Journal of Materiomics 9 (2023) 798e816

Contents lists available at ScienceDirect

Journal of Materiomics
journal homepage: www.journals.elsevier.com/journal-of-materiomics/

Generative artificial intelligence and its applications in materials

science: Current situation and future perspectives
Yue Liu a, d, Zhengwei Yang a, Zhenyao Yu a, Zitu Liu a, Dahui Liu a, Hailong Lin b,
Mingqing Li b, Shuchang Ma a, Maxim Avdeev e, f, Siqi Shi b, c, *
a
School of Computer Engineering and Science, Shanghai University, Shanghai, 200444, China
b
State Key Laboratory of Advanced Special Steel, School of Materials Science and Engineering, Shanghai University, Shanghai, 200444, China
c
Materials Genome Institute, Shanghai University, Shanghai, 200444, China
d
Shanghai Engineering Research Center of Intelligent Computing System, Shanghai, 200444, China
e
Australian Nuclear Science and Technology Organisation, Sydney, 2232, Australia
f
School of Chemistry, The University of Sydney, Sydney, 2006, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Generative Artificial Intelligence (GAI) is attracting the increasing attention of materials community for
Received 2 May 2023 its excellent capability of generating required contents. With the introduction of Prompt paradigm and
Received in revised form reinforcement learning from human feedback (RLHF), GAI shifts from the task-specific to general pattern
10 May 2023
gradually, enabling to tackle multiple complicated tasks involved in resolving the structure-activity re-
Accepted 12 May 2023
Available online 25 May 2023
lationships. Here, we review the development status of GAI comprehensively and analyze pros and cons
of various generative models in the view of methodology. The applications of task-specific generative
models involving materials inverse design and data augmentation are also dissected. Taking ChatGPT as
Keywords:
Machine learning
an example, we explore the potential applications of general GAI in generating multiple materials
Artificial intelligence content, solving differential equation as well as querying materials FAQs. Furthermore, we summarize six
Generative artificial intelligence challenges encountered for the use of GAI in materials science and provide the corresponding solutions.
Materials science This work paves the way for providing effective and explainable materials data generation and analysis
Novel materials discovery approaches to accelerate the materials research and development.
Deep learning © 2023 The Authors. Published by Elsevier B.V. on behalf of The Chinese Ceramic Society. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
2. Various GAI methods and their applications to materials science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
2.1. Generative Adversarial Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
2.2. Variational autoencoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
2.3. Diffusion model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
2.4. Flow model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
2.5. General GAI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805
3. Future applications of general GAI in materials science: ChatGPT example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
3.1. Generating materials data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
3.2. Solving differential equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
3.3. Querying materials FAQs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 810
4. Analysis of problems of GAI in materials science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812
4.1. High-quality data and domain knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812
4.2. Generalization ability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812

* Corresponding author. State Key Laboratory of Advanced Special Steel, School of

Materials Science and Engineering, Shanghai University, Shanghai 200444, China.
E-mail address: [email protected] (S. Shi).
Peer review under responsibility of The Chinese Ceramic Society.

https://doi.org/10.1016/j.jmat.2023.05.001
2352-8478/© 2023 The Authors. Published by Elsevier B.V. on behalf of The Chinese Ceramic Society. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

4.3. Interpretability and credibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813

4.4. Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
4.5. Resource cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
4.6. Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814
Declaration of competing interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814

1. Introduction issues, e.g., high training and maintenance costs, high-quality data
sparseness, poor domain knowledge integration, model interpret-
Data-driven machine learning (ML) brought the possibility of ability and credibility, resource sharing, security, etc.
the “4th paradigm” of materials research and development due to In this paper, we review the status of GAI development sys-
the effective ability of ML to uncover the structure-activity re- tematically and summarize the cons and pros of different GAI
lationships in materials data [1]. However, the paradigm shift is models and scope of their applications. Then, the studies of using
hindered because of slow progress in data quality governance and GAI for materials science are reviewed and the future applications
the lack of the guidance for using domain knowledge in combina- of GAI in materials science are discussed, aiming to aid researchers
tion with data-driven analysis, which can be summarized as three to rapidly select the proper generative models. Following this, we
key issues, i.e., high dimensionality of feature space vs. small perform trails, taking ChatGPT as an example, to explore how
sample, model accuracy vs. usability, and ML results vs. domain general GAI can solve the high-profile materials issues, e.g., novel
knowledge [2]. Embedding domain knowledge into models with material data generation, the solution of differential equations and
generative ability would resolve the above-mentioned issues and query of materials FAQs. Finally, the challenges of the development
enable such models to mine structure-activity relationships more of GAI for materials science are analyzed and discussed in detail,
accurately. such as the requirements of high-quality data and domain knowl-
Generative Artificial Intelligence (GAI), an unsupervised or edge, the improvement of model generalization, interpretability
partially supervised ML framework [3], enables the input samples and credibility, usability, security, and resource cost.
to be semantically edited or disassembled through disturber of the The remainder of this paper is organized as follows: Section 2
input distribution or generation of the distribution for fitting the summarizes the development process of various GAI models from
posterior distribution of the target samples. Such operations can the perspective of specific tasks and general types and introduces
assist in analyzing the inherent relationship between features their applications in materials science. Sections 3 explores possible
while generating target samples accurately, which has been widely applications of general GAI for dealing with the domain key issues.
employed in image and video generation, sequence modeling, The challenges of development of GAI in materials science will be
speech enhancement, etc. [4,5]. For materials science, the target discussed in Section 4. Finally, the significant conclusions from this
materials data can be edited through embedding physical rules and review are presented in Section 5.
operations (e.g., translation and rotation) into the generation pro-
cess of GAI, which has been applied to material performance pre-
2. Various GAI methods and their applications to materials
diction and new material discovery [6e8]. For example,
science
Hoogeboom et al. [9] introduced different Euclidian groups into the
generation process of GAI, to improve quality of generated samples.
Nowadays, researchers modify the learning algorithms or model
Zhao et al. [10] modified the generation process by the description
structures of GAI to fit specific domains to tackle specific tasks.
of crystal space information, to efficiently generate novel materials
Thanks to the accumulation of big data and the abundance of
with high crystal structural diversity. These studies illustrate that
computing resources, large-scale models such as ChatGPT and GPT-
GAI is conducive to the deep coupling of AI for Science (AI4Science)
4 are proposed to improve the generalization ability to solve
[11] with materials science, thus accelerating the progress.
various tasks, which can be regarded as the beginning of general
Nowadays, researchers tend to train complicated generative
GAI. In this section, the development process of task-specific and
models on large datasets, leveraging the Prompt paradigm [12] to
general GAI research is reviewed, especially the applications of GAI
accomplish the few-shot learning [13] even zero-shot learning [14].
in materials research, as shown in Fig. 1.
Then, through the combination of reinforcement learning from
human feedback (RLHF) and such models (e.g., ChatGPT) [15],
domain knowledge can participate in the model training process in 2.1. Generative Adversarial Network
a highly coupled way, which drives the transfer of the paradigm
from “Fitting-Generation” to “Pretraining-Prompting-Generation”. Generative Adversarial Networks (GANs) [18] were first intro-
There have been studies that demonstrated high accuracy of the duced by Goodfellow et al. in 2014. As shown in Fig. 2(a), GAN is a
content generation in multimodal data generation and analysis, deep learning architecture comprised of a generator and a
multi-context reasoning, and code design [16,17], which is prom- discriminator. The generator creates synthetic data that mimics the
ising for achieving realization of the general GAI. These models can real data, while the discriminator distinguishes between real and
be employed to modify and generate materials data of interest fake data. During the adversarial training process, the generator
through the multimodal data generation and domain-knowledge- and discriminator engage in a competition, wherein the generator
embedded generation patterns, thereby breaking the barriers of strives to produce more realistic data and the discriminator aims to
different scales, systems, and types of materials. However, the accurately classify whether the data is real or fake. As a generative
current development of general GAI still suffers from some key model, GAN is often used to generate samples for data augmenta-
tion and data pre-processing methods in deep learning, and has
799
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 1. Development of GAI and its application in materials science. The left arrow means the research heat in computer science. The right arrow means the employment frequency
in materials science.

been widely used in image processing, biomedicine, and other issue of an overly unconstrained training process. To address the
fields. However, the Vanilla GAN is immature, it still has some issue of unstable training and mode collapse, others have proposed
limitations such as data mismatch, unstable training, and mode alternative GAN models, including Wasserstein GAN (WGAN) [21],
collapse. To address these issues, researchers have taken various WGAN with gradient penalty (WGAN-GP) [22], and CycleGAN [23].
approaches and conducted extensive analysis and research. For These models optimize the loss function to improve the perfor-
instance, researchers have developed conditional GAN (CGAN) [19] mance of GAN. In addition, to adapt to tabular data, some re-
and deep convolutional GAN (DCGAN) [20] to improve model searchers proposed TableGAN [24], TabularGAN (TGAN) [25], and
performance by introducing additional constraints, mitigating the conditional tabular GAN (CTGAN) [26] to tackle generating different

800
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 2. Schematic diagram of vanilla generative models: (a) Generative adversarial network; (b) variational autoencoder; (c) diffusion model; (d) flow model.

801
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

types of data by optimizing training schemes. Fig. 3 lists the real data. Furthermore, GAN have been used in materials reverse
detailed comparison of GAN flavors. design to more efficiently sample materials space and generate
Nowadays, GAN and its variants are employed to generate data hypothetical materials. For instance, Dan et al. [30] proposed a
and reverse design in the field of materials science. In data generative model MatGAN for data augmentation in the materials
augmentation, GAN and TGAN are two popular tools for generating field. Based on this model, they sampled the materials databases
additional training data for performance prediction models. For ICSD, MP, and OQMD, and finally generated 2 million hypothetical
example, Ma et al. [27] used GAN to expand 47 images of poly- inorganic materials. Subsequently, they explored the two-
crystalline iron to 136, and performance of the model based on the dimensional (2D) material space and used MatGAN to generate
generated data with 35 % of the real data reached the same level of 2.65 million 2D material samples, ultimately discovering 26,489
performance as that based only on the real data, indicating that new potential 2D materials, of which 1 485 2D materials had an
GAN have certain potential in generating images. Marani et al. [28] area under the curve (AUC) score greater than 0.95 and 12 2D/
used TGAN to generate 6513 new data points based on 810 exper- layered materials were confirmed through DFT calculations [31].
imental values to predict the compressive strength of ultra-high These works demonstrate the potential of GAN in expanding ma-
performance concrete. The results demonstrated that the ML terials reverse design space and screening large-scale inorganic
model trained on the generated data exhibited comparable pre- materials datasets. Kim et al. [32] directly used GAN to predict
diction performance to that trained with real data, which shows ternary compounds containing only Mg, Mn, and O atoms, ulti-
that TGAN can accurately capture the distribution of real datasets mately discovering 23 crystal structures suitable for photo-
and generate high-quality synthetic data. Yang et al. [29] proposed electrochemical anode applications. Long et al. [33] developed a
a two-step data augmentation method based on GAN to enhance reverse design framework CCDCGAN based on DCGAN for pre-
the hardness prediction data of high-entropy alloys. The experi- dicting crystal structures with target properties, and successfully
mental results showed that the RMSE of the ML model constructed generating crystal structures with low formation energy that had
based on the original training set and generated data is 0.123, not been previously reported, thus accelerating the discovery of
which is 6.1 % less accurate than the model constructed using only new phases.

Fig. 3. Development of GAN and its examples.

802
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

In summary, GAN has shown promising capability in image representation, thus effectively searching for candidate materials
generation, tabular data augmentation, and new materials discov- with higher target performance. Noh et al. [46] proposed a VAE
ery and design. The trained GAN can better predict and optimize strategy framework for the inverse design of solid-state materials,
the materials properties, overcoming the limitations of small which efficiently explores the chemical compositional space by
experimental datasets. New materials can be generated by learning learning the distributions of known materials for crystal structure
the structure information from existing materials and further prediction. This method successfully discovered high-performance
exploring the latent space of materials. However, challenges still vanadium oxides that were previously identified in experiments
exist in the application of GAN in materials science. Sometimes the and generated about 20 000 hypothetical materials, generating
new materials generated by GAN are difficult to synthesize and more than 40 new metastable high-performance vanadium oxides
apply experimentally. In addition, due to the lack of reliable and with energy above hull
80 meV/atom, i.e., potentially synthesiz-
consistent evaluation indicators, the prediction results are some- able. Additionally, due to the natural disentanglement phenome-
times inaccurate. Thus, optimizing the network architecture, non exhibited by the latent space of VAEs, Sardeshmukh et al. [47]
objective function, and proposing new improvement methods are designed a VAE model based on style loss to learn interpretable
necessary to advance the application of GAN in materials science. disentangled latent representations of material microstructures,
capturing important attributes responsible for mechanical prop-
2.2. Variational autoencoder erties and generating structures with particular properties. This
approach of using latent variables to explore the connection be-
The Variational Autoencoder (VAE) [34] was first introduced by tween the microstructure and properties of materials holds
Diederik P. Kingma et al. in 2013. This model provides a concise way promise for devising reverse inference and is of great significance
of capturing the essential low-dimensional information from the for designing materials and products with targeted performance.
data, which can be used to generate new samples through simple When representing compounds as vectors of components’ pro-
manipulation of the learned low-dimensional representations via a portions, the problem of discrete features and severe sparsity ari-
decoder. VAE has emerged as a crucial method for both represen- ses. The traditional VAE model used for processing continuous
tation learning and generative modeling. As shown in Fig. 2(b), VAE image data cannot directly generate discrete data. To address this,
is composed of an encoder and a decoder. The encoder extracts the Oubari et al. [48] proposed Binded-VAE, a model specifically
mean and variance of the latent variables that determine its designed for generating highly sparse discrete datasets.
properties from the data through a neural network. The decoder The prospects of using VAEs for material design are promising,
adds Gaussian noise to the encoded information to generate new as they have shown to be effective in generating novel materials
data. Unlike traditional autoencoders [35], which only compress with desired properties. However, there are still challenges that
the image to a point in the latent space, variational autoencoders need to be addressed for VAEs to be fully effective in material
force the latent variables of the latent space to a standard normal design. Firstly, the VAE architecture needs to be optimized for
distribution. The encoder is no longer given a point but a distri- materials data, which is different from image or text data
bution, allowing the model to learn a smooth latent state repre- commonly used in VAE applications. Secondly, the objective func-
sentation of the input data. This enables the model to cover unseen tions used to optimize the VAE should be carefully designed to
samples in the input data. However, the original VAE has limita- accurately reflect the desired material properties. Thirdly, VAEs
tions such as unsatisfactory generation effect, unsuitability for should be able to handle and generate data with high dimension-
discrete data, and an insufficient disentangling effect. In order to ality, such as crystal structures, and effectively model the correla-
solve these problems, researchers proposed various improvements, tions between different properties of a material. Finally, it is
such as VAE-GAN [36], VLAE [37], NVAE [38], etc., which enhance important to validate the generated materials to ensure that they
the effect of model generation. Additionally, VQ-VAE [39], JointVAE both have desired properties and are stable in real-world condi-
[40], VQ-VAE-2 [41], and other approaches have been developed to tions. Despite these challenges, VAEs have the potential to revolu-
adapt to discrete data. Lastly, a variety of methods have been sug- tionize the way materials are designed and discovered with further
gested to improve the model's disentangling effect, including advancements.
b-VAE [42], FactorVAE [43], and b-TCVAE [44]. Fig. 4 provides a
detailed comparison of VAE and its related improved models. 2.3. Diffusion model
In materials science, VAE has become a promising tool to solve
material problems. The VAE-based models transform discrete ma- The Denoising Diffusion Probabilistic Model (DDPM) [49], pro-
terial data representation into multi-dimensional continuous var- posed by Jonathan Ho and Pieter Abbeel in 2020, outperformed
iables through an encoder, and then convert the multi-dimensional GAN [50] in image synthesis upon its release. As a result, many
continuous variables into material representation through a subsequent studies on image generation have shifted towards
decoder. Due to the distribution constraints imposed by the model DDPM-based research, leading to impressive results for the diffu-
on its latent space, the latent space becomes a data distribution sion model in various generative modeling tasks. The diffusion
rather than being discrete data points. Therefore, in terms of ma- model acquires the distribution of some crucial parameters through
terial generation, new material structures can be generated by a neural network, as illustrated in Fig. 2 (c). During the training
performing simple operations on latent space distribution, to process, the model first performs a forward pass by gradually
effectively explore the compound space, which was not originally adding noise to the original data. At each step, the parameters are
covered. For example, Go  mez-Bombarelli et al. [45] applied VAE to iterated through a Markov process until the data is transformed
the design of drug-like molecules. By using an encoder network to into a form that can be approximately represented by pure
convert a discrete molecular representation into a continuous Gaussian noise. Following the forward process, the model then
vector in the latent space, and then performing simple operations engages in a backward pass whereby the recently acquired repre-
on the latent continuous vector, such as perturbing known chem- sentation form undergoes a gradual reduction of noise through
ical structures, or interpolating between molecules. Then the decoding, resulting in the generation of new data. By aiming to
modified vector can be converted back into a new discrete molec- achieve a greater resemblance between the new and original data,
ular representation through a decoder. Finally, a predictor module the model's parameters are continuously optimized, leading to an
is used to predict the chemical properties of the new molecular ongoing improvement in the quality of the generated content. The
803
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 4. The development of VAE and its application examples.

training process of diffusion models is known to provide a higher generating multiple ligands with high binding affinity for specific
level of stability when compared to GANs, while its ability to protein targets. Similarly, to design proteins of high fidelity as
generate a broad range of high-quality samples surpasses that of regards both sequence and structure, Shi et al. [65] proposed a new
VAEs. However, the original diffusion model has three main collaborative design method of protein sequence and structure
shortcomings: slow sampling speed, poor maximum likelihood, based on diffusion model, PROTSEED, which can translate protein
and weak data generalization ability. Researchers have addressed sequence and structure from random initialization to specific
these three shortcomings with various methods tailored to the task desired state based on given prior context features.
at hand. For instance, to improve sampling speed, methods such as In summary, the diffusion model has garnered increasing
DDIM [51], DP [52], DDSS [53], GENIE [54] and Two-step distillation attention and has been applied to material generation due to its
[55] have been proposed. For enhancing data generalization, outstanding performance. The diffusion model, after being adapted
methods like LSGM [56] and D3PMs [57] have been developed. And to material properties, can generate specific materials that meet the
Improved DDPM [58], VDM [59], ScoreFlow [60] and Analytic-DPM desired requirements, thereby exploring novel materials with
[61] have been put forward to strengthen the maximum likelihood. improved properties. However, compared with other generative
Fig. 5 shows the details of the diffusion model and its related models, the diffusion model is still hindered by issues such as large
modifications. memory consumption and computational costs, and is not yet
The diffusion model, with its unique network structure, requires mature enough in handling non-image data such as material data.
only one forward and corresponding inverse process training to Therefore, further optimization of the diffusion model's network
achieve stable training loss and excellent model performance. It can architecture, objective functions, and data applicability is required
be combined with material characteristics to generate new mate- to effectively tackle material-related challenges.
rials that meet the desired targets. Lim et al. [62] used the diffusion
model to generate an optimal microstructure with multi- 2.4. Flow model
functionality. By conducting multi-objective optimization, the light
sensitivity and fracture toughness of the material were improved, Dinh Laurent et al. [66] introduced Non-Linear Independent
offering a novel approach to characterize performance-based Components Estimation (NICE) in 2015, which marked the birth of
composite material microstructures. For proteins as macromolec- flow models. Unlike other generative models that attempt to
ular compounds, Anand et al. [63] introduced a fully data-driven approximate the true distribution of the data, flow models directly
diffusion model for protein structure, sequence, and rotamers, solve the mapping transformation between two distributions by
enabling the generation of highly realistic proteins across the full manipulating the Jacobian determinant. This unique approach has
range of domains in the Protein DataBank. Schneuing et al. [64] made flow models a distinct and noteworthy type of generative
proposed the DiffSBDD model that respects translation, rotation, model. Flow model establishes the mapping relationship between
reflection, and permutation equivariance, which is capable of the distribution of data representation and its latent variable
804
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 5. The development of diffusion models and their applications.

distribution through a series of reversible transformations. Based which is an important factor to hinder the application of ML in
on this, new content can be generated by sampling from the latent materials science. Ohno [71] applied normalizing flows to the
variable distribution. The structure of flow model is shown in Fig. 2 dataset of ionic conductivity of materials and augmented the data
(d). Flow model has two characteristics. One is that it can calculate by using a real-valued non-volume preserving model (RealNVP) as
the distribution of latent variables that determine data represen- the normalizing flow. This effectively solves the problem of insuf-
tation, which ensures that the generated content and training data ficient training data in material performance prediction.
come from the same distribution. The second is that the entire The utilization of flow models in materials science may provide
model is built by a series of reversible changes and the data is a promising avenue driving advancements in the field. However, it
converted to a priori distribution through a bijective function, is still limited by many aspects, such as model, computational, and
which improves the interpretability of the model. Researchers have material structure complexity and optimal parameter setting is-
studied the performance improvement of flow models and pro- sues. In summary, the application of flow models in the field of
posed RealNVP [67] and Glow [68] by optimizing the network material application requires comprehensive consideration of ma-
structure based on NICE. In addition, Chen et al. [69] combined terial characteristics, physical mechanisms, computational effi-
differential equations and deep learning to construct continuous ciency, and reliability, so as to realize its full potential in materials
normalizing flows, thus avoiding the single-unit bottleneck prob- research.
lem in traditional normalizing flows. Grathwohl et al. [70] further
constructed an unbiased estimator with O(D) cost, which has been
improved in high-dimensional density estimation, image genera- 2.5. General GAI
tion and variational inference. Fig. 6 shows the comparison of
classic Flow-based models. In contrast to other generative models, The advancement of big data and data representation technol-
the flow model only uses a reversible encoder to complete the ogies has led to the pursuit of generating human-understandable
construction of the model, which makes the design of the model language from input data patterns and structures, with the capa-
more delicate. However, this also raises the issue of more param- bility to accomplish goals across diverse environments. This aim
eters and increased computing cost, which is the crucial challenge seeks to surpass the current language generation paradigm, which
that still needs to be addressed. is limited to fitting sample distributions for specific tasks. Large
Flow models are also useful in solving problems in the field of language models (LLMs), with their powerful self-supervised
materials science. Material data have the issue of small data size, learning ability, have demonstrated a remarkable capacity for
“general” intelligence, enabling them to perform complex tasks and
805
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 6. Comparison of classic Flow-based models.

answer questions without intrinsic motivation or objectives. than 91 % accuracy. The results show that such LLMs can obtain
Beyond its mastery of language, LLMs represented by ChatGPT acceptable performance to explore scientific knowledge and design
based on transformer can solve novel and difficult tasks that span novel materials. However, since it is hard to update the parameters
mathematics, coding, vision, medicine, law, psychology and more, of LLMs, such models fail to solve complicated materials tasks
without needing any special prompting, and are seen as the pro- effectively. Moreover, the lack of domain knowledge aggravates the
totype of general GAI. Nowadays, Google proposed their own LLM gap between model analysis results and real situation.
named Claude, which has impressive dialogue and task processing The advent of Reinforcement Learning from Human Feedback
skills, to compete with ChatGPT or other advanced models pro- (RLHF) facilitates the shift from task-specific to general GAI in the
posed by OpenAI. Meanwhile, some Chinese companies such as case of InstructGPT [15]. RLHF optimizes the original problem by
HUAWEI, Baidu, Alibaba and Tencent all proposed their own LLMs, converting non-differentiable objectives in language generation
aiming to build industrial models and help industrialize AI large task into sequential decision processes. Compared to traditional RL,
models. Here, we take ChatGPT as an example to discuss the RLHF is better able to align with human preferences and learn
development process of general GAI. fluently from various types of rewards. Organizing human prefer-
Fig. 7 illustrates the evolution of generic GAI. Initially, Genera- ences on demand is integral to creating a true general-purpose
tive Pre-trained Transformer (GPT) displayed its potential for task- intelligence. By using RLHF for fine-tuning, InstructGPT can make
specific natural language generation via unsupervised pre-training decisions that are consistent with human preferences, resulting in
and fine-tuning on downstream tasks. It utilizes 12 transformer- an intelligent system with greater adaptability, interpretability, and
decoder layers to execute next word prediction and coherent text reliability. Furthermore, optimization of the fine-tuning process
generation. The purpose of fine-tuning is to adapt it to a specific improves the exploration efficiency of the model, allowing it to
task based on pre-training. GPT-2 [72] expands the model structure achieve better performance but fewer parameters compared to
and parameters from its predecessor and trains on various datasets GPT-3.
beyond web text. Despite exhibiting advanced results with zero- ChatGPT uses the same methods as InstructGPT but with slight
shot learning, it still falls under the category of task-specific GAI. difference in the data collection setup. As depicted in Fig. 7, an
The GPT-3 [73] leverages Prompt to minimize the need for large initial model is first trained via supervised fine-tuning: human AI
supervised datasets. It models the linguistic structure of text trainers provide conversations in which they play both sidesdthe
probability, enabling the probabilistic predictions. Specifically, the user and an AI assistant. This new dialogue dataset is mixed with
approach allows the language model to be pre-trained on a large the InstructGPT dataset and converted into a Q&A format. Then,
amount of text, enabling it to perform few-shot or zero-shot reward models for reinforcement learning are created by collecting
learning and adapt to new scenarios quickly by defining a new and ranking conversations between the AI trainer and the chatbot,
cue template, even with limited or no labeled data available. along with alternative responses. Finally, the model is fine-tuned
Compared to “Pre-train þ Fine-tune” [74], Prompt extensively taps using proximal policy optimization (PPO) based on these reward
into the prior knowledge of language model, allowing it to “Recall” models. Several iterations of the process resulted in a conversa-
the information learned during the pre-training stage when tional generative model that performed with high efficacy. In
executing generation. Consequently, with appropriate prompts, the addition to RLHF-based training techniques, the deployment of
LLMs can handle numerous downstream tasks in a completely model is also crucial. ChatGPT adopts an iterative deployment
unsupervised manner. In materials science, Xie et al. [75] proposed approach to minimize the risks associated with generative models.
a novel NLP task called structured information inference, which This includes conducting a security assessment before deployment,
aims to address the complexities of information extraction at the executing beta testing with a small user group, and closely moni-
device level in material science. Then, GPT-3 was leveraged and toring usage with retrospective reviews. For materials research,
trained on perovskite solar cell FAIR dataset, which gained more Hong et al. [76] explored some promising application of ChatGPT

806
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 7. Development of General GAI and the details of ChatGPT.

(namely building structures, writing codes for specific scientific formidable capabilities of GPT-4 in multimodal generation and
software, and preparing data visualization scripts). The results conversational interactivity offer a promising outlook for materials
show that such model may generate results with some errors, but it science research. This breakthrough may drive the development of
is still promising. novel studies, such as crystal image analysis, crystallographic in-
Recently, GPT-4, the latest model developed by OpenAI, was formation file (CIF) generation, complex differential equation
trained with an unprecedented scale of computations and data, solving, and material Q&A systems, all of which could facilitate
surprisingly achieved human-like performance across almost all exploration of materials structure-activity relationships and dis-
tasks and significantly outperformed its predecessors. The emer- covery of new materials. Furthermore, ensuring the credibility,
gence of GPT-4 signifies a “general” advancement of general GAI usability, and security of generated content remains a critical
based on the GPT series, encompassing the perception and gener- challenge that must be addressed for the future application of
ation of multimodal data such as text, images, and audio. The general GAIs in materials science.

807
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

3. Future applications of general GAI in materials science: generated by ChatGPT were progressively corrected and improved,
ChatGPT example particularly in terms of their spatial symmetry. As ChatGPT lacks
domain knowledge about atomic arrangement, atomic number, and
High-quality data is critical for materials science research, atomic bonding, even after multiple adjustments and questions, it
providing valuable insights for experiments and samples for ML. only reduces the generation of some keys. Though the results above
The advanced ability of general GAI to generate data, code, and suggest that such general GAI can provide feedback to the user
model now offers materials researchers a great deal of convenience. during the interaction process and has self-correction function,
In this section, we explore various applications of general GAI currently ChatGPT fails to generate compliant CIF for aiding in
(taking ChatGPT as an example), namely generating materials data, materials research and discovery.
solving differential equation, and querying FAQs, and analyze its Experimental data is the fundamental demand for materials
reliability and availability for materials research and discovery. research, but as experiments and measurements are time-
consuming and labor-intensive, the materials sample size is
commonly small. To this end, the powerful data generation capa-
3.1. Generating materials data bility of ChatGPT appears to offer a solution. By extracting data from
published literature, ChatGPT can generate Excel and CSV files in
Materials research generates a large amount of data containing the corresponding format. However, the accuracy of the generated
domain knowledge, which not only is the materials research data sets is not satisfactory. Despite outputting ionic conductivity
foundation but also provide motivation for researchers. As a general and activation energy values based on the search criteria, they do
GAI, ChatGPT has shown great capability of generating data, which not match the literature content. Furthermore, the reference titles
has led to broad application prospects in materials science research. and provenance information are often incorrect.
To this end, in this section ChatGPT is employed to generate CIFs In summary, at present ChatGPT often outputs incorrect data on
and experimental data, and then evaluate its reliability and avail- request, but can be used to quickly retrieve the original corre-
ability of materials data generation. sponding documents, which can save researchers time in data
CIF is a file format proposed by the International Union for collection. With the introduction of GPT-4, it becomes possible to
Crystallography (IUCr) to archive and distribute crystal structure access to more types of data, yet the accuracy of the data also needs
information of materials. Such files include chemical formulas, unit more experimental verification. It is believed that with the devel-
cell parameters, space groups, atomic coordinates, position occu- opment of general GAI, GAI models can be employed to output
pancies, etc., and have been widely employed to report crystal higher-quality multi-modal data for researchers.
structures. Therefore, accurate CIF information is fundamental for
scientific research on materials. Here, we utilize ChatGPT to 3.2. Solving differential equation
generate CIF files for four systems of solid electrolytes i.e.,
Li0.5La0.5(TiO3), Li7La3Zr2O12, Li14ZnGe4O14 and NaZr2(PO4)3, shown In materials science, mathematical equations are constructed to
in Fig. 8. Although ChatGPT can generate the concept of full and express numerical relationships between various attributes and
partial occupancy of atom sites and the generated CIF files can be performance parameters. Differential equations are widely used in
recognized by VESTA, their information is far from the required material science. For example, Chen et al. [77] used Neural ordinary
crystal structure, which is due to the lack of domain knowledge differential equations (ODEs) to simulate behavior in spintronic
about crystal structures related to solid electrolytes. devices with high precision and time efficiency. Li et al. [78]
To this end, we leverage Prompt to edit and improve the gen- investigated the effect of diaphragm thickness and diaphragm
eration process, thus effectively modifying the generated results surface coating on dendrite growth using a phase field model with
guided by the domain knowledge. Taking NaZr2(PO4)3 as an a non-linear partial differential equation (PDE). In this process of
example, we provide corresponding domain knowledge e.g., rules construction of a suitable differential equation, domain knowledge
related to space groups and key-value pairs, to modify the gener- plays a significant role.
ated results, as follows: Naturally, it is crucial to solve these equations so that they
approximate the true values. Nowadays, mainstream simulation
Rule 1: The space group of NaZr2(PO4)3 should be “R3C”. software, such as ANSYS, Hyperworks, and COMSOL, have achieved
Rule 2: The migration ion Na occupies three different sites, high levels of success in simulation computing. ChatGPT also owns
namely Na1 sites (6b), Na2 sites (18e) and Na3 sites (36f), Zr ion the powerful capability of mathematics, which can greatly reduce
occupies 12c site and the P ion occupies the 18e site. computational costs while providing accurate results. To this end,
Rule 3: The crystal structure should be a three-dimensional here we present validation experiments using ChatGPT to solve
skeleton that forms a crystal structure through the vertices- differential equations as an example, to explore the potential of this
sharing of ZrO6 octahedra and PO4 tetrahedra. The sodium capability in materials science.
ions are distributed in different interstices of the skeleton, As illustrated in Fig. 9, after inputting the equation and its
forming Na (1) O6 polyhedron, Na (2) O8 polyhedron and Na (3) relevant parameters in the dialogue box, ChatGPT performs the
O5 polyhedron. necessary calculation steps and analysis. For simple differential
equations, conversion to a computer-comprehensible language is
Prompted by above rules, ChatGPT gradually corrected the CIF not necessary, and ChatGPT can automatically convert and solve the
file. As shown in Fig. 8, Rule 1 can correct the crystal structure of equations. While evaluating the discrete error of the differential
NaZr2(PO4)3 effectively. This emphasizes that spatial symmetry equation, due to computing power constraints, ChatGPT cannot
reflected by the space group is critical for the crystal structure. provide a complete solution and calculation process. However, after
Moreover, owing to Rule 2, ChatGPT can effectively modify the sites entering the command “Generate a visualization file of the discrete
of skeleton atoms. After modifications following Rule 1 and 2, the errors,” ChatGPT generates the corresponding Python code that can
crystal structure has better approximated the target structure. produce the required result after execution. It is also possible to
Compared to the unconstrainted bond generation, Rule 3 effec- request an alternative approach to the problem for better out-
tively prevented erroneous bonds. As the number of questions and comes. In the case of more complex Partial Differential Equations
the application of domain knowledge increased, the CIF files (PDEs), ChatGPT fails to produce the full solution process, indicating
808
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 8. Comparison of generative and true CIF files and the revising process of NaZr2(PO4)3.

809
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 9. Process of solving differential equations by ChatGPT.

its low capability to PDE solution. 3.3. Querying materials FAQs

Since many areas of materials science involve differential
equations involving physics and chemistry knowledge, at present Modification of materials according to application requirements
the ChatGPT cannot directly support materials computations or is one of the most important tasks in materials science. Thereinto,
simulations, but its capability to understand and generate limited one of the crucial steps in material modification research is to
computational content and programming, e.g., Python, code, which analyze and employ publicly available modification strategies,
can be used with external software, is evident. which is the foundation of experiment design. However, this task
can be challenging and time-consuming as it requires researchers
to read numerous papers, reviews, and patents. Furthermore,
important publications may be missing, e.g., due to limited access to

810
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

journals, which can result in suboptimal experimental designs or disadvantages of various dopants.
even duplication of completed work. Therefore, if ChatGPT could be The results show that ChatGPT has the potential to serve as a
used to query materials-science-related FAQs (Frequently Asked “document steward” for materials researchers, which can efficiently
Questions), such as the modification strategy of the material, it gather relevant information from numerous literature sources.
would alleviate the effort to review the literature for researchers Nevertheless, its current limitations include the inability to provide
and thus facilitate experiment design. insight into the value of retrieved information and the difficulty in
The perovskite-type solid electrolyte Li0.5La0.5(TiO3) (LLTO) collecting important relevant details. This leads to having to
possesses stable physical and chemical properties, making it a continuously ask follow-up questions. Therefore, improvements in
promising candidate for the next generation of all-solid-state these areas through iterative upgrades can significantly enhance the
lithium-ion batteries. However, the ionic conductivity of LLTO utility of such general GAI models for materials researchers.
falls short of meeting the requirements for high charge/discharge The current applications of ChatGPT in materials science
rates of lithium-ion batteries. Therefore, it necessary to enhance the demonstrate its rich knowledge base, making it suitable for
ionic conductivity of LLTO. Herein, we take LLTO solid electrolyte as providing users with basic knowledge of materials science. The
an example, to validate the feasibility of querying material modi- Prompt paradigm enables users to obtain corresponding knowl-
fication strategy through ChatGPT. As shown in Fig. 10 (left), edge and correct misconceptions, while its data processing and
ChatGPT collects LLTO's modification schemes from various per- code writing capabilities provide researchers with great conve-
spectives, which provides reference for researchers. However, nience. However, its limitations are also apparent, i.e., it lacks
ChatGPT fails to offer a practical solution by considering the ad- specialized material domain knowledge and most of the answers
vantages and disadvantages of each scheme and cannot provide fail to meet the needs of material researchers. Additionally, the
specific methods to determine whether we should increase or availability of material text data generated by ChatGPT is limited
decrease the temperature, prolong, or shorten the annealing time. and often contrary to material domain knowledge. The ability of
Moreover, ChatGPT responds that the ionic conductivity of LLTO can ChatGPT to process complex data is also constrained by input
be improved by controlling the grain size and introducing a sec- format and computing power, often leading to incomplete output
ondary phase but not how that can be achieved. When queried how of calculation results. Despite these challenges, ChatGPT has
to improve the ionic conductivity of Li0.5La0.5(TiO3) through doping, demonstrated impressive potential for knowledge output and data
ChatGPT not only provided several doping schemes, but also calculation in the field of materials. Its future development is highly
described the mechanism of each doped ion, as shown in the right anticipated.
part of Fig. 10. However, ChatGPT failed to provide advantages and

Fig. 10. Results of FAQs for ChatGPT. (Left) Question about How to improve the ionic conductivity of Li0.5La0.5(TiO3); (Right) Question about How to improve the ionic conductivity of
Li0.5La0.5(TiO3) via doping.

811
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Fig. 11. Contradiction between model learning ability and interpretability.

4. Analysis of problems of GAI in materials science generated samples leveraging the low-quality samples will amplify
the defect of the original samples and result in serious contradic-
GAI has participated in materials research and discovery widely. tion with the original samples or domain knowledge. To this end,
The feasibility of embedding domain knowledge (e.g., fundamental domain knowledge can be symbolized and integrated into data
physical laws) into GAI models brings new possibilities for AI4S- representation methods, which enables GAI models to learn basic
cience to drive the research and development of novel materials. physical laws while retaining the effective information and elimi-
The experiments described above demonstrate that general GAIs, nating redundant information. Moreover, the data quality detection
such as ChatGPT, seem to be able to solve materials science prob- method with embedded domain knowledge [83] can effectively
lems from multiple aspects, which illustrates the potential to improve the data quality via identifying possible abnormal data in
change the paradigm of materials research and development. samples and remove abnormal samples under the guidance of
In the past, with the increase of model scale, the interpretability materials domain knowledge. Through these approaches, GAI can
of the model greatly decreased, and its characteristics of “black generate new samples following domain knowledge, and thus build
box” profoundly affected the reliability, controllability, and credi- a virtuous circle that original and generated samples can be
bility, as shown in Fig. 11. In contrast, the most recent developments merged.
in GAI revealed new behavior emerging with further model scale On the other hand, the objectivity, authenticity, and represen-
increase. The latest models, such as GPT, demonstrate extraordi- tation pattern of domain knowledge are key issues for the appli-
nary ability in data representation, data analysis and few-shot even cation of GAI in materials science. Embedding wrong domain
zero-shot learning, which never occurs in smaller models [79]. This knowledge into GAI models will hinder them from mining the
induces that people begin to consider whether and where GAI structure-activity relationships latent in the materials data. Domain
should be utilized. To this end, Liu et al. [2,80] proposed the knowledge is mostly contained in scientific literature. Although
framework of machine learning embedded with materials domain researchers have employed large-scale language models to mine
knowledge, which aims to transfer “black box” to “gray box” or historical data or empirical knowledge from materials scientific
even “white box” via embedding various domain knowledge into literature [84e87], insufficient attention is often paid to the quality
the entire process of ML. In this way, three key issues in materials of domain knowledge. Therefore, it will be helpful for the
science can be effectively reconciled and the interpretability, improvement of the domain knowledge quality to quantify the
credibility and robustness of ML can be greatly improved. In 2017, credibility of materials scientific literature and perform the credi-
we have pointed out several common problems of ML in materials bility analysis, which can be conducive to researchers to screen out
science [1], which is still applicable to current GAI in materials domain knowledge with higher credibility from massive literature
science. Following this, we here discuss and analyze the problems according to task requirements [88].
of the future development of GAI and its applications in materials
science, then provide possible solutions.
4.2. Generalization ability

4.1. High-quality data and domain knowledge Generalization ability is a vital metric for measuring the per-
formance of AI models. The generalization ability of GAI models is
Data representation affects the amount of valid information in usually expressed as the distribution of generated samples
samples, which in turn determines the data quality [81,82]. Effi- approximating the inputs with some extrapolation capability.
cient data representation for unstructured materials data facilitates When a GAI model is successful, it should interpolate or even
GAI to fit the characteristics of the original samples more accurately extrapolate beyond the original training data which will be a
through Gaussian distribution. Note that the paradigm from continuation of the original data distribution. The generated ma-
“Fitting-Generation” is appropriate for the samples with strong terials data should have a similar or the same structure as the
correlation between features (e.g., image, video). However, the high original and it should be the interpolation or extrapolation of in-
complexity of the structure-activity relationship latent in materials puts. For example, Ref. [33] developed a GAN-based inverse design
data makes it difficult to accurately fit the association between framework to generate distinct crystal structures and applied it to
descriptors (features) through Gaussian distribution. Hence, the the binary Bi-Se system. Ref. [89] proposed an approach combining
812
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

computational chemistry and GAN to generate new catalytic sur- order to improve the usability of GAI models, it is necessary to
faces in an extrapolative manner. It is worth noting that the introduce Neural Architecture Search (NAS) to achieve the best
discrepancy in the generalization ability between the GAI models possible performance using limited computing resources in an
focusing on specific materials and general GAI models. The former automated manner with minimal manual intervention [94,95].
have poor generalization ability to other types of materials yet Moreover, the complexity of the preparation and data labeling will
generate highly accurate content for the target materials while the reduce the usability of the GAI model. To this end, besides
latter have strong generalization ability but with questionable embedding physical information into generative process for
reliability of the generated content. Moreover, how to effectively manipulating generated samples, active learning is an effective
evaluate the generalization ability of the GAI models is an urgent approach to the selection of generated unlabeled samples [96],
problem to be handled. Therefore, it is necessary to comprehen- which not only can improve the performance of candidate samples
sively consider the sample quality, sample content, and GAI models. and reduce the number of candidates enormously but also reduce
Embedding domain knowledge and common-sense constraints the demand of human cost via screening out high-quality unlabeled
into GAI should also be explored, as that would make outputs more samples.
accurate, diverse, and controllable.

4.3. Interpretability and credibility

4.5. Resource cost
Although GAI models have the potential to revolutionize the
While GAI models have powerful generation ability, its resource
current paradigm of materials research and discovery, there is an
consumption is a noteworthy problem, which includes training
ongoing concern about the credibility and interpretability of GAI
costs and operating costs. Thereinto, the training cost refers to the
models. Model interpretability refers to the ability of a model to
cost required for the model to produce a certain generation ca-
interpret the correlation between the input data and the target, as
pacity, e.g., computing power requirements, equipment re-
well as the semantic representation of the feature space and the
quirements, and training time costs. The operating cost of the
behavior of the model during the learning process, which is widely
model mainly considers the energy consumption required for the
discussed as a key factor of credibility. This enables the model
model operation. The advanced generation ability of GAI models is
builder or user to understand the learned features, which in turn
due to the utilization of high computing power, which enables such
can inform decisions about the model's performance and how to
models to process vast amounts of data and learn complicated
improve it. Therefore, how to explain and improve the interpret-
patterns. Therefore, improving the learning efficiency in a reason-
ability of generative models is an urgent problem that needs to be
able and effective manner is key to reducing these costs. For
addressed. Researchers [42,43,90,91] have attempted to focus on
example, Ref. [97] proposed a standard pruning technique to
the semantic features contained in the learned latent space and
compress and optimize the GAI models without sacrificing gener-
manipulating them to explain the internal feature processing of the
ation quality. Ref. [98] explored the relationship between opti-
model. Note that the GAI model has extremely high latent space
mizing the training efficiency and resource consumption of GAI
dimension and complex internal semantics, so how to interpret its
models. Furthermore, the accumulation of large amount of labeled
internal semantic features is a key problem. Therefore, in addition
data requires significant time and human cost. Hence, it is impor-
to understanding the internal mechanism of the GAI models, it is
tant to deal with computing resources and the efficiency of model
necessary to examine their generation process and understand how
output also from the perspective of model compression and usage
it obtains the results step by step, e.g., a dynamic evaluation
costs, to facilitate GAI to be highly integrated with different fields
mechanism can be introduced to improve the interpretability,
and produce new research paradigms in various fields in the future.
which will enable continuous interaction with the model and
provide self-explanatory outputs. Moreover, embedding domain
knowledge [92] into GAI models can also be regarded as an effective
way. By these means, researchers can gain more accurate and 4.6. Security
reliable insights from their outputs, facilitating the development of
new materials. Through such models we can also comprehend GAI models mostly own the characteristics of “black box”, which
which features (descriptors) are crucial for the model. raises security concerns [99]. For example, attackers can exploit
generative models to create convincing fake text [98] (e.g., false
4.4. Usability news articles, social media posts, and phishing schemes), thus
deceiving people into engaging in unlawful activities. Moreover,
Usability measures the difficulty of employing GAI models to GAI models are vulnerable to attack with respect to data privacy
deal with practical problems, which manifests in two aspects. 1) and intellectual property, such as training data or model parame-
Complex and unpredictable generation process requires to carefully ters. As a result, it is essential to conduct in-depth research on se-
screen out reliable results from numerous candidate generated curity of generative models and to strengthen defense measures
data. For example, Ref. [31] leveraged GAN to generate 267 489 new against potential attacks. For example, Ref. [100] proposed defense
potential compositions while only 92 novel materials were techniques to safeguard GAI models, considering a trade-off be-
confirmed by DFT formation energy calculation. Ref. [93] utilized tween the utility and security of generative models. For materials
GAN to generate 10 million samples, where appropriate 4.1 million science, the inputs of GAI models are mainly obtained from ex-
samples can be obtained for downstream screening. Finally, only periments or computation, for which it is necessary to consider the
506 materials were verified by phonon dispersion calculation. 2) intellectual property rights and security. In the future, the devel-
Though deeper GAI models have significant ability to data repre- opment of GAI will involve more content generation, including
sentation and exploration of latent structure-activity relationships, intellectual property, therefore, how to enhance the security of GAI
the optimization space of hyperparameters and the demand of models and safeguard both the generated content and internal data
labeled data can experience an exponential surge. Therefore, in of the models is an urgent issue.

813
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

5. Conclusion References

The appearance of large-scale generative models driven by [1] Liu Yue, Zhao Tianlu, Ju Wangwei, Shi Siqi. Materials discovery and design
using machine learning. J. Materiom. 2017;3:159e77.
Prompt paradigm and RLHF algorithm shifts the generation para- [2] Liu Yue, Guo Biru, Zou Xinxin, Li Yajie, Shi Siqi. Machine learning assisted
digm from “Fitting-Generation” to “Pretraining-Prompting-Gener- materials design and discovery for rechargeable batteries. Energy Storage
ation”. Such models offer the hope of general GAI realization, that Mater 2020;31:434e50.
[3] Jovanovi c MlaCan, Campbell Mark. Generative artificial intelligence: trends
can facilitate the deep fusion between AI4Science and materials and prospects. Computer 2022;55:107e12.
science and accelerate innovation in materials science research. [4] Cao Hanqun, Tan Cheng, Gao Zhangyang, Chen Guangyong, Ann Heng Pheng,
Through reviewing the status of GAI, it can be seen that in computer Stan Z Li. A survey on generative diffusion model. arXiv:220902646 2022.
[5] Jabbar Abdul, Li Xi, Omar Bourahla. A survey on generative adversarial net-
science researchers improve the multi-aspect abilities of GAI works: variants, applications, and training. ACM Comput Surv 2021;54:1e49.
models according to specific demands through different ap- [6] Sanchez-Lengeling Benjamin, Aspuru-Guzik Ala n. Inverse molecular design
proaches (e.g., objective function optimization, condition addiction, using machine learning: generative models for matter engineering. Science
2018;361:360e5.
etc.) and numerous improved models are proposed. The pros and
[7] Luo Shitong, Shi Chence, Xu Minkai, Tang Jian. Predicting molecular
cons of various generative models are summarized, aiming to offer conformation via dynamic graph score matching. Adv Neural Inf Process Syst
a guidance to their applications in materials science. Following that, 2021;34:19784e95.
[8] Chen Litao, Zhang Wentao, Nie Zhiwei, Li Shunning, Pan Feng. Generative
we review the research in GAI for materials science and find that
models for inverse design of inorganic solid materials. J. Mater. Inform.
materials researchers prefer to employ basic GAI models, while 2021;1:4.
there are many promising improved and advanced GAI methods, [9] Hoogeboom Emiel, Garcia Satorras Victor, Vignac Cle ment, Welling Max.
which have not been explored yet. Specifically, we discuss the Equivariant diffusion for molecule generation in 3D. In: Proceedings of the
39th international conference on machine learning, 162; 2022. p. 8867e87.
prospects of general GAI in materials science and leverage ChatGPT [10] Zhao Yong, Edirisuriya M, Siriwardane Dilanga, Wu Zhenyao, Fu Nihang, Al-
as an example for exploring whether it can solve the key issues in Fahdi Mohammed, Hu Ming, et al. Physics guided deep learning for gener-
materials science (e.g., novel materials generation, solving differ- ative design of crystal materials with symmetry constraints. npj Comput
Mater 2023;9:38.
ential equations, and querying FAQs). The results show that [11] Rick Stevens, Taylor Valerie, Nichols Jeff, Arthur Barney Maccabe,
ChatGPT can effectively deal with materials problems and modify Katherine Yelick, Brown David. AI for science: report on the department of
the generated results via Prompt to some extent. However, because energy (DOE) town Halls on artificial intelligence (AI) for science. 2020.
United States.
of the lack of specific domain knowledge in its training dataset, [12] Liu Pengfei, Yuan Weizhe, Fu Jinlan, Jiang Zhengbao, Hayashi Hiroaki,
ChatGPT currently still cannot be employed to handle the compli- Graham Neubig. Pre-train, prompt, and predict: a systematic survey of
cated or domain-knowledge-required materials problems. Finally, prompting methods in natural language processing. ACM Comput Surv
2023;55:1e35.
the six challenges of the use of GAI in materials science are dis-
[13] Wang Yaqing, Yao Quanming, James T, Kwok. Generalizing from a few ex-
cussed, i.e., the requirements of high-quality data and domain amples: a survey on few-shot learning. ACM Comput Surv 2020;53(3):1e34.
knowledge, the improvement of model generalization ability, [14] Xian Yongqin, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-
shot learningda comprehensive evaluation of the good, the bad and the
interpretability and credibility, usability, security, and resource
ugly. IEEE Trans Pattern Anal Mach Intell 2018;41:2251e65.
cost. We believe that embedding domain knowledge into GAI [15] Ouyang Long, Wu Jeffrey, Jiang Xu, Almeida Diogo, Wainwright Carroll,
models would resolve many of these problems. The improvement Mishkin Pamela, et al. Training language models to follow instructions with
of the data and knowledge quality can provide solid foundation for human feedback. Adv Neural Inf Process Syst 2022;35:27730e44.
[16] Bubeck Se bastien, Chandrasekaran Varun, Ronen Eldan, Gehrke Johannes,
materials research and development. AI experts, materials domain Horvitz Eric, Kamar Ece, et al. Sparks of artificial general intelligence: early
experts, and non-experts will need to combine effort to resolve the experiments with GPT-4. arXiv:230312712 2023.
identified issues. Overall, a general GAI is bound to promote the [17] Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiaao Chen, Michihiro Yasu-
naga, and Diyi Yang. Is chatgpt a general-purpose natural language pro-
rapid development of AI4Science, and then open a new era of cessing task solver?. arXiv:230206476.
materials science research. [18] Goodfellow Ian, Pouget-Abadie Jean, Mirza Mehdi, Xu Bing, Warde-
Farley David, Ozair Sherjil, et al. Generative adversarial networks. Commun
ACM 2020;63:139e44.
[19] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets.
arXiv:14111784.
Declaration of competing interest
[20] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation
learning with deep convolutional generative adversarial networks. arXiv:
The authors declare that they have no known competing 151106434.
[21] Martin Arjovsky, Chintala Soumith, Bottou Le on. Wasserstein generative
financial interests or personal relationships that could have
adversarial networks. In: Proceedings of the 34th international conference
appeared to influence the work reported in this paper. on machine learning, 70; 2017. p. 214e23.
[22] Gulrajani Ishaan, Ahmed Faruk, Martin Arjovsky, Vincent Dumoulin, Aaron C
Courville. Improved training of wasserstein GANs. Adv Neural Inf Process
Syst 2017;30.
Acknowledgment [23] Zhu Junyan, Park Taesung, Isola Phillip, Alexei A. Efros. Unpaired image-to-
image translation using cycle-consistent adversarial networks. In: Pro-
ceedings of the IEEE international conference on computer vision. ICCV);
This work was supported in part by National Natural Science 2017. p. 2223e32.
Foundation of China [grant number 92270124, 52073169] and Na- [24] Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hon-
gkyu Park, and Youngmin Kim. Data synthesis based on generative adver-
tional Key Research and Development Program of China [grant sarial networks. arXiv:180603384.
number 2021YFB3802101] the Key Research Project of Zhejiang [25] Lei Xu and Kalyan Veeramachaneni. Synthesizing tabular data using gener-
Laboratory [grant number 2021PE0AC02]. We appreciated the ative adversarial networks. arXiv:181111264.
[26] Xu Lei, Skoularidou Maria, Cuesta-Infante Alfredo, Veeramachaneni Kalyan.
High-Performance Computing Center of Shanghai University and
Modeling tabular data using conditional GAN. Adv Neural Inf Process Syst
Shanghai Engineering Research Center of Intelligent Computing 2019;32.
System for providing the computing resources and technical [27] Ma Boyuan, Wei Xiaoyan, Liu Chuni, Ban Xiaojuan, Huang Haiyou, Wang Hao,
support. et al. Data augmentation in microscopic images for material data mining. npj

814
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

Comput Mater 2020;6:1e9. based diffusion models. Adv Neural Inf Process Syst 2021;34:1415e28.
[28] Marani Afshin, Jamali Armin, Moncef L, Nehdi. Predicting ultra-high- [61] Bao F, Li C, Zhu J, et al. Analytic-dpm: an analytic estimate of the optimal
performance concrete compressive strength using tabular generative reverse variance in diffusion probabilistic models. arXiv preprint arXiv:
adversarial networks. Materials 2020;13:4757. 220106503 2022.
[29] Yang Zhiyuan, Li Shu, Li Shuai, Jia Yang, Liu Dongrong. A two-step data [62] Lim HJ, Lee K-H, Yun GJ. Microstructure design of multifunctional particulate
augmentation method based on generative adversarial network for hardness composite materials using conditional diffusion models. arXiv preprint
prediction of high entropy alloy. Comput Mater Sci 2023;220:112064. arXiv:230109051 2023.
[30] Dan Yabo, Zhao Yong, Xiang Li, Li Shaobo, Hu Ming, Hu Jianjun. Generative [63] Anand N, Achim T. Protein structure and sequence generation with equiv-
adversarial networks (GAN) based efficient sampling of chemical composi- ariant denoising diffusion probabilistic models. arXiv preprint arXiv:
tion space for inverse design of inorganic materials. npj Comput Mater 220515019 2022.
2020;6:84. [64] Schneuing A, Du Y, Harris C, et al. Structure-based drug design with equiv-
[31] Song Yuqi, Edirisuriya M, Siriwardane Dilanga, Zhao Yong, Hu Jianjun. ariant diffusion models. arXiv preprint arXiv:221013695 2022.
Computational discovery of new 2D materials using deep learning genera- [65] Shi C, Wang C, Lu J, et al. Protein sequence and structure co-design with
tive models. ACS Appl Mater Interfaces 2021;13:53303e13. equivariant translation. arXiv preprint arXiv:221008761 2022.
[32] Kim Sungwon, Noh Juhwan, Gu Geun Ho, Aspuru-Guzik Alan, Jung Yousung. [66] Dinh L, Krueger D, Bengio Y. Nice: non-linear independent components
Generative adversarial networks for crystal structure prediction. ACS Cent estimation. arXiv preprint arXiv:14108516 2014.
Sci 2020;6:1412e20. [67] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation
[33] Long T, Fortunato NM, Opahle I, et al. Constrained crystals deep convolu- using real nvp. arXiv:160508803.
tional generative adversarial network for the inverse design of crystal [68] Kingma Durk P, Dhariwal Prafulla. Glow: generative flow with invertible 1x1
structures. npj Comput Mater 2021;7:66. convolutions. Adv Neural Inf Process Syst 2018;31:10236e45.
[34] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv: [69] Chen Ricky TQ, Rubanova Yulia, Bettencourt Jesse, Duvenaud David K. Neural
13126114. ordinary differential equations. Adv Neural Inf Process Syst 2018;31:
[35] Michelucci U. An introduction to autoencoders. arXiv preprint arXiv: 6572e83.
220103898 2022. [70] Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David
[36] Larsen ABL, Sønderby SK, Larochelle H, et al. Autoencoding beyond pixels Duvenaud. FFJORD: free-form continuous dynamics for scalable reversible
using a learned similarity metric. In: International conference on machine generative models. arXiv:181001367.
learning. PMLR; 2016. p. 1558e66. [71] Ohno Hiroshi. Training data augmentation: an empirical study using
[37] Chen X, Kingma DP, Salimans T, et al. Variational lossy autoencoder. arXiv generative adversarial net-based approach with normalizing flow models for
preprint arXiv:161102731 2016. materials informatics. Appl Soft Comput 2020;86:105932.
[38] Vahdat A, Kautz J. NVAE: a deep hierarchical variational autoencoder. Adv [72] Radford Alec, Wu Jeffrey, Child Rewon, Luan David, Amodei Dario,
Neural Inf Process Syst 2020;33:19667e79. Sutskever Ilya. Language models are unsupervised multitask learners.
[39] Van Den Oord A, Vinyals O. Neural discrete representation learning. Adv OpenAI blog 2019;1:9.
Neural Inf Process Syst 2017;30. [73] Brown Tom, Mann Benjamin, Ryder Nick, Subbiah Melanie, Kaplan Jared D,
[40] Dupont E. Learning disentangled joint continuous and discrete representa- Dhariwal Prafulla, et al. Language models are few-shot learners. Adv Neural
tions. Adv Neural Inf Process Syst 2018;31. Inf Process Syst 2020;33:1877e901.
[41] Razavi A, Van den Oord A, Vinyals O. Generating diverse high-fidelity images [74] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT:
with vq-vae-2. Adv Neural Inf Process Syst 2019;32. pre-training of deep bidirectional transformers for language understanding.
[42] Higgins I, Matthey L, Pal A, et al. beta-vae: learning basic visual concepts with arXiv:181004805.
a constrained variational framework. 2016. [75] Tong Xie, Yuwei Wan, Wei Huang, Yufei Zhou, Yixuan Liu, and Qingyuan
[43] Kim H, Mnih A. Disentangling by factorising. In: International conference on Linghu et al. Large Language models as master key: unlocking the secrets of
machine learning. PMLR; 2018. p. 2649e58. materials science with GPT. arXiv:230402213.
[44] Chen RT, Li X, Grosse RB, et al. Isolating sources of disentanglement in [76] Hong Zijian. ChatGPT for Computational Materials Science: A Perspective.
variational autoencoders. Adv Neural Inf Process Syst 2018;31. Energy Mater Adv. 2023;4:0026.
[45] Gomez-Bombarelli R, Wei JN, Duvenaud D, et al. Automatic chemical design [77] Chen Xing, Abreu Araujo Flavio, Riou Mathieu, Jacob Torrejon,
using a data-driven continuous representation of molecules. ACS Cent Sci Ravelosona Dafine , Wang Kang, et al. Forecasting the outcome of spintronic
2018;4(2):268e76. experiments with neural ordinary differential equations. Nat Commun
[46] Noh J, Kim J, Stein HS, et al. Inverse design of solid-state materials via a 2022;13:1016.
continuous representation. Matter 2019;1(5):1370e84. [78] Li Yajie, Sha Liting, Lv Peili, Qiu Na, Zhao Wei, Chen Bin, et al. Influences of
[47] Sardeshmukh A, Reddy S, Gautham B, et al. TextureVAE: learning inter- separator thickness and surface coating on lithium dendrite growth: a
pretable representations of material microstructures using variational phase-field study. Materials 2022;15:7912.
autoencoders. In: AAAI spring symposium. MLPS; 2021. [79] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, and Sebastian
[48] Oubari F, De Mathelin A, De catoire R, et al. A binded VAE for inorganic Borgeaud et al. Emergent abilities of large language models. arXiv:
material generation. arXiv preprint arXiv:211209570 2021. 220607682.
[49] Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models. Adv Neural [80] Liu Yue, Zou Xinxin, Yang Zhengwei, Shi Siqi. Machine learning embedded
Inf Process Syst 2020;33:6840e51. with materials domain knowledge. J Chin Ceram Soc 2022;50:863e76.
[50] Dhariwal P, Nichol A. Diffusion models beat GANs on image synthesis. Adv. [81] Fuhr Addis S, Sumpter Bobby G. Deep generative models for materials dis-
Neural Inform. Process. Sys. 34 2021;34(Neurips 2021):8780e94. covery and machine learning-accelerated innovation. Frontiers in Materials
[51] Song J, Meng C, Ermon S. Denoising diffusion implicit models. arXiv preprint 2022;9:865270.
arXiv:201002502 2020. [82] Liu Yue, Yang Zhengwei, Zou Xinxin, Ma Shuchang, Liu Dahui,
[52] Watson D, Ho J, Norouzi M, et al. Learning to efficiently sample from diffu- Avdeev Maxim, et al. Data quantity governance for machine learning in
sion probabilistic models. arXiv preprint arXiv:210603802 2021. materials science. Natl Sci Rev 2023;10. nwad125.
[53] Watson D, Chan W, Ho J, et al. Learning fast samplers for diffusion models by [83] Shi Siqi, Sun Shiyu, Ma Shuchang, Zou Xinxin, Qian Quan, Liu Yue. Detection
differentiating through sample quality. In: International conference on method on data accuracy incorporating materials domain knowledge. J Inorg
learning representations; 2022. Mater 2022;37:1311.
[54] Dockhorn T, Vahdat A, Kreis K. GENIE: higher-order denoising diffusion [84] Weston L, Tshitoyan V, Dagdelen J, Kononova O, Trewartha A, Persson KA,
solvers. arXiv preprint arXiv:221005475 2022. et al. Named entity recognition and normalization applied to large-scale
[55] Meng C, Gao R, Kingma DP, et al. On distillation of guided diffusion models. information extraction from the materials science literature. J. Chem. Inf.
arXiv preprint arXiv:221003142 2022. Model. 2019;59:3692e702.
[56] Vahdat A, Kreis K, Kautz J. Score-based generative modeling in latent space. [85] Tshitoyan Vahe, Dagdelen John, Weston Leigh, Dunn Alexander, Rong Ziqin,
Adv Neural Inf Process Syst 2021;34:11287e302. Kononova Olga, et al. Unsupervised word embeddings capture latent
[57] Austin J, Johnson DD, Ho J, et al. Structured denoising diffusion models in knowledge from materials science literature. Nature 2019;571:95e8.
discrete state-spaces. Adv Neural Inf Process Syst 2021;34:17981e93. [86] Nie Zhiwei, Zheng Shisheng, Liu Yuanji, Chen Zhefeng, Li Shunning, Lei Kai,
[58] Nichol AQ, Dhariwal P. Improved denoising diffusion probabilistic models. et al. Automating materials exploration with a semantic knowledge graph
In: International conference on machine learning. PMLR; 2021. p. 8162e71. for Li-ion battery cathodes. Adv Funct Mater 2022;32:2201437.
[59] Kingma D, Salimans T, Poole B, et al. Variational diffusion models. Adv Neural [87] Liu Yue, Ge Xianyuan, Yang Zhengwei, Sun Shiyu, Liu Dahui, Avdeev Maxim,
Inf Process Syst 2021;34:21696e707. et al. An automatic descriptors recognizer customized for materials science
[60] Song Y, Durkan C, Murray I, et al. Maximum likelihood training of score- literature. J Power Sources 2022;545:231946.

815
Y. Liu, Z. Yang, Z. Yu et al. Journal of Materiomics 9 (2023) 798e816

[88] Liu Yue, Ding Lin, Yang ZhengWei, Ge XianYuan, Liu DaHui, Liu Wei, et al. [100] Hu Hailong, Pang Jun. Stealing machine learning models: attacks and
Domain knowledge discovery from abstracts of scientific literature on countermeasures for generative adversarial networks. In: Annual computer
Nickel-based single crystal superalloys. Sci China Technol Sci 2023. https:// security applications conference; 2021. p. 1e16.
doi.org/10.1007/s11431-022-2283-7.
[89] Ishikawa Atsushi. Heterogeneous catalyst design by generative adversarial
network and first-principles based microkinetics. Sci Rep 2022;12:11657.
Yue Liu obtained her B.S. and M.S. in computer science
[90] Shen Yujun, Yang Ceyuan, Tang Xiaoou, Zhou Bolei. Interfacegan: inter-
from Jiangxi Normal University in 1997 and 2000. She
preting the disentangled face representation learned by GANs. IEEE Trans
finished her Ph.D. in control theory and control engi-
Pattern Anal Mach Intell 2020;44:2004e18.
neering from Shanghai University (SHU) in 2005. She has
[91] Liu Yue, Liu Zitu, Li Shuang, Yu Zhenyao, Guo Yike, Liu Qun, et al. Cloud-VAE:
been working with the School of Computer Engineering
variational autoencoder with concepts embedded. Pattern Recogn 2023;140:
and Science of SHU since July 2000. During that time, she
109530.
has been a curriculum R&D manager at the Sybase-SHU IT
[92] Laura von Rueden, Sebastian Mayer, Katharina Beckh, Bogdan Georgiev, Sven
Institute of Sybase Inc. from July 2003 to July 2004 and a
Giesselbach, and Raoul Heese et al. Informed machine learning–A taxonomy
visiting scholar at the University of Melbourne from Sep.
and survey of integrating knowledge into learning systems. arXiv:
2012 to Sep. 2013. At present, she is a professor of SHU.
190312394.
Her current research interests focus on research of data
[93] Zhao Yong, Al-Fahdi Mohammed, Hu Ming, Edirisuriya M, Siriwardane D,
mining, machine learning, and AI for materials science.
Song Yuqi, Nasiri Alireza, et al. High-throughput discovery of novel cubic
crystal materials using deep generative neural networks. Adv Sci 2021;8:
e2100566.
[94] Ren Pengzhen, Xiao Yun, Chang Xiaojun, Huang Po-yao, Li Zhihui,
Chen Xiaojiang, et al. A comprehensive survey of neural architecture search: Siqi Shi obtained his B.S. and M.S. from Jiangxi Normal
challenges and solutions. ACM Comput Surv 2021;54:1e34. University in 1998 and in 2001, respectively. He finished
[95] Liu Yue, Zhu Kai, Liu Zitu. SSRNAS: search space reduced one-shot NAS by a his Ph.D. from Institute of Physics, Chinese Academy of
recursive attention-based predictor with cell tensor-flow diagram. In: 2021 Sciences in 2004. During this period, focused on the
international joint conference on neural networks; 2021. p. 1e8. electrolyte, electrode materials and relevant interfaces for
[96] Chen Chun-Teh, Gu Grace X. Generative deep neural networks for inverse lithium-ion batteries, he carried out the first-principles
materials design using backpropagation and active learning. Adv Sci 2020;7: calculation and design on the ionic transport physics,
1902607. cooperative electron/ion transport control problem
[97] Jonathan Frankle, and Michael Carbin. The lottery ticket hypothesis: finding earliest in China. After that, he joined the National Insti-
sparse, trainable neural networks. arXiv:180303635. tute of Advanced Industrial Science and Technology of
[98] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training Japan and Brown University of USA as a senior research
for high fidelity natural image synthesis. arXiv:180911096. associate until joining Shanghai University as a professor
[99] Ganguli Deep, Hernandez Danny, Lovitt Liane, Askell Amanda, Bai Yuntao, in early 2013. His current research interests focus on the
Chen Anna, et al. Predictability and surprise in large generative models. In: fundamentals and multiscale calculation of electrochemi-
2022 ACM conference on fairness, accountability, and transparency; 2022. cal energy storage materials and materials design and per-
p. 1747e64. formance optimization using machine learning.

816