Mater. 2017, 3, 54. 9, 800.
[18] K T. Butler, D W. Davies, H. Cartwright, O. Isayev, A. Walsh, Nature [54] P. B. Jørgensen, K. W. Jacobsen, M. N. Schmidt, Neural message
2018, 559, 547. passing with edge updates for predicting properties of molecules and
[19] J. Schmidt, M R. G. Marques, S. Botti, M A. L. Marques, npj materials. arXiv preprint arXiv:1806.03146, 2018.
Comput. Mater. 2019, 5, 83. [55] A. Jain, K A. Persson, G. Ceder, APL Mater. 2016, 4, 053102.
[20] R. Batra, L. Song, R. Ramprasad, Nat. Rev. Mater. 2020, 6, 655. [56] J. Lee, A. Seko, K. Shitara, K. Nakayama, I. Tanaka, Phys. Rev. B
[21] B. Zhang, X.-Q. Zheng, T.-Y. Zhao, F.-X. Hu, J.-R. Sun, B.-G. Shen, 2016, 93, 115104.
Chin. Phys. B 2018, 27, 067503. [57] S K. Kauwe, T. Welker, T D. Sparks, Integr. Mater. Manuf. Innov.
[22] L. Weston, C. Stampfl, Phys. Rev. Mater. 2018, 2, 085407. 2020, 9, 213.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (11 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

[58] J. Schmidt, L. Chen, S. Botti, M A. L. Marques, J. Chem. Phys. 2018, [85] S. Gražulis, A. Daškevič, A. Merkys, D. Chateigner, L. Lutterotti,
148, 241728. M. Quirós, N R. Serebryanaya, P. Moeck, R T. Downs, A. Le Bail,
[59] X. Zhu, J. Yan, M. Gu, T. Liu, Y. Dai, Y. Gu, Y. Li, J. Phys. Chem. Lett. Nucleic Acids Res. 2012, 40, D420.
2019, 10, 7760. [86] P. Gorai, D. Gao, B. Ortiz, S. Miller, S A. Barnett, T. Mason, Q. Lv,
[60] M. Misawa, S. Fukushima, A. Koura, K. Shimamura, F. Shimojo, V. Stevanović, E S. Toberer, Comput. Mater. Sci. 2016, 112, 368.
S. Tiwari, K.-.I. Nomura, R K. Kalia, A. Nakano, P. Vashishta, J. [87] S. Haastrup, M. Strange, M. Pandey, T. Deilmann, P. S. Schmidt,
Phys. Chem. Lett. 2020, 11, 4536. N. F. Hinsche, M. N. Gjerding, D. Torelli, P. M. Larsen, A. C. Riis-
[61] T. Toyao, K. Suzuki, S. Kikuchi, S. Takakusagi, K.-I. Shimizu, Jensen, J. Gath, K. W. Jacobsen, J. Jørgen Mortensen, T. Olsen,
I. Takigawa, J. Phys. Chem. C 2018, 122, 8315. K. S. Thygesen, 2D Mater. 2018, 5, 042002.
[62] P. Raccuglia, K. C. Elbert, P. D. F. Adler, C. Falk, M. B. Wenny, [88] C. Draxl, M. Scheffler, MRS Bull. 2018, 43, 676.
A. Mollo, M. Zeller, S A. Friedler, J. Schrier, A. J. Norquist, Nature [89] W. Setyawan, R M. Gaume, S. Lam, R S. Feigelson, S. Curtarolo,
2016, 533, 73. ACS Comb. Sci. 2011, 13, 382.
[63] C. Draxl, Stepping Stones towards the Fourth Paradigm of Mate- [90] M. Hellenbrandt, Crystallogr. Rev. 2004, 10, 17.
rials Science. in Seminar, Research Training Group QM3, Universität [91] S. Feng, H. Zhou, H. Dong, Mater. Des. 2019, 162, 300.
Bremen. [92] C. O. S. Sorzano, J. Vargas, A. P. Montano, A Survey of Dimension-
[64] X. Zhang, Polymer 2001, 42, 8179. ality Reduction Techniques. arXiv preprint arXiv:1403.2877, 2014.
[65] P. H. C. Camargo, K. G. Satyanarayana, F. Wypych, Mater. Res. [93] M. Taradeh, M. Mafarja, A. A. Heidari, H. Faris, I. Aljarah,
2009, 12, 1. S. Mirjalili, H. Fujita, Inf. Sci. 2019, 497, 219.
[66] G. Krauss, Steels: Processing, Structure, and Performance, ASM [94] K. Yu, X. Guo, L. Liu, J. Li, H. Wang, Z. Ling, X. Wu, ACM Comput.
International, Almere, 2015. Surveys (CSUR) 2020, 53, 111.
[67] S. Curtarolo, G. L. W. Hart, M. B. Nardelli, N. Mingo, S. Sanvito, [95] Y. Liang, D. Niu, W.-C. Hong, Energy 2019, 166, 653.
O. Levy, Nat. Mater. 2013, 12, 191. [96] Z. Liu, Z. Lai, W. Ou, K. Zhang, R. Zheng, Signal Process. 2020, 170,
[68] J. J. De Pablo, N. E. Jackson, M. A. Webb, L.-.Q. Chen, J. E. Moore, 107456.
D. Morgan, R. Jacobs, T. Pollock, D. G. Schlom, E. S. Toberer, [97] M. Nixon, A. Aguado, Feature Extraction and Image Processing for
J. Analytis, I. Dabo, D. M. Delongchamp, G. A. Fiete, G. M. Grason, Computer Vision, Academic press, Waltham 2019.
G. Hautier, Y. Mo, K. Rajan, E. J. Reed, E. Rodriguez, V. Stevanovic, [98] X. Huang, L. Wu, Y. Ye, Int. J. Pattern Recognition Artif. Intell. 2019,
J. Suntivich, K. Thornton, J.-C. Zhao, npj Comput. Mater. 2019, 5, 33, 1950017.
41. [99] A. Karami, Int. J. Knowledge Engineering Data Mining 2019, 6, 289.
[69] S. Rana, R. Fangueiro, Advanced Composite Materials for Aerospace [100] Y. Alibert, Astron. Astrophys. 2019, 624, A45.
Engineering: Processing, Properties and Applications, Woodhead Pub- [101] M. Modaresnezhad, A. Vahdati, H. Nemati, A. Ardestani, F. Sadri,
lishing, Cambridge 2016. Comput. Biol. Med. 2019, 106, 84.
[70] J. K. Pandey, K. R. Reddy, A. K. Mohanty, M. Misra, Handbook of [102] N. Rabin, M. Kahlon, S. Malayev, A. Ratnovsky, Exp. Syst. Appl.
Polymernanocomposites: Processing, Performance and Application, 2020, 149, 113281.
Springer, Myrtle Beach 2013. [103] I. El Moudden, M. Ouzir, B. Benyacoub, S. El Bernoussi, Contemp.
[71] W. D. Callister, D. G. Dhavalikar, Materials Science and Engineering: Eng. Sci. 2016, 9, 1031.
An Introduction, Wiley, Hoboken, NJ 2021. [104] A. O. Oliynyk, L. A. Adutwum, J. J. Harynuk, A. Mar, Chem. Mater.
[72] R. Kulkarni, S. Dhavalikar, S. Bangar, Traffic Light Detection and 2016, 28, 6672.
Recognition for Self Driving Cars Using Deep Learning. in 2018 Fourth [105] A. O. Oliynyk, L. A. Adutwum, B. W. Rudyk, H. Pisavadia, S. Lotfi,
International Conference on Computing Communication Control and V. Hlukhyy, J. J. Harynuk, A. Mar, J. Brgoch, J. Am. Chem. Soc. 2017,
Automation (ICCUBEA), 2018, IEEE. 139, 17870.
[73] R. Wu, et al., Deep image: Scaling up image recognition. arXiv pre- [106] J. D. Evans, F.-X. Coudert, Chem. Mater. 2017, 29, 7833.
print arXiv:1501.02876, 2015, 7. [107] A. Mansouri Tehrani, A. O. Oliynyk, M. Parry, Z. Rizvi, S. Couper,
[74] I. Serban, et al. Building end-to-end dialogue systems using generative F. Lin, L. Miyagi, T. D. Sparks, J. Brgoch, J. Am. Chem. Soc. 2018,
hierarchical neural network models, in Proceedings of the AAAI Con- 140, 9844.
ference on Artificial Intelligence, 2016. [108] K. Choudhary, B. Decost, F. Tavazza, Physs Rev. Mater. 2018, 2,
[75] L. Ward, A. Agrawal, A. Choudhary, C. Wolverton, npj Comput. 083801.
Mater. 2016, 2, 16028. [109] Y. Kiarnejad, S. Abdollahramezani, A. Adibi, npj Computat. Mater.
[76] J. Im, S. Lee, T.-W. Ko, H. W. Kim, Y. Hyon, H. Chang, npj Comput. 2020, 6, 12.
Mater. 2019, 5, 37. [110] N. Artrith, A. Urban, Comput. Mater. Sci. 2016, 114, 135.
[77] T.-T. Le, J. Compos. Mater. 2021, 55, 787. [111] C. Chen, W. Ye, Y. Zuo, C. Zheng, S. P. Ong, Chem. Mater. 2019, 31,
[78] N. Science, T. Council, Materials genome initiative for global com- 3564.
petitiveness, 2011, Executive Office of the President, National Sci- [112] M. Attarian Shandiz, R. Gauvin, Comput. Mater. Sci. 2016, 117, 270.
ence and Technology Council. [113] T. Lam Pham, H. Kino, K. Terakura, T. Miyake, K. Tsuda,
[79] J. Ling, W. Chen, Y. Sheng, W. Li, L. Zhang, Y. Du, Mater. Sci. Eng., I. Takigawa, H. Chi Dam, Sci. Technol. Adv. Mater. 2017, 18, 756.
C 2020, 106, 110265. [114] G. Pilania, C. Wang, X. Jiang, S. Rajasekaran, R. Ramprasad, Sci.
[80] Y. Wang, Y. Liu, S. Song, Z. Yang, X. Qi, K. Wang, Y. Liu, Q. Zhang, Rep. 2013, 3, 2810.
Y. Tian, Nat. Commun. 2018, 9, 2444. [115] X. Li, Y. Zhang, H. Zhao, C. Burkhart, L. Catherine Brinson,
[81] D. Sholl, J. A. Steckel, Density Functional Theory: A Practical Intro- W. Chen, Sci. Rep. 2018, 8, 13461.
duction, John Wiley & Sons, Hoboken, NJ 2011. [116] X. Wu, V. Kumar, J. Ross Quinlan, J. Ghosh, Q. Yang, H. Motoda,
[82] J. Behler, J. Chem. Phys. 2016, 145, 170901. G. J. Mclachlan, A. Ng, B. Liu, P. S. Yu, Z.-H. Zhou, M. Steinbach,
[83] S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, D J. Hand, D. Steinberg, Knowl. Inf. Syst. 2008, 14, 1.
M. Aykol, S. Rühl, C. Wolverton, npj Comput. Mater. 2015, 1, 15010. [117] K. Arulkumaran, M. P. Deisenroth, M. Brundage, A. A. Bharath,
[84] A. Zakutayev, N. Wunder, M. Schwarting, J D. Perkins, IEEE Signal Process. Mag. 2017, 34, 26.
R. White, K. Munch, W. Tumas, C. Phillips, Sci. Data 2018, 5, [118] C. Szepesvári, Synth. Lectures Artif. Intell. Mach. Learn. 2010,
180053. 4, 1.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (12 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

[119] J. Oh, et al., Discovering reinforcement learning algorithms. arXiv pre- [153] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu,
print arXiv:2007.08794, 2020. X. Wang, G. Wang, J. Cai, T. Chen, Pattern Recognition 2018, 77,
[120] T. G. Dietterich, Ensemble Methods in Machine Learning, in Interna- 354.
tional Workshop on Multiple Classifier Systems. 2000, Springer. [154] S. Hochreiter, J. Schmidhuber, Neural Comput. 1997, 9, 1735.
[121] Z.-H. Zhou, Ensemble Methods: Foundations and Algorithms, CRC [155] M. Schuster, K. K. Paliwal, IEEE Trans. Signal Process. 1997, 45,
Press, Boca Raton, FL 2012. 2673.
[122] J. Schmidhuber, Neural Networks 2015, 61, 85. [156] Y. Yu, X. Si, C. Hu, J. Zhang, Neural Comput. 2019, 31, 1235.
[123] S. Kearnes, L. Li, Richard N. Zare, P. Riley, Sci. Rep. 2019, 9, 10752. [157] C. Rao, Y. Liu, Comput. Mater. Sci. 2020, 184, 109850.
[124] Z. Li, D. Wu, T. Yu, J. Manuf. Sci. Eng. 2019, 141, 031003. [158] S. N. S. Mortazavi, A. Ince, Comput. Mater. Sci. 2020, 185, 109962.
[125] J. R. Quinlan, C4. 5: Programs for Machine Learning, Elsevier, [159] C. Rudin, Nat. Mach. Intell. 2019, 1, 206.
Amsterdam 2014. [160] M. T. Ribeiro, S. Singh, C. Guestrin. “Why should i trust you?”
[126] I. Rizkya, et al. Analysis of Defective Causes in Coffee Product Using Explaining the predictions of any classifier. in Proceedings of the 22nd
Decision Tree Approach. in IOP Conference Series: Materials Science ACM SIGKDD international conference on knowledge discovery and
and Engineering, 2020, IOP Publishing. data mining, 2016.
[127] M. Martinez-Arroyo, L. E. Sucar, Learning an optimal naive bayes [161] W. J. Murdoch, et al., Interpretable machine learning: definitions,
classifier. in 18th International Conference on Pattern Recognition methods, and applications. arXiv preprint arXiv:1901.04592, 2019.
(ICPR’06), 2006, IEEE. [162] Y. Iwasaki, R. Sawada, V. Stanev, M. Ishida, A. Kirihara, Y. Omori,
[128] V. Solov’ev, A. Tsivadze, G. Marcou, A. Varnek, Mol. Inf. 2019, 38, H. Someya, I. Takeuchi, E. Saitoh, S. Yorozu, npj Comput. Mater.
1900002. 2019, 5, 103.
[129] C. Cortes, V. Vapnik, Mach. Learn. 1995, 20, 273. [163] B. Kailkhura, B. Gallagher, S. Kim, A. Hiszpanski, T. Y.-J. Han, npj
[130] S. H. Lee, J. Mazumder, J. Park, S. Kim, J. Manuf. Processes 2020, Comput. Mater. 2019, 5, 108.
55, 307. [164] S. Hooker, et al., A benchmark for interpretability methods in deep
[131] D. Jahed Armaghani, P. G. Asteris, B. Askarian, M. Hasanipanah, neural networks. arXiv preprint arXiv:1806.10758, 2018.
R. Tarinejad, V. V. Huynh, Sustainability 2020, 12, 2229. [165] B. Mittelstadt, C. Russell, S. Wachter. Explaining explanations in AI.
[132] T. Cover, P. Hart, IEEE Trans. Inf. Theory 1967, 13, 21. in Proceedings of the conference on fairness, accountability, and trans-
[133] T. Adithiyaa, D. Chandramohan, T. Sathish, Mater. Today: Proc. parency. 2019.
2020, 21, 1000. [166] L. H. Gilpin, et al. Explaining explanations: An overview of interpret-
[134] T. Adithiyaa, D. Chandramohan, T. Sathish, Mater. Today: Proc. ability of machine learning. in 2018 IEEE 5th International Conference
2020, 21, 108. on data science and advanced analytics (DSAA), 2018, IEEE.
[135] R. E. Schapire, Explaining Adaboost, in Empirical Inference, 2013, [167] K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko,
Springer, 37–52. K.-R. Müller, J. Chem. Phys. 2018, 148, 241722.
[136] Ji Li, C. Zhang, X. Zhang, H. He, W. Liu, C. Chen, IEEE Access [168] L. Himanen, M. O. J. Jäger, E V. Morooka, F. Federici Canova,
2020, 8, 12413. Y S. Ranawat, D Z. Gao, P. Rinke, A S. Foster, Comput. Phys.
[137] W. Y. Loh, Wiley Interdiscip. Rev.: Data Mining Knowl. Discovery Commun. 2020, 247, 106949.
2011, 1, 14. [169] L. Ward, A. Dunn, A. Faghaninia, N. E. R. Zimmermann, S. Bajaj,
[138] S. Kadali, S. M. Naushad, A. Radha Rama Devi, V. L. Bodiga, Mol. Qi Wang, J. Montoya, J. Chen, K. Bystrom, M. Dylla, K. Chard,
Cell. Biochem. 2019, 458, 27. M. Asta, K A. Persson, G. J Snyder, I. Foster, A. Jain, Comput.
[139] S. Na, L. Xumin, G. Yong. Research on k-means clustering algorithm: Mater. Sci. 2018, 152, 60.
An improved k-means clustering algorithm. in 2010 Third Interna- [170] S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia,
tional Symposium on intelligent information technology and security D. Gunter, V. L. Chevrier, K. A. Persson, G. Ceder, Comput. Mater.
informatics, 2010, Ieee. Sci. 2013, 68, 314.
[140] J. Lin, T. N. Pappas, Structural Texture Similarity for Material Rec- [171] T. Ueno, T. D. Rhone, Z. Hou, T. Mizoguchi, K. Tsuda, Mater. Dis-
ognition. in 2019 IEEE International Conference on Image Processing covery 2016, 4, 18.
(ICIP), 2019, IEEE. [172] E. Gossett, C. Toher, C. Oses, O. Isayev, F. Legrain, F. Rose,
[141] G. Weber, M. Pinz, S. Ghosh, JOM 2020, 72, 4404. E. Zurek, J. Carrete, N. Mingo, A. Tropsha, S. Curtarolo, Comput.
[142] I. C. Gormley, T. B. Murphy, Encyclop. Stat. Quality Reliab. 2008, 2. Mater. Sci. 2018, 152, 134.
[143] T. Masuda, T. Nishio, J. Kataoka, M. Arimoto, A. Sano, [173] M. Abadi, et al. Tensorflow: A system for large-scale machine learning.
K. Karasawa, Phys. Med. Biol. 2019, 64, 175011. in 12th {USENIX} symposium on operating systems design and imple-
[144] R. Agrawal, R. Srikant, Fast Algorithms for Mining Association Rules, mentation ({OSDI} 16). 2016.
in Proc. 20th Int. Conf. Very Large Data Bases, VLDB, 1994, Citeseer. [174] A. Paszke, et al., Pytorch: An imperative style, high-performance deep
[145] S. Ren, X. Zhao, B. Huang, Z. Wang, X. Song, J. Ambient Intell. learning library. arXiv preprint arXiv:1912.01703, 2019.
Humanized Comput. 2019, 10, 1093. [175] A. Gulli, S. Pal, Deep learning with Keras, Packt Publishing Ltd,
[146] A. Langville, C. Meyer, Internet Math. 2004, 1, 335. Birmingham 2017.
[147] G. Bastos, et al., Development of an inverse identification method [176] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
for identifying constitutive parameters by metaheuristic optimization O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg,
algorithm: Application to hyperelastic materials, in Residual Stress, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot,
Thermomechanics & Infrared Imaging and Inverse Problems, 2020, E. Duchesnay, J. Mach. Learn. Res. 2011, 12, 2825.
Springer, p. 141. [177] A. Géron, Hands-on machine learning with Scikit-Learn, Keras, and
[148] B. Liu, S. Jiang, Q. Zou, Brief. Bioinf. 2020, 21, 298. TensorFlow: Concepts, tools, and techniques to build intelligent sys-
[149] G. E. Hinton, S. Osindero, Y.-W. Teh, Neural Comput. 2006, 18, tems, 2019, O’Reilly Media.
1527. [178] V. K. Verma, et al. Towards zero-shot learning with fewer seen class
[150] G. E. Hinton, Deep Belief Networks, 2009, 4, 5947. examples. in Proceedings of the IEEE/CVF Winter Conference on
[151] J. Bouvrie, Notes on Convolutional Neural Networks. 2006. Applications of Computer Vision. 2021.
[152] R. Yamashita, M. Nishio, R. K. G. Do, K. Togashi, Insights Imaging [179] V. Garcia, J. Bruna, Few-shot learning with graph neural networks.
2018, 9, 611. arXiv preprint arXiv:1711.04043, 2017.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (13 of 14) © 2021 Wiley-VCH GmbH
 16163028, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/adfm.202108044 by Stanford University, Wiley Online Library on [26/04/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.afm-journal.de

[180] Z. Li, et al., Meta-sgd: Learning to learn quickly for few-shot learning. [184] F. Scarselli, M. Gori, Ah Chung Tsoi, M. Hagenbuchner,
arXiv preprint arXiv:1707.09835, 2017. G. Monfardini, IEEE Trans. Neural Networks 2008, 20, 61.
[181] B. M. Lake, R. Salakhutdinov, J. B. Tenenbaum, Science 2015, 350, 1332. [185] X. Yi, E. Walia, P. Babyn, Med. Image Anal. 2019, 58, 101552.
[182] E A. Olivetti, J M. Cullen, Science 2018, 360, 1396. [186] A. Vaswani, Advances in Neural Information Processing Systems,
[183] T. Bai, D. Li, K. Sun, Y. Chen, W. Li, Remote Sens. 2016, 8, 715. 2017.

Chaochao Gao received his Bachelor’s degree from School of Materials Science and Technology,
Taiyuan University of Science and Technology in 2018. He is currently pursuing his Master’s
degree under the supervision of Assoc. Prof. Xin Min and Prof. Zhaohui Huang. His research
focuses on the machine learning for material design optimization.

Xin Min received his BS degree (2011) and Ph.D. degree (2016) from School of Materials Science
and Technology, China University of Geosciences (Beijing). Then, he joined the faculty of School
of Materials Science and Technology, China University of Geosciences (Beijing). He is a visiting
scholar in Department of Materials and Metallurgy, University of Cambridge. His main research
interests include designing high-performance functional materials for energy storage and
conversion.

Zhaohui Huang received his Ph.D. degree from School of Materials Science and Engineering,
University of Science and Technology Beijing in 2000. He joined the faculty of China University
of Geosciences (Beijing) (CUGB) in 2004. He has been the associate Dean of the School of
Materials Science and Engineering in CUGB. He has hosted the State Key Program of National
Natural Science of China, the “12th Five-Year” Plan key projects supported by National Science and
technology, and the National Key R&D Program of China. Currently, his main research interests
include advanced ceramics, refractories, and material utilization of solid waste.

Adv. Funct. Mater. 2022, 32, 2108044 2108044 (14 of 14) © 2021 Wiley-VCH GmbH