Review of Materials Research 1 (2025) 100010

Contents lists available at ScienceDirect

Review of Materials Research

journal homepage: [Link]/journal/review-of-materials-research

AI4Materials: Transforming the landscape of materials science

and enigneering
Xue Jiang a, Dezhen Xue b, Yang Bai a, William Yi Wang c, Jianjun Liu d, Mingli Yang e,
Yanjing Su a, f, *
a
Beijing Advanced Innovation Center for Materials Genome Engineering, Institute for Advanced Materials and Technology, University of Science and Technology Beijing,
Beijing, 100083, China
b
State Key Laboratory for Mechanical Behavior of Materials, Xi'an Jiaotong University, Xi'an, 710049, China
c
State Key Laboratory of Solidification Processing, Northwestern Polytechnical University, Xi'an, 710072, China
d
State Key Laboratory of High Performance Ceramics and Superfine Microstructure, Shanghai Institute of Ceramics, Chinese Academy of Sciences, Shanghai, 200050,
China
e
Institute for Nanobiomedical Technology and Membrane Biology, Sichuan University, Chengdu, 610041, China
f
Chinese Materials Research Society, Beijing, 100083, China

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

Keywords: New materials, crucial for economic and technological progress, are prioritized globally with strategies to
Intelligent computation accelerate their advancement through big data and AI. AI for Materials (AI4Mater) serves as an overall framework
Machine learning for integrating AI into Materials Science and Engineering, which is structured around three main elements:
Materials data infrastructure
materials data infrastructure, AI4Mater techniques, and applications. This article reviews the development pro-
Autonomous experiment
Intelligent manufacture
cedure and recent innovations in materials data infrastructure, machine learning in materials, autonomous
experiment, intelligent computation, and intelligent manufacture. These efforts aim to foster open access to AI
resources and enhance the collective advancement of materials science, ultimately accelerating breakthroughs
and elevating the engineering application of new materials in a sustainable manner.

1. Introduction development, transform industry practices, and overcome inefficiencies

arross the field. This article outlines the objectives of AI4Mater, reviews
The development of novel materials is a cornerstone of technological recent advancements in materials data infrastructure, machine learning,
innovation and economic growth, driving transformative advances intelligent computing, autonomous experimentation, and smart
across diverse industrial sectors. Governments and industries worldwide manufacturing. It also proposes a road map for global collaboration,
are investing in strategies that integrate big data and artificial intelli- including the development of unified materials data standards, inte-
gence (AI) to accelerate material innovation. This has led to the emer- grated and shared data resources, open access to AI models and
gence of AI for Materials (AI4Mater), a rapidly evolving interdisciplinary computational tools in materials science.
field that combines computational design, data science and experimental
techniques. The integration of design, production, and application 2. Definition and motivation
through AI, combined with its ability to enhance computational simu-
lations and experimental workflows, is fundamentally reshaping how 2.1. Definition of AI4Mater
new materials are created and utilized [1]. The fusion of machine
learning and materials computation addresses challenges in cross-scale AI4Mater refers to an integrated framework for applying AI into
calculations, while AI-driven automation in experimental setups in- materials field. It is structured around three main elements: materials
creases the efficiency of material discovery. Breakthroughs in AI and data infrastructure, AI4Mater techniques and applications domains, as
robust data infrastructures will accelerate material research and illustrated in Fig. 1. At the foundation, the materials data infrastructure

* Corresponding author. Beijing Advanced Innovation Center for Materials Genome Engineering, Institute for Advanced Materials and Technology, University of
Science and Technology Beijing, Beijing, 100083, China.
E-mail address: yjsu@[Link] (Y. Su).

[Link]
Received 1 April 2025; Received in revised form 5 April 2025; Accepted 5 April 2025
Available online 6 April 2025
3050-9130/© 2025 The Authors. Published by Elsevier B.V. on behalf of Chinese Materials Research Society. This is an open access article under the CC BY-NC-ND
license ([Link]
X. Jiang et al. Review of Materials Research 1 (2025) 100010

such as regression, classification, clustering, dimensional reduction,

active learning, and reinforcement learning. During this period, the focus
was primarily on building foundational knowledge and developing
methodologies to analyze and interpret material datasets effectively.
Phase 2 (2016–2020) was the emergence of a new data-driven
paradigm in materials research. This phase was marked by the rapid
adoption of data-centric approaches to material composition design,
process optimization, and property prediction and improvement. Re-
searchers began to mine structure-property relationships using large
datasets, leading to more efficient materials discovery and development
processes. As data generation being accelerated, the ability to leverage
this information for informed decision-making became a crucial aspect of
materials research.
Phase 3 (2020-Present) has been defined by an even deeper inte-
gration of AI into materials R&D. During this phase, intelligent
computing, autonomous experimentation, big data analytics, and
generative AI have collectively transformed the landscape of materials
science. This integration represents a significant shift toward the fifth
R&D paradigm in materials, emphasizing not only efficiency and opti-
mization but also approaches to theory innovation and materials dis-
covery. As AI technologies continue to evolve, their application in
materials science promises to further revolutionize the field, enabling the
rapid acceleration of material innovation and tailored solutions for
various applications.
Fig. 1. The components of AI4Mater.
This evolution of AI in materials research highlights a continuous
journey of integrating fragmented technologies into a cohesive system.
The application of AI technology in materials computation, experimen-
supports the system through tools for data processing, data repositories,
tation, and production has significantly enhanced the intelligence level of
e-collaborative platforms, and standards and protocols. Built on this
computational design, experimental research, and manufacturing pro-
foundation, AI4Mater integrates machine learning, autonomous experi-
cesses. Leveraging its robust capability to handle high-dimensional data,
mentation, intelligent computation and intelligent manufacturing. The
AI effectively integrates computational, experimental, and production
techniques enable applications accross material science, materials R&D,
technologies through data-driven approaches. By establishing a closed-
manufacturing, and performance evaluation, enhancing innovation, ef-
loop feedback iteration and autonomous optimization across the mate-
ficiency and scalability throughout the materials value chain.
rial design-experimentation-production cycle, AI dramatically improves
the efficiency of new material development and engineering applica-
2.2. History of AI4Mater tions. From the initial application of basic machine learning algorithms to
the sophisticated AI-driven approaches of today, each phase has built
Over the past decade, AI has significantly advanced the landscapte of upon the previous one, fostering a more integrated and effective R&D
materials science, evolving from isolated apllications of isolated algo- framework. The gradual unification of AI tools and methodologies has
rithms to a integrated, data-driven paradigm, as shown in Fig. 2. This laid the foundation for a new era of materials innovation, where AI not
progress can be divided into three phases: only supports but drives the discovery and development of new mate-
Phase 1 (Before 2016) was characterized by an increasing awareness rials. This transition sets the stage for the next phase, where AI will play
of new technologies, as researchers began to understand and employ an even more critical role in addressing the challenges and opportunities
various machine learning algorithms and tools. This included techniques in materials science.

Fig. 2. The history of AI4Mater.

2
X. Jiang et al. Review of Materials Research 1 (2025) 100010

2.3. Motivation for AI4Mater In 2011, the Materials Genome Initiative (MGI) was launched as part
of the Advanced Manufacturing Partnership, aiming to halve the time
The discovery and development of advanced materials lie at the heart and cost of new materials development through high-throughput
of technological progress, enabling breakthroughs in renewable energy, computation, high-throughput experimentation, and materials data
aerospace, high-performance electronics, and beyond. However, con- infrastructure. Over the years, MGI has played a pivotal role in shaping
ventional approaches—dominated by trial-and-error experimentation the U.S. materials innovation ecosystem, fostering collaboration among
and computational simulations—are slow, costly, and often constrained government agencies, research institutions, and industries. The initia-
by existing theoretical frameworks. These limitations hinder innovation tive's Strategic Plan updates in 2014 and 2021 reinforced investments in
and delay critical solutions to global challenges such as energy sustain- AI-driven materials discovery, advancing intelligent materials R&D.
ability and resource scarcity. AI4Mater is reshaping this paradigm by Recognizing AI's transformative potential, the U.S. National Academy of
accelerating innovation, reducing risks, generating new scientific in- Sciences released the 2019 Materials Research Frontiers: A Decadal
sights, enabling disruptive material applications, and bridging the gap Survey, highlighting AI and data science as fundamental tools for over-
between materials design and manufacturing. coming inefficiencies in materials development. More recently, the 2024
Turning the Impossible into Reality: Unprecedented R&D Effi- Supercharging Research report by the President's Council of Advisors on
ciency. AI dramatically enhances research efficiency by predicting ma- Science and Technology identified A Phase Change for Materials Dis-
terial properties with high accuracy before experimental validation, covery and AI for Designing Advanced Semiconductors as key areas
minimizing costly trial-and-error processes. Data-driven insights enable where AI can accelerate scientific breakthroughs, underscoring its role in
the rapid identification of promising compositions, while AI-guided next-generation materials innovation.
simulations and automated experimentation optimize synthesis path- Europe has long been at the forefront of data-driven materials
ways. These advancements unlock previously inaccessible material research. In 2011, the Accelerated Metallurgy (AccMet) program,
spaces, making discoveries feasible at an unprecedented pace. managed by the European Space Agency (ESA) and involving 31 Euro-
Turning Hesitation into Confidence: Risk Reduction Across the pean institutions, pioneered high-throughput alloy design and simula-
R&D Chain. AI mitigates risks throughout the materials development tion, significantly reducing materials R&D timelines and costs. In 2012,
pipeline by providing predictive insights into material performance, the Metallurgy Europe initiative (2012–2022) identified 17 priority
manufacturability, and long-term stability. Early-stage screening iden- material needs and 50 cross-industry metallurgical research themes,
tifies unviable pathways before costly investments, while AI-driven covering applications in clean energy, green transportation, healthcare,
process optimization enhances yield, reduces waste, and ensures sus- and next-generation manufacturing. The European Commission's Hori-
tainability. Real-time process control further enables cost-effective, zon 2020 research and innovation framework (2014–2020) further in-
scalable production with minimized economic and environmental risks. tegrated research efforts across EU member states, promoting cross-
Unraveling the Unknown: Fundamental Breakthroughs in Ma- disciplinary collaboration and accelerating the adoption of AI-driven
terials Science. AI is not just an accelerant—it is a knowledge generator. materials innovation. Meanwhile, Germany's Industry 4.0 strategy,
By integrating multi-scale simulations, experimental data, and theoret- introduced in 2013, emphasized the use of IoT-driven information sys-
ical models, AI uncovers hidden correlations and extracts interpretable tems to digitize and optimize production processes, laying the foundation
design principles. Explainable machine learning sheds light on complex for AI-enhanced smart manufacturing and materials design.
structure-property relationships, driving new scientific theories and The UK has taken a decisive step toward AI-driven materials inno-
expanding our fundamental understanding of materials. vation with the launch of the National Materials Innovation Strategy in
From Discovery to Disruption: Revolutionizing High- January 2025. Drafted by the Henry Royce Institute—the UK's national
Performance Materials. AI-powered innovation paves the way for the hub for materials science—the strategy reflects input from over 2,000
development of next-generation materials with unprecedented functio- experts across 270 organizations. It prioritizes sustainability, energy ef-
nality—superior strength, extreme resilience, enhanced conductivity, or ficiency, and healthcare applications, aiming to position the UK as a
tailored surface properties. By accelerating the ideation-to- leader in advanced materials research. By fostering collaboration be-
implementation cycle, AI enables disruptive breakthroughs in energy tween academia, industry, and government, the strategy seeks to accel-
storage, lightweight composites, self-healing materials, and beyond. erate the translation of cutting-edge research into real-world
Bridging Materials and Manufacturing: Lifecycle-Centric Mate- applications. Key initiatives include the establishment of AI-powered
rial Management. AI seamlessly integrates materials design with materials innovation hubs, the development of high-throughput
manufacturing, ensuring end-to-end lifecycle management. From raw computational models and intelligent material databases, and the
material selection to process optimization, from real-time monitoring to deployment of AI-optimized experimental platforms. These efforts aim to
recyclability assessment, AI enables smart, adaptive control over the enhance the UK's global competitiveness in materials science while
entire materials value chain. This convergence enhances sustainability, addressing critical challenges such as climate change and resource
facilitates digital twin applications, and ensures materials innovation scarcity.
translates into scalable, real-world impact. At the international level, the International Science Council (ISC)
With these powerful capabilities, AI4Mater is transforming how we released the 2025 edition of Preparing National Research Ecosystems for
discover and design materials. More importantly, it bridges the gap be- AI, a report aimed at guiding nations in adapting their research infra-
tween innovation and manufacturing, ensuring that breakthroughs in the structure to AI's transformative potential. Published from ISC head-
lab translate into practical, scalable solutions. AI4Mater is not only quarters in Paris, the report provides actionable recommendations for
accelerating progress—it's redefining the boundaries of what science and integrating AI into national research frameworks, fostering interdisci-
technology can achieve. plinary collaboration, and addressing ethical, regulatory, and infra-
structural challenges. By showcasing best practices and case studies from
3. Foundations and frontiers: Global and domestic perspectives leading nations, the ISC underscores the importance of building inclu-
sive, equitable, and sustainable AI-driven research ecosystems to address
3.1. Global perspective global challenges while ensuring responsible AI deployment.

The integration of AI and data-driven approaches has become a 3.2. Domestic perspective
defining force in materials science worldwide, with major economies
implementing national strategies to accelerate discovery, design, and China has simultaneously developed targeted national strategies to
application of advanced materials. harness AI's potential in materials science. The Chinese government has

3
X. Jiang et al. Review of Materials Research 1 (2025) 100010

placed significant emphasis on the development of new materials, these databases primarily offered data hosting and search services,
recognizing their strategic importance for technological advancement enabling scientists to query and utilize data effectively. Additionally,
and industrial competitiveness. In 2015, it launched the national key they encouraged materials scientists to share their data within a broader
R&D program Key Technologies and Supporting Platforms for Materials community, including platforms like PAULING FILE, Granta Design,
Genome Engineering, with the ambitious goal of halving both the R&D MatNavi, and Springer Materials. Following 2011, the U.S. Materials
cycle and costs — commonly referred to as the “two halves” strategy. This Genome Initiative fostered the development of a data-driven model for
program focuses on establishing three major demonstration platforms to materials innovation, marking a transition into a new phase for materials
support research and innovation in materials genome engineering: high- databases. They evolved into data centers that provide raw data,
throughput computation, high-throughput synthesis and characteriza- computational simulations, and data analysis services, ultimately trans-
tion, and specialized materials databases. forming into materials discovery platforms that facilitate the R&D of new
To validate the effectiveness of these approaches, the program has materials.
facilitated demonstration applications across five key material cate- This shift in demand has raised new requirements for the storage and
gories: energy materials, biomedical materials, rare earth functional management of materials data. The process of constructing materials
materials, catalytic materials, and special alloys. These efforts have led to databases is an abstraction process aimed at addressing materials science
significant breakthroughs, reinforcing the applicability and impact of issues, and this process has traditionally involved two steps: (1) the
materials genome engineering in accelerating new materials develop- abstraction from the real world to the information world and (2) the
ment. In 2020, further initiatives were launched to promote the adoption abstraction from the information world to the machine world. The first
of materials genome engineering technologies, driving their transition step of abstraction is carried out collaboratively by materials scientists
from fundamental research to industrial application. and database engineers. Materials scientists create conceptual models
More recently, in 2024, China introduced the Construction Plan for that can describe the entities, properties, and relationships of materials,
the New Materials Big Data Center, aiming at integrating and developing while database engineers formalize and impose constraints on these
material data resources to establish a robust foundation for AI-driven conceptual models to ultimately develop a logical model of the infor-
materials research. These initiatives have propelled advancements in mation world. The second step of the abstraction process involves data-
fundamental materials genome engineering research, the development of base engineers using database management systems to convert the
key equipment and core software, and the engineering application of logical model into a software-supported physical model, such as MySQL,
intelligent materials technologies. The growing synergy between mate- Oracle, PostgreSQL, and MongoDB. Driven by AI and machine learning
rial data resources and AI has played a crucial role in shaping new par- technologies, the demand for the integration of materials data has
adigms for intelligent materials R&D, paving the way for innovative intensified, leading to increasing functional requirements for heteroge-
breakthroughs. neous data patterns. This makes the traditional approach of predefining
In parallel, leading innovation hubs such as Beijing and Shanghai materials data schemas for database construction obsolete for several
have unveiled their AI þ Materials Action Plans, marking a pivotal step in reasons: (1) frequent and close communication between materials sci-
integrating AI with materials science. These initiatives seek to accelerate entists and database engineers raises the design and construction barriers
the discovery, design, and application of advanced materials by of material databases; (2) changes to logical and physical data storage
leveraging AI-driven technologies. By establishing dedicated AI-powered models must be completed by database engineers, resulting in low
materials innovation hubs, both cities are prioritizing the development of database scalability; and (3) accessing the database through applications
high-throughput computational models, intelligent material databases, requires prior understanding of the database's design patterns, creating
and AI-empowered experimental platforms. Furthermore, these efforts difficulties in data exchange and integration.
emphasize cross-industry collaboration, bringing together academia, As a result, the field of materials databases has begun to develop
research institutions, and enterprises to build a dynamic ecosystem for templated database systems that allow for user-defined data descriptions,
AI-powered materials innovation. As these initiatives gain momentum, gradually becoming the foundation for the open interconnection and
they are set to redefine the landscape of materials research, driving seamless sharing of multiple data resources and applications. By
technological breakthroughs and industrial transformation in the era of providing reusable data types for template customization, these systems
intelligent materials. meet the personalized data expression needs of different engineering
applications, groups, institutions, and individuals, allowing for the post-
4. Prior efforts on innovation definition and standardized storage of material data schemas. Examples
include the Materials Data Curation System in the United States and
4.1. Materials data infrastructure MGEDATA in China.
The Materials Data Curation System (MDCS) [2], developed by Na-
With the rapid growth of big data and AI technologies, the demands tional Institute of Standards and Technology (NIST) in the U.S., offers
on materials databases have shifted significantly, evolving from simple reusable general base types, including double-precision floating-point
data hosting platforms to dynamic, data-driven systems. This trans- types, unit types, and uncertainty types. It allows users to combine these
formation has necessitated a new approach to database design and base types to create new complex data types, forming customized ma-
management, as traditional methods of predefining schemas and rigid terials data templates for data uploads. The data is stored in a document
database structures no longer meet the needs of heterogeneous and database in the format of Extensible Markup Language (XML). MGEDATA
multimodal materials data. Instead, templated database systems that [3,4], developed by Researchers from University of Science and Tech-
allow for user-defined data descriptions and flexible data integration are nology Beijing developed, is a materials database system. It has devel-
emerging as the new standard. This shift facilitates seamless data sharing, oped dynamic container system that supports user-defined data storage
interconnection, and post-definition of material data schemas, ultimately structures, providing 10 advanced general data types, including string,
driving forward the open collaboration and innovation in materials array, table, container, and generator types (Fig. 3). Users can combine
science. these types to create custom data templates, and the data is stored in
Binary JSON (BSON) format within a MongoDB document database,
4.1.1. Repositories fulfilling the description and storage needs for multi-source heteroge-
With the advancement of big data and AI technologies, the charac- neous materials data. Fig. 3 illustrates the graphical user interface of the
teristics of material data—such as heterogeneity and multi- dynamic container system.
modality—have become increasingly prominent. Consequently, the goals
of materials databases are undergoing significant changes. Prior to 2010,

4
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Fig. 3. The graphical user interface of the dynamic container system of MGEDATA.

4.1.2. E-collaborative platforms datasets. This platform facilitates the sharing of computational science
An e-collaboration platform is a web-based environment that hosts resources driven by materials modeling workflows, creating a new ma-
open-source software tools, enhancing collaboration and providing terials data ecosystem that integrates data publishing, storage, discovery,
distributed access to experimental and computational data. These plat- resource registration, online development, model integration, and
forms allow for the tracking of collaborative workflows, making it easier application feedback.
to monitor progress and contributions. Specifically designed to boost the In 2021, China lunched the National New Materials Database Platform
productivity of cross-disciplinary project teams, open-source, open-ac- (Demonstration) to explore a distributed construction concept of decen-
cess e-collaboration environments facilitate the seamless sharing and tralized development and centralized sharing of materials data resources.
exchange of information, knowledge, and expertise among team mem- It has created a prototype of a digital collaboration platform that integrates
bers who are geographically and organizationally distributed across data accumulation, publishing and sharing, integration, retrieval and
various stakeholder groups. discovery, intelligent analysis, and design applications, providing cloud-
Since 2002, with the support of the National Science Foundation, based self-service analysis, tool integration, application development,
Purdue University has established the Network for Computational and rapid delivery services for materials R&D. The design workflows
Nanotechnology (NCN) and developed the nanoHUB infrastructure [5]. provide graphical, customizable materials data processing and analysis
nanoHUB provides a self-service analysis, application development, and processes, meeting the needs for low-code materials machine learning and
rapid delivery system, featuring a rapidly growing collection of simula- data mining. Additionally, the platform incorporates an online coding
tion tools. It supports cloud-based operations and internet access, offer- environment using Jupyter Notebook and offers a containerized environ-
ing a user-centered, scientific end-to-end cloud computing environment. ment for materials design software and tools integration.
This platform serves as an open and free resource for computational
research, education, and collaboration in nanotechnology, materials 4.1.3. Standards and protocols
science, and related fields. Since 2000, there has been an international effort to initiate and
The U.S. government has effectively harnessed the expertise of NIST promote standardized protocols for materials data. MatML, an XML
in data integration and management. Central to this initiative, NIST has specifically developed for the storage and exchange of materials infor-
partnered with nearly a hundred research institutions and companies, mation, was initiated by NIST in 2000 [6]. MatML defines a coherent and
including Northwestern University. By consolidating materials data, consistent document structure for its tags, ensuring that any program-
codes, computational tools, and various software and hardware resources ming language can parse and process the data in any required manner,
from multiple entities, they have developed a series of materials database thereby enhancing applications’ understanding and operability with the
systems, data service platforms, repositories, and resource registration data. Initially, NIST established a working group to describe and develop
systems. Notable examples include the MDCS, the Materials Data Facility the scope and specifications of the new MatML language, and released
(MDF), the Open Quantum Materials Database (OQMD), the NIST Ma- draft documents of the Annotated MatML DTD (versions 1.0, 2.0, and
terials Resource Registry (NMRR), and the CHiMaD Phase Field. Through 3.0) for public comment. In 2003, the MatML Coordination Committee
the materials resource registration platform, they facilitate the registra- was formed, led by ASM International and the Materials Information
tion of multi-source heterogeneous data, establish data exchange pro- Society, comprising experts from academia and industry. The standard-
tocols, and bridge the gap between data resources, databases, software, ization of MatML is currently managed by the Organization for the
and end users. This initiative paves the way for materials data integration Advancement of Structured Information Standards (OASIS), with a
and application both in the U.S. and globally, promoting broader sharing dedicated schema development working group responsible for the
and reuse of materials data. development and maintenance of the MatML markup language.
After 2011, propelled by the European Union's “Accelerating Mate- MARVEL, in collaboration with 31 leading international materials
rials” and “Metallurgy Europe” research programs, the field of compu- data platforms, including AFLOW, the Materials Project, and NOMAD,
tational materials data underwent rapid development. This initiative created the Open Databases Integration for Materials Design (OPTI-
brought together computational resources and research teams from MADE) international alliance. This initiative introduced a universal
multiple European countries to establish the NOMAD database. Swit- application programming interface (API) for accessing and exchanging
zerland's National Center for Materials Computing, Design, and Discov- materials databases, promoting the FAIR principles for global materials
ery (MARVEL) also developed the Materials Cloud platform, offering a data—ensuring that data is Findable, Accessible, Interoperable, and
comprehensive solution built on open-source workflows and open Reusable [7].

5
X. Jiang et al. Review of Materials Research 1 (2025) 100010

In 2019, the Zhongguancun Materials Testing Technology Alliance in in materials science, encouraging researchers to explore more advanced
China released the group standard “General Principles for Materials methods for materials prediction and discovery [7]. For example, to
Genome Engineering Data”, which standardized the information that efficiently screen stable ternary compounds, researchers at Northwestern
must be collected during the data generation process. In October 2024, University employed ensemble learning to develop a model capable of
under the standardization system of the China Materials and Testing predicting formation energies. This model successfully identified 4,500
Standardization Committee (CSTM), the “Materials Data Standardization new stable materials from a pool of 1.6 million candidate compositions,
Field Committee (CSTM/FC93)” was established. This committee aims to achieving computational speeds six orders of magnitude faster than
develop universal technical standards, protocols, and rules covering density functional theory (DFT) calculations [13]. A landmark 2016
various aspects of materials data, including storage, identification, study published in Nature demonstrated the innovative application of
evaluation, integration, exchange, sharing, and management. support vector machines (SVMs) to guide chemical reaction design using
Simultaneously, blockchain systems possess features such as immu- both successful and failed experiments. This approach significantly
tability, complete traceability, retrievability, transparency, and collective reduced experimental workload and improved predictive accuracy,
maintenance. In the process of sharing and exchanging materials data, illustrating the practical benefits of machine learning in real-world ma-
blockchain technology can be utilized to monitor data activities. By terials research [14].
combining consensus mechanism algorithms and smart contract, a In recent years, machine learning-driven materials research has also
complete record of the data lifecycle can be established, enabling flourished in China, resulting in the development of a range of high-
traceable and trustworthy data access, retrieval, and application. This performance advanced materials. At Shanghai University, researchers
approach has become an important data exchange model. Based on the employed Random Forest (RF) and symbolic regression to predict the
records of data exchanges, dynamic evaluations of data resource quality properties of amorphous metal alloys, successfully estimating their glass-
and reliability can be implemented, promoting a healthy ecological cycle forming ability and elastic modulus with strong predictive performance
of data. and generalizability [15]. Similarly, researchers at Northeastern Uni-
Materials databases and E-collaborative platforms have undergone versity applied a general machine learning framework to predict the
substantial transformations to keep pace with evolving demands in the tensile properties of Reduced Activation Ferritic/Martensitic (RAFM)
field. The integration of AI, machine learning, and big data technologies steels, demonstrating improved accuracy and broader applicability
has highlighted the need for flexible, user-customizable data systems. compared to traditional physical metallurgy models [16]. At Southeast
These systems support the creation of dynamic, user-defined data tem- University, researchers applied machine learning to screen perovskite
plates, enabling more efficient storage, retrieval, and sharing of complex, materials, ultimately using a gradient boosting classifier to select stable
heterogeneous materials data. E-collaborative platforms have further candidates from 19,841 compositions. Combined with a downstream
expanded the scope of materials R&D by offering cloud-based environ- regression model, they identified 151 promising ferroelectric photovol-
ments for seamless data exchange and collaboration. Meanwhile, inter- taic (FPV) perovskites with optimal bandgaps [17].
national efforts to standardize materials data have promoted global Today, machine learning models for materials science increasingly
interoperability. The combination of these advancements, along with the need to evolve into deployable tools that can seamlessly integrate into
integration of blockchain technology for secure and traceable data research workflows. A collaboration including the University of Chicago
sharing, is fostering a more open, collaborative, and efficient ecosystem developed Matminer, an open-source Python software platform that
for materials discovery and innovation. provides modular tools to retrieve large datasets from external databases,
perform feature extraction, and generate interactive charts for visuali-
4.2. Machine learning in materials science zation. Matminer integrates closely with existing Python machine
learning and data analysis packages, such as scikit-learn and pandas,
The rapid advancement of AI is transforming materials science, streamlining data-driven materials research [18]. To further simplify the
accelerating the discovery, design, and optimization of advanced mate- adoption of machine learning in materials science, Shanghai University
rials. Traditional trial-and-error experimentation and computational created the machine learning for materials design (MLMD) platform,
simulations often struggle with the complexity of modern materials providing programming-free AI tools for material designers [19].
systems. AI offers a powerful alternative by uncovering hidden patterns, Despite the successes, challenges remain. In particular, ML models
predicting complex property relationships, and efficiently navigating still struggle to generalize across diverse materials systems and cope with
vast design spaces. This section explores how traditional machine the vast, complex candidate spaces inherent to materials discovery.
learning, deep learning, heuristic algorithms, generative AI, knowledge- Future advancements will likely focus on integrating machine learning
guided models, and large language models are reshaping materials with high-throughput experimental and computational techniques, and
research, providing new pathways for innovation across the entire ma- developing adaptive models capable of learning from new data in real
terials development lifecycle. time while continually refining their predictions.

4.2.1. Conventional machine learning approaches for materials 4.2.2. Heuristic algorithms and active learning for materials searching
The discovery and optimization of advanced materials is essential for The discovery of novel materials often involves navigating vast search
technological innovation across industries. However, traditional process spaces and addressing multi-objective optimization challenges. To tackle
for discovering new materials has largely relied on labor-intensive trial- these complexities, researchers have increasingly adopted heuristic al-
and-error experimentation, significantly slowing process. Machine gorithms and active learning techniques. These approaches have proven
learning (ML) has transformed this landscape by enabling faster, more particularly effective due to their global search capabilities and efficient
accurate predictions of material properties, thereby accelerating the exploration of experimental spaces.
development process and reducing reliance on costly experimentation Heuristic algorithms, such as Genetic Algorithms (GA), are rule-based
[8–10]. methods that efficiently explore large search spaces to identify promising
Regression models were among the earliest ML tools applied in ma- candidates when exhaustive searches are impractical.
terials science, particularly for predicting straightforward properties. Among heuristic algorithms, stand out for their robust global search
Early efforts employed simple artificial neural networks to develop capabilities. By mimicking the evolutionary processes of natural selec-
regression models capable of predicting properties, such as, the me- tion, GA excels at identifying optimal combinations. For instance, re-
chanical properties of ceramic matrix composites and the glass transition searchers at the University of Ottowa applied GA to optimize functional
temperature of polymers [11,12]. The launch of the Materials Genome groups in metal-organic frameworks (MOFs), significantly enhancing
Initiative (MGI) in 2011 catalyzed a rapid expansion of ML applications CO2 adsorption capability [20]. Their method screened 141

6
X. Jiang et al. Review of Materials Research 1 (2025) 100010

experimentally characterized MOFs and successfully identified de- large chemical databases [28,29]. More recently, Shanghai University
rivatives with improved performance with CO2 adsorption capacity and Central South University have continued to expand the applications
increased by over 400 %. This successful application highlighted the of BO, addressing a wide range of materials design challenges [29,30].
power of GA in handling complex, high-dimensional materials optimi- Despite their proven advantages, both GA and BO face ongoing
zation tasks. challenges. These include the need for large number of iterations in
In parallel, active learning, particularly within the Bayesian Optimi- highly complex systems and the difficulty of managing high-dimensional,
zation (BO) frame work, has become a powerful tool for materials dis- multi-component materials spaces. However, emerging approaches that
covery, especially in data-scarce experimental settings. BO builds integrate these methods with advanced machine learning techniques,
predictive models that actively guide experimentation by identifying the such as deep learning-based surrogate models, are expected to further
most informative candidate materials to investigate next, thereby mini- enhance their efficiency and predictive power. By providing effective
mizing the number of required experiments or calculations. Researchers strategies for exploring vast design spaces and guiding experimental
from Kyoto University employed support vector regression (SVR) within campaigns, heuristic algorithms and active learning are set to remain
a BO framework, together with density function theory (DFT) calcula- essential tools in accelerating the discovery of next-generation materials.
tions to optimize thermal conductivity from vast candidate pools [21].
Researchers at Xi'an Jiaotong University advanced the application of BO 4.2.3. Deep learning for complexities in materials
for material property prediction and optimization. Their work introduced The aforementioned conventional ML methods, while effective for
an adaptive design framework to balance exploration and exploitation specific tasks, often struggle to handle the high-dimensional, complex,
(as shown in Fig. 4(a) and (b)), incorporating machine learning predic- and large-scale data sets seen in materials science. Deep learning,
tion, uncertainty quantification and iterative experimental validation. It inspired by the biological neural networks, has emerged as a trans-
ensures efficient searches across vast material design spaces while formative tool that can automatically extract high-level features and
maintaining robust performance even in data-limited conditions capture intricate nonlinear relationships through multi-layer neural
[22–26]. network architectures [31,32]. By leveraging this ability, deep learning
In the realm of alloy design, University of Science and Technology offers new solutions to some of the most computationally challenging
Beijing researchers leverage BO to optimize the composition and pro- problems in materials research.
cessing of Cu-Ni-Co-Si alloys, aiming to reduce reliance on expensive A pioneering example of deep learning's impact in materials science is
[27]. Their BO-guided optimization process required only 32 experi- the Deep-learning DFT Hamiltonian (DeepH) developed by Tsinghua
ments to refine deformation-aging parameters, resulting in an alloy with University. This approach directly learns from DFT (Density Functional
less than half the cobalt content of ASTM C70350 while maintaining Theory) data to predict Hamiltonian quantities (as shown in Fig. 4(c)),
comparable mechanical performance and electrical conductivity. Build- enabling efficient and accurate electronic structure calculations across a
ing on these successes, Tsinghua University applied BO to automate broad range of materials [33]. By significantly reducing the computa-
materials discovery workflows, successfully optimizing CO2 adsorption tional cost while maintaining accuracy, DeepH has been extended to
processes and screening for high-performance photosensitizers from applications such as magnetic superstructures and hybrid density

Fig. 4. (a) [22] and (b) [24] for Bayesian Optimization; (c) [33] and (d) [40] for the application of deep learning in first-principles electronic structure calculation and
crystal structure assignment respectively.

7
X. Jiang et al. Review of Materials Research 1 (2025) 100010

functional calculations, demonstrating its adaptability and versatility in analysis that enhances both efficiency and accuracy. This method has
electronic structure analysis [34–36]. proven particularly valuable for tracking the structural evolution of
Deep learning has also transformed how researchers identify and materials during synthesis and application processes [37,38]. Further
analyze microscopic crystal structures. At Shenzhen Graduate School of enriched by AI techniques, this graph-based methodology has been
Peking University, researchers integrated graph-theoretic mathematics applied to lithium solid-state battery fast-ion conductors and X-ray
with structural chemistry, creating a novel approach for structural diffraction (XRD) analysis, unlocking new opportunities for crystal

Fig. 5. (a) for GAN to copper-based metallic glasses [49]; (b) for CVAE to design SMAs [52]; (c) for physical metallurgy-guided ML approach [66]; (d) for physical
mechanism knowledge integrated method [67].

8
X. Jiang et al. Review of Materials Research 1 (2025) 100010

structure mining and materials discovery, as shown in Fig. 4(d) [39,40]. Generative AI has also proven valuable in designing functional ma-
Oriented towards property-oriented materials design, researchers at terials such as high-temperature superconductors and metamaterials.
Shanghai Jiao Tong University introduced a deep transfer learning Researchers at Harbin Institute of Technology, Shenzhen combined VAE
framework for predicting adsorption energies, a critical factor for un- and GAN techniques to systematically generate undiscovered supercon-
derstanding physical and chemical reactions at nano-interfaces. This ductors under specific high Tc (critical temperature) constraints [48].
method drastically reduces computational costs, and it has already been They later expanded this work by integrating diffusion models and a
applied areas such as lithium-sulfur battery cathode materials and heavy doping-adaptive representation into a three-channel matrix framework,
metal ion adsorption processes [41,42]. In addition, by applying graph which proposed 200 potential high-Tc superconductors not yet docu-
neural networks (GNNs) to model adsorption energies across various 2D mented in the literature [53,54]. Their method provides a powerful tool
materials, they identified promising catalysts for hydrogen evolution, for guiding the experimental discovery of new superconductors, opening
reducing the materials screening cycle by approximately ten years [43]. new possibilities for material design.
This illustrates how deep learning not only enhances accuracy but also In the field of metamaterials, which present even more complex
accelerates the pace of materials discovery. design challenges due to their intricate structural requirements, genera-
Despite these transformative advances, the application of deep tive AI has also shown promise. Researchers at Tsinghua University
learning in materials science still faces challenges, particularly the reli- introduced an active learning framework that integrates VAE, multi-
ance on high-quality, large-scale datasets, which are often scarce for objective optimization, and 3D printing to design orthopedic implants
novel materials. Additionally, the black-box nature of many deep with tailored microscale architectures [55]. Researchers at Huazhong
learning models limits their physical interpretability, creating a gap be- University of Science and Technology developed a VAE-based generative
tween prediction and scientific understanding. Addressing these limita- design framework for thermal metamaterials, efficiently mapping topo-
tions will require the development of more efficient and interpretable logical unit cells to desired thermal properties [56]. These innovations
deep learning frameworks, which embed materials domain knowledge illustrate how generative AI can bridge the gap between structural to-
into model architectures. Nevertheless, deep learning has already pology optimization and functional property design, unlocking new op-
demonstrated exceptional potential to handle the complexity, scale, and portunities for advanced metamaterial engineering.
nonlinearity inherent in materials data, and as these technologies Generative AI represents a paradigm shift in materials science, of-
mature, they are poised to further accelerate the discovery, design, and fering powerful tools for the intelligent design of complex materials. As
optimization of new materials. these technologies continue to evolve, they promise to further accelerate
materials innovation, making the design process more efficient, accurate,
4.2.4. Generative AI for inverse designing new materials and aligned with real-world applications. However, generative AI still
Generative AI introduced a transformative approach to materials faces challenges, particularly regarding the availability of high-quality
design, enabling the inverse design of materials tailored to meet specific training data and the black-box nature of many generative models,
performance criteria. Unlike conventional ML, generative AI models which can hinder physical interpretability. Moving forward, the inte-
usually constructs and explores a latent design space, where materials gration of large language models (LLMs) for enhanced data curation and
performance constraints guide the discovery of new materials. Key ad- knowledge extraction, combined with expert domain input, will be
vances in generative models, such as Variational Autoencoders (VAE) critical for refining generative processes and ensuring that material de-
[44], and Generative Adversarial Networks (GAN) [45], have been signs are not only innovative but also practically viable.
pivotal in optimizing material composition and designing materials
structures. These foundational methods, along with techniques like 4.2.5. Fusion of materials knowledge with AI
Reinforcement Learning (RL) [46] and diffusion models [47], have As materials science continues to embrace data-driven methods, the
dramatically expanded the potential for capturing complex nonlinear need to combine domain-specific knowledge with AI techniques becomes
relationships within high-dimensional the material spaces. increasingly critical, especially when experimental data is scarce. While
Generative AI has found broad application in alloy and metallic glass generative AI and other machine learning methods excel at navigating
design. To address the vast compositional design space in multi-principal vast design spaces, they often lack the physical interpretability required
element alloys, researchers at City University of Hong Kong applied un- to fully understand complex material behaviors. To address this, re-
supervised GAN Boosted Trees to map composition-property relation- searchers are developing hybrid approaches that embed physical laws,
ships, advancing the inverse design of bulk metallic glasses (BMGs) [48]. expert knowledge, and theoretical models directly into AI workflows,
They later extended this approach by combining transfer deep learning enhancing both predictive accuracy and scientific reliability [57].
with generative and supervised learning models, successfully designing AI models are particularly effective at uncovering hidden patterns in
copper-based metallic glasses with exceptional hardness and low large datasets that often remain undetected by conventional methods
modulus, as shown in Fig. 5(a) [49]. Further, the team developed an [58]. However, prediction accuracy can be further improved by incor-
explainable AI framework integrating Conditional Variational Autoen- porating physical features and domain-specific descriptors into these
coders (CVAE) and Artificial Neural Networks (ANN), enabling the direct models. For example, researchers have successfully applied phase dia-
generation of eutectic compositionally complex alloys across quaternary grams and Landau-Devonshire theory to construct new descriptors that
to senary systems [50]. These advances provide a new paradigm for in- guide the discovery of high-performance ferroelectric materials,
verse designing high-performance metallic glasses and alloys. At Xi'an demonstrating how physical insights can directly improve the quality of
Jiaotong University, researchers combined Reinforcement Learning (RL) AI predictions [59,60].
and CVAE to efficiently explore large compositional spaces for alloy To handle the vast number of potential descriptors in materials sci-
design. They applied RL to design a TiNi-based phase change alloy with ence, feature engineering techniques have been increasingly employed to
the high transformation enthalpy, surpassing traditional design methods select, refine, and optimize descriptors before model training. At Uni-
through iterative optimization coupled with experimental validation versity of Science and Technology Beijing, researchers developed a
[51]. The team introduced a generative learning-enabled inverse design hybrid method combining feature engineering with genetic algorithms,
framework that translates hand-drawn target heat flow curves into which enabled the simultaneous optimization of machine learning model
tailored compositions and processing parameters for TiNi alloys, as selection and material descriptors. This method achieved high accuracy
shown Fig. 5(b) [52]. These works demonstrated the potential for inte- in predicting phase stability in high-entropy alloys (HEAs), including
grating human design intent directly into machine-guided materials both two-phase and three-phase systems [61]. They further proposed an
discovery workflows. innovative metric to quantify the efficiency of active learning (AL),

9
X. Jiang et al. Review of Materials Research 1 (2025) 100010

allowing the identification of optimal feature subsets, which helped structures and properties, leading to the discovery of new thermoelectric
accelerate the discovery of HEAs with exceptional compressive strength materials [74]. Researchers at New York University also adopted the
at elevated temperatures [62]. Another effective approach for identifying skip-gram algorithm on 6.4 million materials-related abstracts plus ab-
informative material descriptors comes from Shanghai University, where stracts on metallic materials, and successfully represented high-entropy
researchers applied the Sure Independence Screening and Sparsifying alloys by 200-dimensional word vectors. The approach identified the
Operator (SISSO) technique [63]. Using compressed sensing principles, representative FCC Cantor and BCC Senkov alloys as the most promising
SISSO efficiently identifies a sparse set of descriptors most relevant to the high-entropy alloys, long before they had been discovered and synthe-
target properties. This approach has already been applied to successfully sized [75].
predict perovskite stability and provided valuable insights into Furthermore, LLMs accelerate materials design by automating the
metal-oxide interactions, illustrating how mathematical sparsity tech- design, planning, and execution of complex experiments, thereby
niques can effectively complement materials knowledge in AI pipelines enabling autonomous research on materials. This is achieved through the
[64,65]. integration of AI agents that utilize memory retention, logical analysis,
Incorporating quantitative physical laws directly into machine task decomposition, and summarization capabilities. Together, these
learning models offers another path to improving both predictive accu- advancements highlight the significant impact of LLMs in advancing
racy and interpretability. Researchers at Northeastern University devel- materials science and facilitating innovative research directions. Re-
oped a physical metallurgy-guided machine learning model that searchers at Carnegie Mellon University developed an AI system driven
explicitly incorporates alloy composition, crystal structure, and phase by GPT-4 that autonomously designs, plans and performs complex
stability into its framework, significantly improving the accuracy of alloy chemical synthesis experiments of organic compounds [76]. With prompt
property predictions, as shown in Fig. 5(c) [66]. Beyond prediction, AI input, it can use tools to browse the internet and relevant documentation,
has also contributed to the refinement of physical laws themselves. For use robotic experimentation application programming interfaces and
example, the classical Hall-Petch relationship linking grain size to ma- leverage other LLMs for various tasks. Researchers from Korea Advanced
terial strength was enhanced by incorporating intrinsic factors identified Institute of Science and Technology built an AI system called ChatMOF to
through machine learning, resulting in more accurate predictions and manage data retrieval, property prediction, and structure generations
deeper insights into the mechanisms governing grain boundary tasks by leveraging a large-scale language model. ChatMOF shows high
strengthening, as shown in Fig. 5(d) [67]. In the search for novel mate- accuracy rates of 96.9 % for text-based data searching, 95.7 % for
rials, particularly in high-dimensional compositional spaces, traditional property predicting, and 87.5 % for structure generating tasks with
optimization techniques often get trapped in local minima. By integrating GPT-4 [77].
domain knowledge into AI search algorithms, researchers can constrain LLMs face significant challenges in directly supporting material
and guide the search process [59], steering the exploration towards design, particularly when it comes to quantitatively reasoning about
physically meaningful solutions. properties based on composition and processing conditions. This raises
The fusion of domain knowledge with AI has already demonstrated critical questions: How accurate can LLMs be in predicting material
success in enhancing both the predictive power and physical interpret- properties, and can they actually facilitate more efficient design and
ability of machine learning models, and accelerating materials discovery. optimization of materials? Researchers at Georgia Institute of Technol-
By combining physical insights, expert-crafted descriptors, and advanced ogy trained polyBERT using DeBERTa on a dataset comprising 100
algorithms, researchers can better navigate complex design spaces and million hypothetical polymers generated by exhaustively combining
uncover new materials more efficiently. As automated experimentation chemical fragments derived from a database of over 13,000 synthesized
platforms and LLMs continue to evolve, the seamless integration of expert polymers. Through training, polyBERT learns to convert input PSMILES
knowledge into AI workflows will become even more essential. Looking strings into numerical representations as polymer fingerprints, and a
ahead, this hybrid approach is expected to drive new breakthroughs, multitask machine learning framework then predicts polymer properties
enabling the rapid design of new materials tailored to meet specific [78]. Researchers from the University of Science and Technology Beijing
performance requirements [68,69]. have developed an innovative end-to-end pipeline that transforms
materials-related text into precise property predictions for steels,
4.2.6. Large language models leveraging the capabilities of an LLM called SteelBERT (Fig. 6). This
The advancement of natural language processing (NLP) and LLMs is model has been pretrained on an extensive corpus of 4.2 million abstracts
transforming the way we process and apply extensive materials litera- and 55,000 full-text articles in the field of materials science, specifically
ture, emerging as a hot topic in AI4Mater science. LLMs have been focusing on steel [79]. To enhance its predictive power, SteelBERT em-
effectively utilized for the automatic extraction of materials data from ploys a disentangled attention mechanism that enables it to offer an ac-
journal articles, focusing on key aspects such as chemical composition, curate representation of steel-related knowledge when given natural
properties, and synthesis procedures. As early as 2011, researchers began language input. When examining three key mechanical properties—yield
to recognize chemical entities in the chemical literature, such as Chem- strength, tensile strength, and elongation—the researchers achieved
DataExtractor [70]. Researchers at University of Science and Technology impressive determination coefficients (R2) of around 80 %. Furthermore,
Beijing proposed an automated NLP pipeline to capture the dataset they demonstrated the model's ability to be fine-tuned with a modest
covering chemical compositions, physical properties, and synthesis and laboratory dataset tailored to specific steel designs. With just 64 exper-
processing routes in alloys [71,72]. One notable method, ChatExtract, imental samples of austenitic stainless steels, their pipeline effectively
enables automated and precise data extraction from tables with minimal predicted a novel 15Cr austenitic stainless steel with yield strength of
setup, utilizing conversational LLMs. By leveraging carefully crafted 960 MPa, tensile strength of 1138 MPa, and elongation of 32.5 %, sur-
prompts, ChatExtract identifies, extracts, and verifies information passing those of all previously reported. This groundbreaking research
through follow-up questions [73]. highlights the potential of LLMs to tackle quantitative prediction chal-
Beyond merely extracting information, the insights found in materials lenges derived from natural language inputs, paving the way for ad-
science literature can be efficiently encoded into information-dense word vancements in material design across various steel types and processing
embeddings using LLMs. These embeddings can then be employed for routes.
new materials discovery and accurate property predictions. For example, The integration of AI into materials science has already delivered
researchers at Lawrence Berkeley National Laboratory have successfully substantial breakthroughs, ranging from the discovery of new materials
adopted unsupervised learning techniques, effectively encoding scientific to the refinement of physical models and the development of autonomous
knowledge from materials literature into information-dense word vec- design frameworks. As techniques such as generative AI, LLMs, and
tors. This approach has identified relationships between material knowledge-guided learning continue to mature, AI will not only enhance

10
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Fig. 6. Steel design based on a large language model [79].

Fig. 7. The workflow of autonomous experiments [80–84].

11
X. Jiang et al. Review of Materials Research 1 (2025) 100010

predictive accuracy and efficiency but also deepen our understanding of reaction parameters for experimental iteration. More than 600 experi-
complex material behaviors. Looking ahead, the seamless fusion of AI ments were conducted in an autonomous closed-loop mode, ultimately
algorithms, domain knowledge, and real-time experimental data will successfully producing carbon nanotubes at different preset growth rates.
unlock faster, smarter, and more sustainable materials innovation, The system also reveals a completely new synthesis mechanism and
ensuring that future materials are tailored to meet the ever-evolving discovers a new "kinetic trap" phenomenon.
demands of technology, energy, and industry. In 2021, an autonomous experimental system for additive
manufacturing based on ARES was released [87]. This system combined
4.3. Autonomous experiments syringe extrusion printing technology with cloud-based machine learning
optimization algorithms, autonomously adjusting printing parameters to
Autonomous experimentation is an innovative experimental method achieve direct writing of single-layer printing features that meet target
that integrates various advanced technologies. Its core lies in the high specifications. It employed automatic image analysis as a closed-loop
integration of AI, robotics, data technology, and optimization algorithms feedback mechanism for online Bayesian optimization, achieving the
with experimental equipment, enabling the automation and intelligence desired printing goals in fewer than 100 experimental iterations.
of the experimental process. Fig. 7 shows the workflow of autonomous Inspired by the work of ARES, some other autonomous experimental
experiments. The system first uses AI to analyze and process experi- systems are designed and developed in some universities for new mate-
mental data based on predefined goals and conditions, providing a basis rials research, which can be divided into categories of flow, table and
for the formulation of experimental plans. Then, robots perform the mobile according to the architecture of automation equipment. A typical
relevant experimental operations and collect data. The system analyzes example of a table system is developed by University of British Columbia
the results again and adjusts the experimental plan. This process con- in Canada, named Ada [88]. It is a modular robotic platform to accelerate
tinues in a feedback loop, automatically adjusting the material compo- the self-driving research and development of thin-film materials. This
sition and preparation process to explore new materials with better platform is driven by model-based optimization algorithms and can
performance. Additionally, human experts can interact with the platform autonomously optimize the optical and electronic properties of thin films
to provide domain knowledge and experience, guiding the optimization by modifying their composition and processing parameters. The research
of experimental plans. At the same time, the system can also feedback the team demonstrated the workflow and advantages of the Ada platform by
experimental results and analysis conclusions to the researchers, pro- optimizing the hole mobility of spiro-OMeTAD, an organic hole transport
moting mutual progress. The collaborative application of materials material used in perovskite solar cells. The Ada platform, based on a
computation, experiment and data further enhances the intelligence level modular robotic system, can automatically carry out preparation pro-
of the experimental system. On this basis, the integration of information cesses such as solution formulation, spin coating, thermal treatment, and
technology, the internet, big data, and other technologies enables the property testing. The platform can install corresponding experimental
modular collaborative scheduling and management of autonomous ex- modules based on specific experimental workflows, making it easily
periments, achieving efficient management and utilization of experi- adaptable to various material research scenarios. The Ada platform fea-
mental and computational resources, thereby improving experimental tures fully automated processes, allowing for the rapid generation of
efficiency and quality. high-quality, well-organized datasets. It also enables the control or
optimization of usually uncontrollable variables (such as the time be-
4.3.1. Autonomous laboratory tween process steps or the height of the spin-coating dispensing nozzle),
In 2010, the U.S. Air Force Research Laboratory collaborated with the enhancing the accuracy and reliability of experiments.
Air Force Institute of Technology to initiate the development of the Further, the hardware and software of the Ada platform was upgraded
world's first autonomous materials research system (ARES). The goal was to achieve multi-objective optimization of material properties [89]. This
to achieve autonomous experimental design, operation, and data analysis allowed for the rapid and efficient identification of the optimal annealing
through automated experimental apparatus and AI, significantly short- temperature and conductivity for combustion-synthesized palladium thin
ening the materials R&D cycle and fundamentally changing the paradigm films. By upgrading the hardware and software of the original
of materials research. The system integrates automated experiments, AI- self-driving platform, a 7-step automated workflow was implemented. A
driven decision-making, and closed-loop data processing, achieving un- precise 4-axis robot handled the entire process of thin-film deposition
precedented research efficiency. In 2014, the results and research capa- and characterization, while a 6-axis robot was used to transfer samples
bilities of ARES were announced [85]. The system is composed of a between different modules. The preparation of four different variable
six-axis robotic arm, a multi-channel chemical vapor deposition (CVD) thin-film samples and their performance testing were achieved through
reactor, and an in-situ characterization module (Raman spectroscopy, the coordination of dual-robot positions. The processing parameters and
SEM/TEM). By integrating hybrid AI algorithms with random forests and testing results were fed into the qEHVI multi-objective optimization al-
Bayesian optimization models, it conducted approximately 100 experi- gorithm to design the next experiment. The upgraded Ada platform al-
ments daily, with a prediction accuracy of 92 %. The system successfully lows for the control of more variables, eliminating the need for human
identified the temperature and hydrocarbon pressure conditions that intervention during the experimental process. It can run 40–60 experi-
determine the number of wall layers in carbon nanotubes from a ments autonomously, reducing labor costs, minimizing human error, and
ten-dimensional parameter network affecting their growth. ensuring the accuracy and stability of the experiments.
Although machine learning and predictions were performed manu- A typical example of a mobile system is the “Mobile Chemist” for
ally offline at that time, the potential for closed-loop autonomous photocatalysis experiments developed by the University of Liverpool in
experimentation was already demonstrated. Using logistic regression for the UK [90]. This mobile robot can move freely within the laboratory and
dimensionality reduction and Latin hypercube sampling, the system utilizes a combination of laser scanning and touch feedback to achieve
identified critical wall-layer control parameters through 237 experi- precise positioning within a laboratory space measuring 7.3 m  11 m,
ments, about 300 times faster than conventional methods. This break- with spatial positioning accuracy of 0.12 mm and orientation posi-
through stemmed from its innovative "parameter sensitivity tioning accuracy of θ  0.005 . With high-precision control, the robot can
stratification" strategy, which prioritizes primary effect parameters perform dexterous operations at various stations in the laboratory,
before applying response surface methodology for precision optimiza- achieving a level of execution precision comparable to that of re-
tion. In 2016, ARES was used for in-depth studies on the growth rate of searchers. The robotic system can simultaneously consider ten di-
carbon nanotubes [86]. By conducting 84 preliminary experiments to mensions of variables and operates for 21.5 h a day (with the remaining
gain prior knowledge, a model was established by using the random time used for charging). It significantly enhances experimental effi-
forest algorithm, and genetic algorithms were employed to determine the ciency, being 1000 times faster than manual methods and at least 10

12
X. Jiang et al. Review of Materials Research 1 (2025) 100010

times faster than semi-automated robotic workflows. this system inte- path and, through the same modular flow-synthesis platform, continu-
grated with Bayesian optimization independently completed 688 ex- ously feed reactants to achieve the large-scale production of target
periments and identified the optimal catalyst formulation within an perovskite quantum dots.
8-day experimental cycle, increasing the hydrogen production rate by In 2024, the University of Amsterdam in the Netherlands designed a
six times. Additionally, the platform can explore a larger experimental multifunctional RoboChem robotic platform based on a flow system for
space and handle complex multi-component experiments. It is highly the self-optimization and process intensification of photocatalytic re-
flexible, capable of operating in dark environments suitable for photo- actions [94]. Users input experimental parameters through a graphical
sensitive reactions, and meets safety standards, allowing it to work in the user interface (GUI), and the experimental equipment, controlled by
same space as humans. Its modular design makes it easy to expand with Python code, begins to operate. The obtained experimental data is fed
other experimental modules. back to a Bayesian optimization algorithm in real time. This reduces
Collaborated with the University of Liverpool in the UK, East China manual intervention, decreases dependence on specialists, and enhances
University of Science and Technology and Zhejiang Normal University experimental safety. At the same time, the platform efficiently explores
established a materials acceleration platform for discovering molecular vast reaction parameter space, accurately identifies the optimal reaction
nanocluster photocatalysts [91]. This platform combines combinatorial conditions for different substrates, significantly improves the reaction
molecular libraries, high-throughput automated screening and flow yield and space-time yield. Additionally, its modular design facilitates
synthesis technology, accelerating the discovery and preparation of future integration with other equipment, and the accumulated
molecular nanocluster photocatalysts for hydrogen evolution. The high-quality experimental data contributes to the advancement of the
advantage of this platform lies in effectively bridging the gap between digitalization of synthetic chemistry.
small-scale high-throughput screening and large-scale production. The Combining with 3D printing technology is another important devel-
optimized molecular nanocluster MTPA-CA:CNP147 exhibits excellent opment direction of autonomous experimental system. Boston University
hydrogen evolution performance. Theoretical calculations revealed the in the United States developed a Bayesian autonomous experimental
relationship between the moderate binding energy between the molec- researcher (BEAR) for the design of mechanical metamaterials in the
ular donor and acceptor and the high catalytic activity. This research vastness of additive manufacturing design space [95]. This system in-
opens up a new paradigm for integrated research and production, and tegrates five fused deposition modeling (FDM) 3D printers within the
provides an important technical pathway for the large-scale preparation working radius of a six-axis robotic arm and is equipped with a balance to
of efficient catalysts in the future. weigh and a universal testing machine to test toughness. Custom software
Similarly, the University of Science and Technology of China devel- controls the coordinated operation of these devices. This system enables
oped an AI Chemist system and created a data-driven, automated the autonomous optimization of metamaterials and their high-throughput
chemistry platform [92]. This platform integrates literature reading, experiments, resulting in a rapid iteration to find optimal mechanical
autonomous experimental design, chemical experiments, and chemical performance. It quickly quantifies the uncertainty inherent to properties
product optimization, aiming at studying catalysts for the automatic such as toughness and explores the correlation between structure and
synthesis of oxygen-evolving reactions from Martian meteorites. The performance. Therefore, an almost 60-fold reduction was achieved in the
system consists of mobile robots, a computational "brain", cloud servers, number of experiments needed to identify high-performing structures
and 14 dedicated workstations. It achieves full automation from ore relative to a grid-based search. It provides new pathways for research in
analysis to catalyst optimization through a dual-layer workflow. The the mechanical fields and can help discover new properties when com-
system efficiently screens formulations and quickly identifies the best bined with methods like topology optimization.
catalyst from over 3 million possible compositions, significantly
improving the efficiency compared to traditional trial-and-error 4.3.2. Software platform
methods. The screened catalyst can stably and efficiently produce oxy- In 2018, the ARES development team pointed out that autonomous
gen under simulated Martian conditions, providing a reliable solution for experimental systems could lead to discoveries of new materials growing
oxygen supply in Mars exploration. Moreover, the system is highly at a rate akin to “Moore's Law”, enabling them to complete research tasks
automated, with the entire experimental process eliminating no human in multidimensional parameter spaces faster than human researchers and
intervention, reducing human error and improving experimental accu- more effectively address the increasingly complex and high-dimensional
racy and efficiency. It is also highly versatile, and the established demands of materials development [87]. In 2021, the open-source
experimental processes and systems can be applied to other extraterres- operating system ARES OS, capable of fully controlling and driving
trial material research and chemical synthesis, advancing the develop- ARES, was introduced, featuring three main modules: interfaces, data
ment of extraterrestrial resource utilization. analysis, and design planning. ARES OS consists of three core modules
A representative of flow system is "Artificial Chemist" developed by [87]. The interface module enables seamless communication between
North Carolina State University, which adopted a smart-flow-based ARES OS and external devices and software systems, ensuring accurate
quantum dot synthesis strategy [93]. This system consists of precursor data transmission and precise command execution. The data analysis
preparation, flow reaction, and in-situ characterization modules. The module uses advanced algorithms to process experimental data, identify
precursor preparation module includes multiple components that can patterns, trends, and potential issues, and provide insights for experi-
precisely mix reaction reagents; the flow reaction module uses a flow mental optimization. The design planning module formulates experi-
reactor to carry out the reactions; and the in-situ characterization module mental strategies by considering experimental goals, available resources,
uses a specially designed short-path flow cell to perform precise in-situ and data analysis results, such as designing the best reaction conditions
measurements of high-concentration quantum dots without disrupting in carbon nanotube synthesis. Currently, ARES OS has been utilized in
the flow uniformity, solving the problem of significantly diluting the autonomous research systems for carbon nanotube synthesis, flow
reaction phase for previous in-situ photoluminescence quantum yield chemistry, and additive manufacturing.
(PLQY) measurements of quantum dots. The system can quickly explore In 2018, International Business Machines Corporation (IBM) created
the chemical space of colloidal quantum dots, learn synthesis pathways, the RXN online chemistry platform [96]. Through independent analysis
identify the components and synthesis routes required to achieve specific of literature, the platform autonomously learns reaction rules, predicts
optoelectronic properties, and archive and transfer the knowledge both forward and reverse reactions along with their yields for iterative
generated internally to subsequent experiments. It can continuously cycles, formulates the synthesis pathway for the target compound, and
synthesize rapidly optimized quantum dots as needed, producing various automates the synthesis process. A key advantage of RXN lies in its
customized quantum dots using a small amount of starting solution in a Transformer model, which is grounded in the Simplified Molecular Input
short time. Additionally, it can autonomously discover the best synthesis Line Entry System (SMILES) database. This model can be trained in an

13
X. Jiang et al. Review of Materials Research 1 (2025) 100010

unsupervised manner using an attention mechanism to rapidly and for addressing these issues [98,99]. MLIPs utilize ML algorithms to
accurately establish atomic mappings to achieve desired outcomes. automatically learn potential energy functions from data, enabling rapid
Additionally, the platform is initially equipped with cloud AI-driven and and accurate prediction of interactions between particles. Compared
remote-control functionalities. with traditional quantum mechanical methods, MLIPs not only offer
significant advantages in computational efficiency but also adapt to a
4.3.3. Large language models driven autonomous experiments wide variety of complex materials, effectively capturing intricate in-
Recently, the rise of large language models has brought new devel- teractions that traditional potentials struggle to describe.
opment directions and strong help to autonomous experiments, In 2023, The initial machine learning potentials primarily utilized ANNs to
University of California introduced the A-Lab, an autonomous laboratory handle specific systems, directly using all atomic coordinates for poten-
for the solid-state synthesis of inorganic powders [97]. This platform uses tial energy predictions [100]. Although these models found application
computations, historical data from the literature, machine learning and in certain fields, they were limited to specified systems from which the
active learning to plan and interpret the outcomes of experiments per- models were derived. Consequently, their universality and accuracy has
formed using robotics. Over 17 days of continuous operation, the A-Lab restricted their applications in wider fields. To address these challenges,
realized 41 novel compounds from a set of 58 targets including a variety deep learning algorithms have emerged and gradually taken a dominant
of oxides and phosphates that were identified using large-scale ab initio role in MLIP models. Deep Neural Networks (DNNs) algorithms can
phase-stability data from the Materials Project and Google DeepMind. capture the complex interactions between atoms through the nonlinear
Synthesis recipes were proposed by natural-language models trained on mappings of multiple layers of neurons, showing exceptional perfor-
the literature and optimized using an active-learning approach grounded mance, especially when handling large-scale datasets and
in thermodynamics. Analysis of the failed syntheses provides direct and high-dimensional systems. Furthermore, graph neural networks (GNNs),
actionable suggestions to improve current techniques for materials as an emerging deep learning algorithm, utilize graph-structured data to
screening and synthesis design. The high success rate demonstrates the automatically identify interatomic relationships, further improving the
effectiveness of artificial-intelligence-driven platforms for autonomous MLIP accuracy.
materials discovery and motivates further integration of computations, With the development of deep learning algorithms, a variety of deep
historical knowledge and robotics. learning-based MLIP models have been developed. Among them, DP
Carnegie Mellon University developed an AI system named Technology® in China has developed DeePMD-kit [101], a tool using
Coscientist, leveraging advanced large-scale language models such as DNNs to learn the complex patterns of many-body interactions and
GPT-4 [76]. This system is capable of autonomously completing a whole accurately predicting the potential energy of material systems, which
set of processes, from retrieving information, planning and designing surpasses the limitations of many traditional classical potential models.
experiments, writing programs, remotely operating automated systems The “Deep Potential” team, composed of Chinese and American scien-
to do experiments, and then analyzing data. The software modules of tists, combined molecular modeling, machine learning, and
Coscientist can retrieve publicly available information about compounds high-performance computing to extend the limit of molecular dynamics
from the Internet, document data, and other available sources. After simulations to the scale of 100 million atoms [102], reducing a compu-
learning, it formulates synthesis pathway, determines the experimental tation task that would normally take 60 years to just one day. This
protocol, and remotely operates the pipetting robot to conduct the ex- achievement earned them the 2020 Gordon Bell Prize [103]. Combined
periments. Furthermore, this system utilizes publicly available chemical with high-performance computing, the team recently extended the
information in the Simplified Molecular Linear Input Specification ability of first-principles accuracy molecular dynamics simulations to the
(SMILES) format to enable communication between different modules by systems containing billions of atoms [104]. GPUMD [105] is a software
writing code autonomously. Coscientist combines laboratory automation package specifically designed for large-scale atomic simulations, using a
with large language models, breaking through the traditional experi- single-layer neural network to construct precise MLIPs. By combining
mental design mode and opening a new direction for the intelligence of GPU acceleration with the neural network potentials (NEP, this method
chemical experiments. enables efficient prediction of atomic interactions when handling com-
plex systems. Despite the relative simplicity of the model, GPUMD is able
4.4. Intelligent computation to provide accurate simulation results for various materials and their
applications. As an emerging field, MLIPs have demonstrated great po-
Materials intelligent computation and design refer to the application tential in the field of material simulations. With the continuous
of AI in multi-scale computing, high-throughput computing, and inte- advancement of ML algorithms, datasets, and computational resources,
grated computational materials engineering (ICME). By utilizing ma- MLIPs will play an increasingly important role in the future materials
chine intelligence, it achieves autonomous setting of trial-and-error R&D.
spaces, selection of computational approaches, adjustment of computa-
tional resources, and optimization of screening criteria during the com- 4.4.2. Efficient computation and screening
putations for new material design, preparation, processing, and Traditional computational methods face significant challenges when
deployment. This forms a closed-loop of autonomous virtual iterations to dealing with complex data and high-dimensional problems, particularly
obtain optimal designs that guide experimental research and engineering in terms of computational efficiency and accuracy. The integration of AI
applications of new materials. Intelligent computational technology with computational methods, especially in areas such as high throughput
breaks through the fixed thinking patterns of the human brain, autono- computation (HTC), molecular dynamics simulations, and DFT, is driving
mously explores new variables beyond known variable spaces, and innovation in material computation and screening.
autonomously optimizes design ideas and technical pathways, signifi- The development of AI, especially ML models, has provided new
cantly improving material computational efficiency and enhances the momentum for DFT, significantly improving the computational effi-
reliability and accuracy of material design and screening. ciency of DFT, enabling ML models to predict the properties of new
materials several orders of magnitude faster than traditional DFT
4.4.1. Machine learning interatomic potentials methods [34,35,106]. This accelerated capability for material discovery
Traditional molecular simulations rely on quantum mechanical cal- and screening is attributed to the ability of ML models to learn from
culations and potential energy functions obtained by fitting experimental DFT-generated data, which in turn supports the rapid iteration of
data. The computational cost and efficiency become a bottleneck as the experimental designs and the development of automated experimental
scale and complexity of material systems continue to increase. The rise of processes. In terms of accuracy, ML models, through training on large
machine learning interatomic potentials (MLIPs) has provided new ideas datasets, reduce prediction errors, thereby enhancing the precision of

14
X. Jiang et al. Review of Materials Research 1 (2025) 100010

DFT predictions. As the training dataset grows, the accuracy of the model platform of ALKEMIE [112], acronyms for Artificial Learning and
approaches the noise level, particularly in compound screening and Knowledge Enhanced Materials Informatics Engineering, which facili-
ranking, where its importance becomes evident. By optimizing feature tates the widespread of data-driven techniques across diverse scientific
sets and incorporating physical principles, ML models effectively miti- communities based on the design principles of AMDIV (automation,
gate the impact of data imperfections, thereby improving the reliability modular, database, intelligence, visualization). All structured and
of DFT predictions in material applications. AI also enhances the trans- multi-modal data from computational processes and results are stored in
ferability of DFT by learning electronic structure information. For the ALKEMIE-Matter Cloud, where they can be accessed, queried, and
example, researchers at Tsinghua University combined deep learning shared via high-level APIs. The Matter Cloud database currently contains
with DFT to train a neural network that efficiently represents the DFT over 600,000 crystal structure entries, more than 10,000 phonon spec-
Hamiltonian, bypassing the computationally intensive self-consistent trum records, over 100,000 DFT band structure datasets, more than 1,
field (SCF) iteration process [33]. This approach demonstrated high ac- 000 GW-accuracy band structure datasets, and over 200,000 sets of
curacy, good transferability, and high efficiency across different material machines learning potential data. The MatCloud platform [113], devel-
systems, particularly in the study of twisted van der Waals materials. In oped by researchers from Chinese Academy of Sciences, employs quan-
addition, AI techniques improve the computational capability of DFT tum mechanics/machine learning (QM/ML) methods to predict material
when applied to large-scale systems by breaking down tasks and utilizing properties and drive the discovery of new materials, providing a fully
local contributions. ML models can learn from localized atomic envi- automated solution for the entire process from modeling to data storage.
ronments and extend this knowledge to larger systems, enabling in-depth Under extreme conditions, the Professional Materials at Extreme
studies of complex material systems. (ProME) platform has been setup, which is consisted of several toolkit
As the complexity of molecular systems increases, traditional MD with IPR (Independent Intellectual Property), as shown in Fig. 8. It can be
simulations face significant challenges, including a huge demand for seen that Similar Atomic Environment (SAE) tool and AI-based Crystal
resources and difficulty in data processing. In recent years, AI technol- (ABC) search tool are designed to construct the order/disorder crystal
ogies have offered new approaches to address these challenges [107, structures, whose atomic and electronic structures together with some
108]. First, AI enables the efficient processing and analysis of large physical properties can be precisely predicted via the classical first-
datasets generated by MD simulations, which often contain underlying principles calculations [112]. Through utilizing the further analysis
patterns and relationships that traditional methods are unable to effec- tools such as the second version Mean-Field Potential (MFP2) tool and
tively extract. Through automated data processing, AI trains models to Auto-Calphad, thermodynamic properties especially for those extreme
recognize these latent relationships, thereby significantly enhancing conditions can be obtained. The High-performance Tool of Elastic
research efficiency. For instance, researchers at Soochow University Modeling (HTEM) and Efficient Calculation of alloy by Coherent Poten-
employed a fully connected neural network model to analyze the rela- tial Approximation (ECPA) tools display their capability in predicting the
tionship between the structure and conductivity of 180,000 alkaline elastic and plastic properties while HiPPOS (High strain rate Plastic
anion exchange membrane variants, quickly identifying promising Phase field mOdel) together with integrated models and algorithms
membrane materials [109]. These predictions provided more precise simulate the microstructure evolutions and optimizations. The Corrected
initial conditions for MD simulations, effectively avoiding the cumber- Assessments & Self-adaptive Tool (CAST) is designed as the fundamental
some assumptions and trial-and-error processes inherent in traditional one organizing the workflows and smartly corrected the errors during
methods. Second, AI plays a critical role in optimizing parameter selec- high-throughput first-principles calculations. It is worth mentioning that
tion in MD simulations. In traditional MD simulations, the accuracy and all these individual tools together with the integrated ProME platform
computational efficiency typically depend on the precise specification of have been tested by third party and validated and verified in various
parameters such as time step, temperature, and force field. However, the materials.
selection of these parameters often requires extensive trial and error and
empirical knowledge. By training ML models on historical data, AI learns 4.4.3. ICME in AI þ era
from past simulations and predict optimal parameter settings based on ICME integrates computational modeling, experimental validation,
simulation goals and the characteristics of the studied system. The inte- and engineering design, employing multi-scale modeling techniques to
gration of AI with MD simulations enhances both the efficiency and ac- unify physical processes from the atomic level to the macroscopic level.
curacy of traditional MD methods, enabling researchers to more This approach enables researchers to optimize processing techniques and
effectively handle complex material systems. predict product performance prior to materials fabrication and product
HTC involves large datasets and complex computational processes, manufacturing. ICME covers key components of product development,
particularly when dealing with high-dimensional data, which often leads including materials property databases, microstructural models, micro-
to the ‘dimension disaster’. ML offers powerful support in this context. By structure–property models, and cost analysis models [115]. However, as
adopting a data-driven approach, ML can extract valuable patterns and materials design becomes more complex, traditional ICME approaches
insights from vast amounts of data, thereby further optimizing the HTC face significant challenges when dealing with high-dimensional, multi-
workflows and enhancing the efficiency and accuracy of material variable, and nonlinear behaviors. The integration of AI offers new so-
screening and design [110]. For example, ML analyzes historical data to lutions for ICME [114]. By learning from large-scale experimental data
identify key variables, optimize the selection workflows. Moreover, AI is and simulation results, AI can automatically identify underlying patterns,
capable of recognizing previously overlooked potential factors, thereby optimize model parameters, and thus improve predictive accuracy and
refining the scientific basis of materials design. For material screening, computational efficiency. AI not only compensates for the limitations of
ML models are trained on existing datasets to evaluate and rank candi- traditional physical models but also handles the interaction of multiple
date materials, enabling the identification of promising candidates variables and complex relationships, positioning itself as a pivotal tech-
within a significantly smaller pool. This new technique may provide nology driving the further advancement of ICME.
precise assessments during the early stages of material design, elimi- In recent years, the integration of AI and ICME has brought revolu-
nating the need for labor-intensive manual selection. Researchers from tionary changes to materials design. AI helps ICME with revealing the
DeepMind [111] utilized a GNoME model, which was rained on a large relationship between microstructure and macroscopic property,
and diverse set of first-principles calculations, to predict 2.2 million analyzing and learn from large volumes of experimental data and simu-
crystal structures, of which 380,000 have been confirmed as stable. This lation results, and enabling more accurate property predictions [116]. AI
model increased the number of stable crystals by more than an order of has demonstrated unique advantages, particularly in situations where
magnitude, opening a new approach to material discovery. explicit physical models are lacking. In multi-scale modeling, AI-aided
One significant achievement of Materials Genome Engineering is the ICME integrates cross-scale data and automatically adjust model

15
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Fig. 8. Tool kits together with their functions embedded in the ProME platform with IPR (Independent Intellectual Property) [114].

parameters, enhancing the accuracy and adaptability of the models evolution of smart design and manufacturing paradigm of advanced
[117]. In materials design and optimization, AI helps quickly identify materials from knowledge-enabled data-driven ICME [115,120] to AIþ
promising material combinations from vast amounts of experimental and era, promoting the transformations from designing the materials to
simulation data, and automatically optimize microstructure, composi- designing with materials and driving the dramatically evolution of AIþ
tion, and processing conditions [118]. This capability is especially paradigm from the Embedding AI to the Copilot AI and the on-going AI
valuable in the development of materials for high-performance or spe- Agent [114]. While the MGI highlights the dominate roles of experi-
cific applications. mental tools, computational tools and databases together with their in-
The integration of AI and ICME has driven a leap forward in materials teractions, the HCPs & MGE emphasis on the interactions of
science. For example, researchers at Central South University [119] used Human-Cyber-Physics [122], the common cross field of which indicate
a multiple loops machine learning framework to establish the composi- the future applications of those advanced AI technologies. The machine
tion–processing–structure–property relationship of cemented carbides, learning algorithms could accelerate the generation of fundamental
enabling the intelligent design. Researchers at Northwestern Poly- materials insights and basic science research through identification of
technical University [120] adopted a data-driven ICME framework to main/value data relationships to strengthen human-physics and
achieve intelligent design and optimization of advanced structural metal human-cyber interpretations and yield scientific knowledge/model
materials, such as light-weight materials (e.g., Ti, Mg, and Al alloys), [123]. Through digitization of knowledge, the Cyber-Physical systems
refractory high-entropy alloys, and superalloys. These successful appli- (CPs) can be integrated by developing tools to digitize the knowledge and
cations highlight the significance of knowledge-based mod- to crat new materials together with their robust manufacturing process
eling/simulations, machine learning algorithms, and the ICME [124]. Industry 4.0 fueled by data and machine learning is building on
ecosystems. Researchers at Texas A&M University [121] have introduced the digital advances by bridging the physical and digital world through
a method that integrates phase-field modeling with deep Variational cyber-physical systems, which has transformed into Materials 4.0 [124].
AuToEncoder for inverse microstructure-centered materials design. This Comparing the top-down engineering routine with the bottom-up
method has been successfully applied to predict the effective thermal computational design one, the twin features between experimental and
conductivity of multi-phase materials. theoretical chains are highlighted by two types of arrows with different
Moreover, it is noted that AI aided design and manufacturing have background colors, which also reveal the corresponding
been considered as one kind of robust data-driven and data-intensive composition-processing-structure-properties-performance (CPSPP) rela-
technologies in the ICME era [115,120]. As presented in Fig. 9, the tionship during the smart design and manufacturing. On one hand,

Fig. 9. The evolution of smart design and manufacturing paradigm of advanced materials from knowledge-enabled data-driven ICME [115,120] to AIþ era, promoting
the transformations from designing the materials to designing with materials and driving the dramatically evolution of AIþ paradigm from the Embedding AI to the
Copilot AI and the on-going AI Agent [114]. The insert description of Materials Innovation Platform was reported in Ref. [132].

16
X. Jiang et al. Review of Materials Research 1 (2025) 100010

digital twins have been considered important tools for realizing the The application prospects of 3D printing in digital manufacturing
real-time interaction and integration between information and the have garnered significant attention. With continuous advancements in
physical world, and represent a key enabling technology for achieving materials science, printing technology, and artificial intelligence (AI), 3D
cyber-physical integration for smart assembly/manufacturing [125], i.e., printing can now process not only traditional metals [137] and polymers
for aircraft, trains and engines [114,120,126]. On the other hand, ac- [138] but also ceramics [139], composites [140], and even biological
cording to the CPSPP relationship, process-structure and materials [141], greatly expanding its scope in intelligent manufacturing.
structure-property models highlight the significant role of materials AI plays a transformative role in optimizing the entire 3D printing pro-
knowledge in the transformation from data and informatics to cess, from material composition [142] to structural design [136] and
manufacturing, which also indicate the high-value chain toward intelli- process optimization [143]. In material science, AI algorithms analyze
gent manufacturing [125,126]. Furthermore, motivated by the LLM and vast datasets to predict and design novel materials with tailored prop-
the generative and the next-generation AI [14,127–129], the role of AI in erties, enhancing the performance of printed parts by identifying optimal
the assignments of accelerating materials discovery and manufacturing material combinations and predicting their behavior under various
would be increased from ~30 % of Embedding AI paradigm to ~50 % of conditions. In structural design, AI facilitates generative design, where
Copilot AI paradigm, and even ~80 % of on-going AI Agent paradigm algorithms propose and optimize complex geometries based on factors
[114]. Therefore, through accelerating materials discovery and like strength, weight, and material usage, enabling the creation of highly
manufacturing enabled by LLM or AI-agents [130], the Human-Cyber efficient, performance-optimized structures [55]. Furthermore, AI drives
interactions in line with the AI-enabled HCPs and MGE will be dramat- process optimization by integrating intelligent control systems with 3D
ically changed. The materials S&T and the scientific discoveries and printing, enabling full automation from design to production [144]. AI
applications in AIþ era will transform from “design the materials” to continuously monitors real-time data during printing, making adjust-
“design with materials” [114,130,131]. ments to printing parameters based on feedback, which allows for
adaptive manufacturing that enhances precision, reduces defects, and
4.5. Intelligent manufacture minimizes errors [143]. This technological breakthrough, powered by AI,
brings higher flexibility, sustainability, and precision to manufacturing,
4.5.1. AI-assisted 3D printing positioning 3D printing as a key driver in the transformation toward
The traditional subtractive manufacturing process often results in smart manufacturing.
materials waste and has limitations in terms of precision and cost control
when processing complex shapes. 3D printing builds three-dimensional 4.5.2. AI-integrated digital twin
objects layer by layer, directly fabricating complex geometries from One of the main challenges in smart manufacturing is seamlessly
digital design files. This technology surpasses the size and shape con- integrating the physical and virtual spaces. Advances in data acquisition,
straints of traditional manufacturing, addressing the challenges of communication, and computing technologies are making this integration
machining and manual methods in producing intricate structures. increasingly feasible. Among these advancements, Digital Twin (DT) has
Additionally, 3D printing reduces production steps, enables mold-free emerged as a key enabler of cyber-physical integration, attracting sig-
manufacturing, and effectively saves time, reduces costs, and shortens nificant interest from both academia and industry. DT integrates real-
production cycles. world and virtual data across the product lifecycle, developing
In the aerospace industry, 3D printing has become a vital technology advanced analytics to enhance decision-making and operational effi-
for producing complex components, offering significant advantages over ciency. Its applications span smart manufacturing, real-time system
traditional manufacturing methods. It enhances design freedom, monitoring, and predictive maintenance. By continuously acquiring real-
enabling the creation of intricate geometries and internal structures that time data and updating virtual models, DT enables predictive adjust-
would be difficult or costly to produce using conventional techniques. ments, improving system reliability and efficiency. As a result, DT not
Additionally, 3D printing reduces material waste, cuts manufacturing only accelerates the adoption of smart manufacturing but also facilitates
costs, and optimizes component designs. Selective Laser Melting (SLM), a autonomous decision-making, real-time process optimization, and
widely used metal 3D printing technique, works by using a high-power enhanced lifecycle management.
laser to fully melt and fuse metallic powder particles layer by layer, Digital Twin (DT) technology has significantly advanced materials
enabling the production of fully dense, high-precision metal parts with science by enhancing the prediction of mechanical properties, optimizing
complex geometries [133]. This method not only supports the fabrication additive manufacturing processes, and improving high-temperature
of topology-optimized designs but also minimizes material waste, mak- material performance assessments. Researchers at the University of Sci-
ing it ideal for lightweight, high-performance aerospace components ence and Technology Beijing developed a DT-driven Hall-Petch model
such as turbine blades or structural brackets. For example, in China's that integrates intrinsic material factors, such as valence electron dis-
C919 aircraft, SLM technology developed by researchers at Northwestern tance and grain boundary energy, to eliminate empirical fitting constants
Polytechnical University was used to produce 23 complex components, and enable real-time optimization of polycrystalline metal properties
including the central winglet bar. Similarly, NASA employed SLA to [67]. In additive manufacturing, DT has been employed to refine process
manufacture a spacecraft fuel nozzle with a complex internal structure, accuracy in laser powder bed fusion (L-PBF) of Inconel 718 alloy by
reducing material waste and improving engine performance and reli- integrating real-time monitoring and experimental data, thereby
ability. Beyond component production, 3D printing offers greater flexi- improving simulation reliability and optimizing the homogenization
bility in material processing, enabling the creation of metamaterials with process. For instance, DT-assisted homogenization facilitated the disso-
extraordinary physical properties [134]. At NASA's Glenn Research lution of 67 % of the Laves phase within 0.5 h, ensuring more uniform
Center, researchers used a model-based alloy design approach and microstructural evolution [145]. Furthermore, researchers at North-
laser-based rapid manufacturing technology to develop a new western Polytechnical University combined DT with Convolutional
oxidation-dispersion-strengthened nickel-cobalt-based alloy, GRX-810. Neural Networks (CNN) to enhance the predictive modeling of creep
Compared to traditional alloys, GRX-810 demonstrated twice the fracture life in high-temperature titanium alloys, significantly improving
strength, 1,000 times the creep resistance, and double the oxidation model accuracy and adaptability through real-time feedback mechanisms
resistance at 1093  C [135]. Similarly, researchers at the China Academy [146]. These advancements underscore the transformative role of DT in
of Engineering Physics used direct-write 3D printing technology to create materials research, enabling precise computational modeling, dynamic
mechanical metamaterials with ultra-long mechanical adaptability and process optimization, and data-driven decision-making for the develop-
negative stiffness [136]. ment of advanced materials.

17
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Defect prediction and control are key to improving product quality in AI-Scientist independently optimized the performance of lithium-rich
smart manufacturing. Researchers at Beihang University proposed a new disordered cathode materials and, after four cycles, the materials pro-
DT-driven approach for End-face Defect Control (EF-DC) in hot-rolled duced in an automated laboratory based on the design reached a capacity
coils [147]. Traditional EF-DC relies on physical-space data for defect of 282 mAh/g with 100 % capacity retention after 20 cycles, breaking
identification, limiting its effectiveness. By integrating DT with deep through the Pareto frontier of the LiMnTiOF system and achieving
learning, they created a multi-dimensional EF-DC model that maps sur- world-leading results.
face quality across geometry, physics, behavior, and rules. This model A preliminary case study of LLM enabled intelligent microstructure
combines operational and virtual data to track and correct defect-causing optimization and defects classification of welded titanium alloys is pre-
abnormalities in real time, enhancing defect detection and prevention. sented in Fig. 10, the technical roadmap of which consists of the LLM,
Additionally, they proposed a DT-based intelligent model for CNC ma- auto-coding via AI agent, image processing, image mosaic, and ML for
chine tools that uses OPC interfaces to ensure interoperability between welding defect detection [114,153]. Based on the welding images ob-
physical and digital spaces [148]. This model predicts potential failures, tained from high-throughput experiments, it is highlighted that the
such as ball screw issues, by analyzing real-time data from vibration and AI-integrated software is able to compile numerous images and extract
temperature sensors, improving equipment reliability and efficiency. crucial features from them, thereby efficiently facilitating the welding
With the continuous advancement of smart manufacturing technol- defect detection [114,153]. These achievements not only highlight the
ogies, DT has emerged as an important technology driving the trans- exceptional potential of agents in materials research but also lay the
formation and upgrading of industrial production models. By enabling foundation for further developments in the field of smart manufacturing,
real-time data feedback and dynamic optimization of virtual models, T signaling a new transformation in manufacturing through the wide-
improves the precision and efficiency of manufacturing processes while spread application of agent technology, with improvements in efficiency,
providing new theoretical foundations and technological support for process optimization, and the overcoming of innovation bottlenecks.
smart manufacturing.
5. Prospect and challenge
4.5.3. AI agent-assisted material manufacturing
With the rapid development of LLM, emerging technologies repre- The rapid advancement of AI is revolutionizing materials science,
sented by intelligent agents are set to significantly accelerate the shift enabling unprecedented capabilities in data management, computational
from automation to smart manufacturing [149]. An agent is a compu- modeling, autonomous experimentation, and interdisciplinary collabo-
tational system capable of autonomously perceiving its environment, ration. The establishment of a robust materials data circulation
making decisions, and executing tasks. Unlike traditional programs or infrastructure is essential for harnessing the full potential of AI,
algorithms, an agent exhibits autonomy, adaptability, collaboration, and ensuring seamless data integration, standardization, and accessibility
learning abilities, allowing it to interact with its surroundings and adjust across global research communities. Advancements in AI power, al-
its behavior based on feedback, forming a closed-loop iterative system gorithms, and tools continue to drive the acceleration of material dis-
[150]. Agents typically gather environmental data via sensors, make covery and design, with deep learning, reinforcement learning, and
decisions using reasoning mechanisms, and engage with the environment generative models enhancing predictive accuracy and optimization effi-
through actuators such as robots or control systems. The major charac- ciency. Federated learning presents a promising approach to training
teristics of agents make them particularly useful in dynamic and complex materials science large language models while preserving data privacy,
environments, where they can effectively handle uncertainty and opti- fostering secure, large-scale collaboration. The development of auto-
mize decision-making processes. mated/intelligent experimental techniques is streamlining high-
Agents improve smart manufacturing at both local and global levels. At throughput material synthesis and characterization, forming smart,
the local level, agents enhance quality control, process optimization, and closed-loop optimization processes that enhance R&D efficiency. Mean-
equipment management [151,152]. Traditional quality checks rely on while, multiscale modeling and digital twin technology are breaking
manual inspection or fixed machines, which can be slow and inaccurate. down spatial and temporal barriers, enabling real-time simulation,
Agents use AI vision systems and sensors to detect defects automatically, optimization, and precise control of materials manufacturing. Finally,
improving accuracy. They also monitor equipment in real time, predicting fostering an interdisciplinary workforce with expertise in both AI and
failures to prevent breakdowns. Additionally, agents enable flexible pro- materials science will be critical for sustaining innovation and translating
duction, adjusting machine settings to quickly switch between different AI-driven advancements into practical industrial applications. Moving
products for customization. By improving quality, efficiency, and flexi- forward, the integration of AI with materials science will not only
bility, agents make smart manufacturing faster, more reliable, and accelerate the discovery of next-generation materials but also drive sus-
cost-effective. At the global level, agents can function as the “smart brain” tainable, intelligent, and highly efficient manufacturing processes,
of a factory, effectively integrating isolated departments, optimizing shaping the future of materials innovation.
operational efficiency through real-time data exchange and intelligent
scheduling, and reducing production costs. Through such global coordi- 5.1. Materials data circulation infrastructure
nation, agents can achieve optimal allocation of production resources,
enhancing the overall intelligence of the manufacturing process. For the future development of AI in materials science, it is essential to
Although the empowerment of smart manufacturing through agents gather multi-scale, high-quality, and extensive materials data, including
is still in its early stages, agents have already proven effective in LLMs and the evolution of machine learning algorithms and software.
advancing material chemistry research at the laboratory level [149]. The implementation of FAIR principles—Findability, Accessibility,
Researchers from the Shanghai Institute of Ceramics, Chinese Academy Interoperability, and Reuse—across various materials communities, both
of Sciences, inspired by brain functions, developed a multi-agent in academia and industry, is becoming increasingly important. The
collaborative system with self-growth capabilities—Material Intelli- construction of National materials big data center is poised to revolu-
gence Scientist System (AI-Scientist). This system includes fundamental tionize how industries manage and utilize materials data in a globalized
functional modules for completing complex material tasks, such as economy. Prospective benefits include improved data accessibility for
autonomous learning, data extraction, material reasoning, and interac- researchers and manufacturers, which can lead to accelerated product
tion with the physical world. In scenarios with limited data, AI-Scientist development and more sustainable practices through better materials
demonstrated superior reasoning capabilities compared to Bayesian selection and usage. Future research should focus on developing an in-
optimization through a dual-driven mode of data and knowledge, tegrated big data cloud platform for automatic data collection, storage,
significantly improving the efficiency of material R&D. Notably, mining, and application, building a diversified, multi-layered, and

18
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Fig. 10. Large language models enabled intelligent microstructure optimization and defects classification of welded titanium alloys [153].

systematic infrastructure for materials data. Establishing circulation power, advanced algorithms, and the development of accessible toolkits.
technologies for material data identification, citation, evaluation, and These advancements enable researchers to address previously insur-
trading will help overcome technical bottlenecks in data sharing, mountable challenges, significantly accelerating material discovery and
knowledge sharing, and information security. This will promote revolu- design.
tionary changes in R&D models, accelerating the development and State-of-the-art computational infrastructure underpins AI's integra-
application of new materials. tion into materials research. Specialized hardware—such as GPUs, high-
Additionally, standardized protocols for data sharing can facilitate performance computing (HPC) systems, and emerging quantum
compliance with international regulations and enhance the interopera- computing platforms—has dramatically increased the capacity to process
bility of systems across borders, ensuring that data flows seamlessly be- large datasets and perform complex calculations. These computational
tween different jurisdictions and stakeholders. This includes advances are essential for supporting data-intensive AI models and
incorporating common data dictionaries, data exchange formats, and enabling rapid simulation and analysis of materials.
pedigree standards. The absence of global or internationally recognized Advancements in AI algorithms and models have expanded the ca-
materials data standards for collection, storage, and exchange signifi- pabilities of materials science beyond traditional methods. Machine
cantly hinders data sharing and community development in materials learning approaches, initially dominated by regression and decision-tree
science. There is a pressing need for standardized materials data sche- models, have evolved to include deep learning, reinforcement learning,
mas, consistent file storage formats, international APIs for databases and generative models. These advanced models can predict material
across different countries, and robust data exchange protocols between properties, optimize synthesis processes, and even design new materials
various software platforms and machines. As new methods and tech- with targeted characteristics. Notably, generative adversarial networks
nologies are rapidly adopted, it becomes impossible to stay current with (GANs) and variational autoencoders (VAEs) are increasingly used to
all the latest developments in the field. To address the diversity of create materials with specific properties, such as enhanced strength or
datasets and machine learning methods, a checklist has been proposed to superconductivity.
guide the assessment with substantial ML content [154]. However, it The development of comprehensive AI toolkits and platforms has
remains challenging to determine whether AI is ready for materials data made these technologies more accessible to the materials science com-
standardization. Achieving informative, exhaustive, and adaptable data munity. Platforms like the Materials Project and AFLOW, along with
standardization in materials science continues to be an outstanding widely used machine learning libraries such as TensorFlow and PyTorch,
challenge. enable seamless integration of AI into research workflows. These re-
sources lower the barrier to entry, allowing researchers to leverage AI
without extensive backgrounds in data science. Cloud-based services
5.2. Advances in AI power, algorithms, models, and tools further facilitate the deployment of scalable models, offering on-demand
computational resources that accelerate the design and optimization of
The continuous evolution of AI technologies is driving transformative new materials.
changes in materials science, propelled by enhanced computational

19
X. Jiang et al. Review of Materials Research 1 (2025) 100010

Looking ahead, four emerging AI-driven design patterns—reflection, platform can significantly improve the efficiency of intelligent robot
tool use, planning, and multi-agent collaboration—are expected to drive operations and reduce ineffective experiments, further enhancing overall
significant progress. These approaches will enhance autonomous dis- experimental efficiency. With the application of the metaverse and dig-
covery, streamline complex workflows, and promote collaborative ital twins, experimental data will be presented in virtual reality (VR) and
innovation, further accelerating advancements in materials science. augmented reality (AR), allowing scientists to immerse themselves in and
direct the experimental process. Moreover, the establishment of an
5.3. Federal training of materials scientific large language models intelligent cloud platform will become the core scheduling and control
center of the entire intelligent experimental system, where all experi-
Federal training for materials scientific large language models marks mental data, control instructions, optimization algorithms, and intelli-
a transformative shift in materials science. This decentralized approach gent decisions are centrally managed and intelligently scheduled. This
enables privacy-preserving collaboration across disparate datasets, networked and cloud-driven experimental model makes the storage,
significantly accelerating material discovery, optimizing preparation processing, and sharing of experimental data more efficient while also
processes, and enhancing material performance. Future trends in this providing strong support for remote monitoring, collaborative optimi-
field will likely emphasize integrating high-throughput simulations, AI- zation, and resource scheduling. This significantly enhances the system's
driven optimization, and multiscale modeling, which together will scalability and flexibility, providing a convenient technological founda-
allow for the design of materials with highly tailored properties. The tion for collaborative experiments across multiple fields and interdisci-
integration of data from experimental studies, real-time sensors, and plinary cooperation. Moreover, when adopting blockchain technology,
large-scale simulations will not only expedite the discovery of novel the data security and traceability will be greatly enhanced, ensuring the
materials but also enable more efficient preparation processes. Further- accuracy, transparency, and protection against tampering of data
more, federated models will facilitate the customization of materials for throughout the experimental process. Next-generation AI technologies
specific industrial needs, such as in aerospace, automotive, energy, and will continuously drive the system's self-learning and optimization,
healthcare, fostering innovation and enhancing product functionality enabling smarter, more automated experimental operations.
across sectors. The ability to optimize resource use and minimize waste
aligns with the growing focus on sustainability, making this technology 5.5. Multiscale modeling and digital twin technology
even more valuable in addressing global challenges.
The key advantage of federal training lies in its capacity to foster Future computational design of materials will focus on breaking
collaboration while protecting data privacy, allowing multiple stake- spatial and temporal scale boundaries, developing cross-scale computa-
holders to contribute to model development without disclosing sensitive tional and autonomous methods and software for multiscale modeling.
information. This decentralized framework supports the scalability of Machine learning molecular dynamics simulations have successfully
models and fosters global cooperation, facilitating the development of achieved quantum mechanical precision simulations of millions of atoms,
more robust and adaptable materials models. However, challenges with the potential to truly transcend the simulation of spacetime
remain, including the complexities of integrating diverse data sources, boundaries, enabling the analysis of structure-property relationships at
the computational expense of training large models, and ensuring model the micro-nano scale and the atomic-scale design of micro-nano device
accuracy and adaptability to evolving materials data. Despite these structures. Establishing on-demand, elastically scalable autonomous
hurdles, the continued evolution of federated learning in materials sci- computational methods for materials will enable distributed, automated,
ence holds immense potential to revolutionize industries, drive break- and efficient computations. The integration of data-driven and compu-
throughs in material design, and promote sustainability across global tational materials engineering will break through bottlenecks in multi-
manufacturing systems. scale modeling and comprehensive computational design that encompass
material composition, microstructure, macrostructure, properties, pro-
5.4. Intelligent experimental techniques cessing, and service behavior, achieving “pre-optimization” before ma-
terial production or product manufacturing [114,130].
Future material experiments will rely on robotics technology, AI, and A substantial accumulation of foundational material data will be
the integration of big data with cloud platform technologies to develop undertaken, covering multidimensional processing parameters, multi-
intelligent robotic systems for material experiments. These systems field environments, multi-process data, and service behavior data. By
feature high sensitivity, multi-dimensional motion perception, multi- integrating computational materials engineering and machine learning
objective recognition, and intelligent decision-making feedback mecha- modeling, a digital model library and knowledge base will be constructed
nisms. Deeply integrating intelligent robots with high-throughput ma- to dynamically describe the complex intrinsic relationships, interactions,
terial preparation, testing, and characterization technologies, they and evolutionary behaviors of material composition, structure, process-
construct a physics platform with features of high-throughput, automa- ing, and performance. Breakthroughs in database technologies for the
tion, and intelligence. Through multi-sensor real-time interaction tech- online collection, processing, storage, and computational interaction of
nology and multi-scale modeling and simulation, the system is able to multi-source heterogeneous data will enable real-time simulation visu-
monitor changes in variables during experiments, providing timely alization of material preparation and processing. Advances in physical
feedback and optimizing the experimental plan. Through deep learning and information system mapping and real-time interaction technologies
of experimental data and multi-objective optimization, the intelligent will drive the construction of digital twin systems, enabling rapid online
system can automatically adjust each experimental step and decision, optimization and precise control of material manufacturing. These ad-
forming an intelligent closed-loop optimization experimental plan aimed vancements will support the digitalization and intelligence of materials
at achieving target performance in material research and development. science and engineering.
With a continuous upgrade of intelligent technologies, autonomous
experimental systems may adopt a modular design. By developing 5.6. Building an interdisciplinary workforce
modules tailored to different experimental scenarios, the system will
match modules according to specific experimental needs, thus enabling As AI becomes increasingly integrated into materials science, there is
full-process automation of various complex experimental tasks. This a growing need for an interdisciplinary workforce that can bridge the gap
design supports complex experimental systems with multiple stages and between traditional materials science and cutting-edge AI technologies.
parameters, making it adaptable to a wide range of application scenarios, This requires a combination of expertise in both domains, with pro-
including solid, liquid, and gas environments. With the enhancement of fessionals who can apply AI methods to solve complex materials prob-
intelligence by using visual assistance technologies, the experimental lems. Materials scientists must gain proficiency in AI techniques, such as

20
X. Jiang et al. Review of Materials Research 1 (2025) 100010

machine learning, data analytics, and simulation tools, while AI practi- [11] H.S. Rao, A. Mukherjee, Artificial neural networks for predicting the
macromechanical behaviour of ceramic-matrix composites, Comput. Mater. Sci. 5
tioners need a strong understanding of the intricacies of materials sci-
(1996) 307–322.
ence, including material properties, fabrication methods, and [12] X.-L. Yu, B. Yi, X.-Y. Wang, Prediction of the glass transition temperatures for
performance testing. Developing this cross-disciplinary skillset is essen- polymers with artificial neural network, J. Theor. Comput. Chem. 7 (2008)
tial to ensuring effective collaboration and innovation. 953–963.
[13] B. Meredig, A. Agrawal, S. Kirklin, J.E. Saal, J.W. Doak, A. Thompson, K. Zhang,
To address this need, educational programs must evolve, offering A. Choudhary, C. Wolverton, Combinatorial screening for new materials in
training that combines materials science with AI, data science, and unconstrained composition space with machine learning, Phys. Rev. B 89 (2014)
computational methods. Collaborative initiatives between universities, 94104.
[14] P. Raccuglia, K.C. Elbert, P.D.F. Adler, C. Falk, M.B. Wenny, A. Mollo, M. Zeller,
research institutions, and industry partners can also help cultivate a S.A. Friedler, J. Schrier, A.J. Norquist, Machine-learning-assisted materials
workforce that is capable of addressing the challenges posed by AI discovery using failed experiments, Nature 533 (2016) 73–76.
integration in materials research. Efforts to promote such interdisci- [15] J. Xiong, S.-Q. Shi, T.-Y. Zhang, A machine-learning approach to predicting and
understanding the properties of amorphous metallic alloys, Mater. Des. 187
plinary training will ensure a future workforce capable of leveraging AI (2020) 108378.
to its fullest potential, accelerating innovation and the development of [16] C. Wang, C. Shen, Q. Cui, C. Zhang, W. Xu, Tensile property prediction by feature
new materials solutions. Building this skilled, collaborative workforce engineering guided machine learning in reduced activation ferritic/martensitic
steels, J. Nucl. Mater. 529 (2020) 151823.
will be a critical factor in driving the next generation of materials science. [17] S. Lu, Q. Zhou, L. Ma, Y. Guo, J. Wang, Rapid discovery of ferroelectric
photovoltaic perovskites and material descriptors via machine learning, Small
CRediT authorship contribution statement Methods 3 (2019) 1900360.
[18] L. Ward, A. Dunn, A. Faghaninia, N.E.R. Zimmermann, S. Bajaj, Q. Wang,
J. Montoya, J. Chen, K. Bystrom, M. Dylla, K. Chard, M. Asta, K.A. Persson,
Xue Jiang: Writing – original draft. Dezhen Xue: Writing – review & G.J. Snyder, I. Foster, A. Jain, Matminer: an open source toolkit for materials data
editing. Yang Bai: Writing – review & editing. William Yi Wang: mining, Comput. Mater. Sci. 152 (2018) 60–69.
[19] J. Ma, B. Cao, S. Dong, Y. Tian, M. Wang, J. Xiong, S. Sun, MLMD: a programming-
Writing – review & editing. Jianjun Liu: Writing – review & editing. free AI platform to predict and design materials, npj Comput. Mater. 10 (2024) 59.
Mingli Yang: Writing – review & editing. Yanjing Su: Supervision. [20] S.P. Collins, T.D. Daff, S.S. Piotrkowski, T.K. Woo, Materials design by
evolutionary optimization of functional groups in metal-organic frameworks, Sci.
Adv. 2 (2016) e1600954.
Declaration of competing interest [21] A. Seko, T. Maekawa, K. Tsuda, I. Tanaka, Machine learning with systematic
density-functional theory calculations: application to melting temperatures of
single-and binary-component solids, Phys. Rev. B 89 (2014) 54303.
The authors declare that they have no known competing financial [22] D. Xue, P.V. Balachandran, J. Hogden, J. Theiler, D. Xue, T. Lookman, Accelerated
interests or personal relationships that could have appeared to influence search for materials with targeted properties by adaptive design, Nat. Commun. 7
the work reported in this paper. (2016) 1–9.
[23] D. Xue, P. V Balachandran, R. Yuan, T. Hu, X. Qian, E.R. Dougherty, T. Lookman,
Accelerated search for BaTiO3-based piezoelectrics with vertical morphotropic
Acknowledgements phase boundary using Bayesian learning, Proc. Natl. Acad. Sci. 113 (2016)
13301–13306.
[24] R. Yuan, Z. Liu, P. V Balachandran, D. Xue, Y. Zhou, X. Ding, J. Sun, D. Xue,
This work is financially supported by the National Key Research and T. Lookman, Accelerated discovery of large electrostrains in BaTiO3-based
Development Program of China (2022YFB3707502), National Natural piezoelectrics using active learning, Adv. Mater. 30 (2018) 1702884.
Science Foundation of China (92270001, 52201061), Guangdong Basic [25] Y. Tian, R. Yuan, D. Xue, Y. Zhou, X. Ding, J. Sun, T. Lookman, Role of uncertainty
estimation in accelerating materials development via active learning, J. Appl.
and Applied Basic Research Foundation (2023A1515140101).
Phys. 128 (2020) 014103.
[26] Y. Tian, R. Yuan, D. Xue, Y. Zhou, Y. Wang, X. Ding, J. Sun, T. Lookman,
References Determining multi-component phase diagrams with desired characteristics using
active learning, Adv. Sci. 8 (2021) 2003165.
[27] H. Zhang, H. Fu, S. Zhu, W. Yong, J. Xie, Machine learning assisted composition
[1] J. Xie, Y. Su, D. Xue, X. Jiang, H. Fu, Machine learning for materials research and
effective design for precipitation strengthened copper alloys, Acta Mater. 215
development, Acta. Met. Sin. 57 (2021) 1343–1361.
(2021) 117118.
[2] Materials Data Curation System, (n.d.). [Link]
[28] X. Yuan, M. Suvarna, J.Y. Lim, J. P
erez-Ramírez, X. Wang, Y.S. Ok, Active
[3] S. Liu, Y. Su, H. Yin, D. Zhang, J. He, H. Huang, X. Jiang, X. Wang, H. Gong, Z. Li,
learning-based guided synthesis of engineered biochar for CO2 capture, Environ.
H. Xiu, J. Wan, X. Zhang, An infrastructure with user-centered presentation data
Sci. Technol. 58 (2024) 6628–6636.
model for integrated management of materials data and services, npj Comput.
[29] S. Xu, J. Li, P. Cai, X. Liu, B. Liu, X. Wang, Self-improving photosensitizer
Mater. 7 (2021) 88.
discovery system via Bayesian search with first-principle simulations, J. Am.
[4] H. Gong, J. He, X. Zhang, L. Duan, Z. Tian, W. Zhao, F. Gong, T. Liu, Z. Wang,
Chem. Soc. 143 (2021) 19769–19777.
H. Zhao, W. Jia, L. Zhang, X. Jiang, W. Chen, S. Liu, H. Xiu, W. Yang, J. Wan,
[30] J. Xiong, J.-C. He, X.-S. Leng, T.-Y. Zhang, Gaussian process regressions on hot
A repository for the publication and sharing of heterogeneous materials data, Sci.
deformation behaviors of FGH98 nickel-based powder superalloy, J. Mater. Sci.
Data 9 (2022) 787.
Technol. 146 (2023) 177–185.
[5] G. Klimeck, M. McLennan, S.P. Brophy, G.B. Adams III, M.S. Lundstrom, nanohub.
[31] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444.
org: advancing education and research in nanotechnology, Comput. Sci. Eng. 10
[32] K. Choudhary, B. DeCost, C. Chen, A. Jain, F. Tavazza, R. Cohn, C.W. Park,
(2008) 17–23.
A. Choudhary, A. Agrawal, S.J.L. Billinge, E. Holm, S.P. Ong, C. Wolverton, Recent
[6] J.G. Kaufman, E.F. Begley, MatML: a data interchange markup language: this
advances and applications of deep learning methods in materials science, npj
article updates and expands upon previously published information on MatML, a
Comput. Mater. 8 (2022) 59.
new extensible markup language designed specifically for the exchange of
[33] H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong, W. Duan, Y. Xu, Deep-learning
materials information. In particular, it describes the s, Adv. Mater. Process. 161
density functional theory Hamiltonian for efficient ab initio electronic-structure
(2003) 35–37.
calculation, Nat. Comput. Sci. 2 (2022) 367–377.
[7] C.W. Andersen, R. Armiento, E. Blokhin, G.J. Conduit, S. Dwaraknath, M.L. Evans,

Fekete, A. Gopakumar, S. Grazulis, A. Merkys, F. Mohamed, C. Oses, G. Pizzi, [34] X. Gong, S.G. Louie, W. Duan, Y. Xu, Generalizing deep learning electronic
A.
structure calculation to the plane-wave basis, Nat. Comput. Sci. 4 (2024) 752–760.
G. Rignanese, M. Scheidgen, L. Talirz, C. Toher, D. Winston, R. Aversa,
[35] Y. Wang, Y. Li, Z. Tang, H. Li, Z. Yuan, H. Tao, N. Zou, T. Bao, X. Liang, Z. Chen,
K. Choudhary, P. Colinet, S. Curtarolo, D.D. Stefano, C. Draxl, S. Er, M. Esters,
S. Xu, C. Bian, Z. Xu, C. Wang, C. Si, W. Duan, Y. Xu, Universal materials model of
M. Fornari, M. Giantomassi, M. Govoni, G. Hautier, V. Hegde, M.K. Horton,
deep-learning density functional theory Hamiltonian, Sci. Bull. 69 (2024)
P. Huck, G. Huhs, J. Hummelshøj, A. ariryaa, B. Kozinsky, S. Kumbhar, M. Liu,
2514–2521.
N. Marzari, A.J. Morris, A.A. Mostofi, K.A. Persson, G. Petretto, T. Purcell, F. Ricci,
[36] Z. Tang, H. Li, P. Lin, X. Gong, G. Jin, L. He, H. Jiang, X. Ren, W. Duan, Y. Xu,
F. Rose, M. Scheffler, D. Speckhard, M. Uhrin, A. Vaitkus, P. Villars,
A deep equivariant neural network approach for efficient hybrid density
D. Waroquiers, C. Wolverton, M. Wu, X. Yang, OPTIMADE, an API for exchanging
functional calculations, Nat. Commun. 15 (2024) 8815.
materials data, Sci. Data 8 (2021) 217.
[37] M. Weng, Z. Wang, G. Qian, Y. Ye, Z. Chen, X. Chen, S. Zheng, F. Pan, Identify
[8] Materials Genome Initiative for Global Competitiveness, (n.d.). [Link]
crystal structures by a new paradigm based on graph theory for building materials
.gov/sites/default/files/documents/materials_genome_initiative-fi[Link].
big data, Sci. China Chem. 62 (2019) 982–986.
[9] A. Agrawal, A. Choudhary, Perspective: materials informatics and big data:
[38] S. Li, Z. Chen, Z. Wang, M. Weng, J. Li, M. Zhang, J. Lu, K. Xu, F. Pan, Graph-based
realization of the “fourth paradigm” of science in materials science, APL Mater. 4
discovery and analysis of atomic-scale one-dimensional materials, Natl. Sci. Rev. 9
(2016) 053208.
(2022) nwac028.
[10] D. Morgan, R. Jacobs, Opportunities and challenges for machine learning in
materials science, Annu. Rev. Mater. Res. 50 (2020) 71–103.

21
X. Jiang et al. Review of Materials Research 1 (2025) 100010

[39] W. Zhang, M. Weng, M. Zhang, Z. Chen, B. Wang, S. Li, F. Pan, Rapid mining of [69] K. Hippalgaonkar, Q. Li, X. Wang, J.W. Fisher III, J. Kirkpatrick, T. Buonassisi,
fast ion conductors via subgraph isomorphism matching, J. Am. Chem. Soc. 146 Knowledge-integrated machine learning for materials: lessons from gameplaying
(2024) 18535–18543. and robotics, Nat. Rev. Mater. 8 (2023) 241–260.
[40] L. Chen, B. Wang, W. Zhang, S. Zheng, Z. Chen, M. Zhang, C. Dong, F. Pan, S. Li, [70] M.C. Swain, J.M. Cole, ChemDataExtractor: a toolkit for automated extraction of
Crystal structure assignment for Unknown compounds from X-ray diffraction chemical information from the scientific literature, J. Chem. Inf. Model. 56 (2016)
patterns with deep learning, J. Am. Chem. Soc. 146 (2024) 8098–8109. 1894–1904.
[41] H. Zhang, Z. Wang, J. Ren, J. Liu, J. Li, Ultra-fast and accurate binding energy [71] W. Wang, X. Jiang, S. Tian, P. Liu, D. Dang, Y. Su, T. Lookman, J. Xie, Automated
prediction of shuttle effect-suppressive sulfur hosts for lithium-sulfur batteries pipeline for superalloy data by text mining, npj Comput. Mater. 8 (2022) 9.
using machine learning, Energy Storage Mater. 35 (2021) 88–98. [72] W. Wang, X. Jiang, S. Tian, P. Liu, T. Lookman, Y. Su, J. Xie, Alloy synthesis and
[42] Z. Wang, H. Zhang, J. Ren, X. Lin, T. Han, J. Liu, J. Li, Predicting adsorption ability processing by semi-supervised text mining, npj Comput. Mater. 9 (2023) 183.
of adsorbents at arbitrary sites for pollutants using deep transfer learning, npj [73] M.P. Polak, D. Morgan, Extracting accurate materials data from research papers
Comput. Mater. 7 (2021) 19. with conversational language models and prompt engineering, Nat. Commun. 15
[43] S. Wu, Z. Wang, H. Zhang, J. Cai, J. Li, Deep learning accelerates the discovery of (2024) 1569.
two-dimensional catalysts for hydrogen evolution reaction, Energy & Environ. [74] V. Tshitoyan, J. Dagdelen, L. Weston, A. Dunn, Z. Rong, O. Kononova,
Mater. 6 (2023) e12259. K.A. Persson, G. Ceder, A. Jain, Unsupervised word embeddings capture latent
[44] D.P. Kingma, Auto-encoding Variational Bayes, 2013. ArXiv Prepr. knowledge from materials science literature, Nature 571 (2019) 95–98.
ArXiv1312.6114. [75] Z. Pei, J. Yin, P.K. Liaw, D. Raabe, Toward the design of ultrahigh-entropy alloys
[45] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, via mining six million texts, Nat. Commun. 14 (2023) 54.
A. Courville, Y. Bengio, Generative adversarial networks, Commun. ACM 63 [76] D.A. Boiko, R. MacKnight, B. Kline, G. Gomes, Autonomous chemical research
(2020) 139–144. with large language models, Nature 624 (2023) 570–578.
[46] B.A. Sutton Rs, Reinforcement Learning: an Introduction, MIT Press, 1998. [77] Y. Kang, J. Kim, ChatMOF: an artificial intelligence system for predicting and
[47] J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Adv. Neural generating metal-organic frameworks using large language models, Nat. Commun.
Inf. Process. Syst. 33 (2020) 6840–6851. 15 (2024) 4705.
[48] Z. Zhou, Y. Shang, X. Liu, Y. Yang, A generative deep learning framework for [78] C. Kuenneth, R. Ramprasad, polyBERT: a chemical language model to enable fully
inverse design of compositionally complex bulk metallic glasses, npj Comput. machine-driven ultrafast polymer informatics, Nat. Commun. 14 (2023) 4099.
Mater. 9 (2023) 15. [79] S. Tian, X. Jiang, W. Wang, Z. Jing, C. Zhang, C. Zhang, T. Lookman, Y. Su, Steel
[49] Y. Shang, Z. Zhou, R. Han, H. Wang, X. Liu, Y. Yang, Customized design of design based on a large language model, Acta Mater. 285 (2025) 120663.
amorphous solids by generative deep learning, Innov. Mater. 2 (2024) 100071. [80] J. He, X. Su, C. Wang, J. Li, Y. Hou, Z. Li, C. Liu, D. Xue, J. Cao, Y. Su, Machine
[50] Z.Q. Chen, Y.H. Shang, X.D. Liu, Y. Yang, Accelerated discovery of eutectic learning assisted predictions of multi-component phase diagrams and fine
compositionally complex alloys by generative machine learning, npj Comput. boundary information, Acta Mater. 240 (2022) 118341.
Mater. 10 (2024) 204. [81] K. Terayama, R. Tamura, Y. Nose, H. Hiramatsu, H. Hosono, Y. Okuno, K. Tsuda,
[51] Y. Xian, P. Dang, Y. Tian, X. Jiang, Y. Zhou, X. Ding, J. Sun, T. Lookman, D. Xue, Efficient construction method for phase diagrams using uncertainty sampling,
Compositional design of multicomponent alloys using reinforcement learning, Phys. Rev. Mater. 3 (2019) 33802.
Acta Mater. 274 (2024) 120017. [82] Y. Zhang, D.W. Apley, W. Chen, Bayesian optimization for materials design with
[52] D.X. Pengfei Dang, Jinlong Hu, Yuehui Xian, Cheng Li, Yumei Zhou, mixed quantitative and qualitative variables, Sci. Rep. 10 (2020) 4924.
Xiangdong Ding, Jun Sun, Elastocaloric thermal battery: ultrahigh heat-storage [83] X. Pei, J. Pei, H. Hou, Y. Zhao, Optimizing casting process using a combination of
capacity based on generative learning-designed phase-change alloys, Adv. Mater. small data machine learning and phase-field simulations, npj Comput. Mater. 11
(2025) 2412198. (2025) 27.
[53] C. Zhong, J. Zhang, Y. Wang, Y. Long, P. Zhu, J. Liu, K. Hu, J. Chen, X. Lin, High- [84] C. Kim, R. Batra, L. Chen, H. Tran, R. Ramprasad, Polymer design using genetic
performance diffusion model for inverse design of high Tc superconductors with algorithm and machine learning, Comput. Mater. Sci. 186 (2021) 110067.
effective doping and accurate stoichiometry, InfoMat 6 (2024) e12519. [85] P. Nikolaev, D. Hooper, N. Perea-Lopez, M. Terrones, B. Maruyama, Discovery of
[54] C. Zhong, Y. Wang, Y. Long, J. Liu, K. Hu, J. Zhang, J. Chen, X. Lin, Enhancing wall-selective carbon nanotube growth conditions via automated experimentation,
superconductor critical temperature prediction: a novel machine learning ACS Nano 8 (2014) 10214–10222.
approach integrating dopant recognition, ACS Appl. Mater. Interfaces 16 (2024) [86] P. Nikolaev, D. Hooper, F. Webber, R. Rao, K. Decker, M. Krein, J. Poleski,
60472–60481. R. Barto, B. Maruyama, Autonomy in materials research: a case study in carbon
[55] B. Peng, Y. Wei, Y. Qin, J. Dai, Y. Li, A. Liu, Y. Tian, L. Han, Y. Zheng, P. Wen, nanotube growth, npj Comput. Mater. 2 (2016) 1–6.
Machine learning-enabled constrained multi-objective design of architected [87] D.P. Tabor, L.M. Roch, S.K. Saikin, C. Kreisbeck, D. Sheberla, J.H. Montoya,
materials, Nat. Commun. 14 (2023) 6630. S. Dwaraknath, M. Aykol, C. Ortiz, H. Tribukait, C. Amador-Bedolla, C.J. Brabec,
[56] Y. Wang, W. Sha, M. Xiao, C.-W. Qiu, L. Gao, Deep-learning-enabled intelligent B. Maruyama, K.A. Persson, A. Aspuru-Guzik, Accelerating the discovery of
design of thermal metamaterials, Adv. Mater. 35 (2023) 2302387. materials for clean energy in the era of smart automation, Nat. Rev. Mater. 3
[57] C.M. Childs, N.R. Washburn, Embedding domain knowledge for machine learning (2018) 5–20.
of complex material systems, MRS Commun 9 (2019) 806–820. [88] B.P. MacLeod, F.G.L. Parlane, T.D. Morrissey, F. H€ase, L.M. Roch, K.E. Dettelbach,
[58] S. Kirklin, J.E. Saal, B. Meredig, A. Thompson, J.W. Doak, M. Aykol, S. Rühl, R. Moreira, L.P.E. Yunker, M.B. Rooney, J.R. Deeth, G.J. Ng, H. Situ, R.H. Zhang,
C. Wolverton, The Open Quantum Materials Database (OQMD): assessing the M.S. Elliott, T.H. Haley, D.J. Dvorak, A. Aspuru-Guzik, J.E. Hein, C.P. Berlinguette,
accuracy of DFT formation energies, npj Comput. Mater. 1 (2015) 1–15. Self-driving laboratory for accelerated discovery of thin-film materials, Sci. Adv. 6
[59] R. Yuan, Y. Tian, D. Xue, D. Xue, Y. Zhou, X. Ding, J. Sun, T. Lookman, Accelerated (2020) eaaz8867.
search for BaTiO3-based ceramics with large energy storage at low fields using [89] B.P. MacLeod, F.G.L. Parlane, C.C. Rupnow, K.E. Dettelbach, M.S. Elliott,
machine learning and experimental design, Adv. Sci. 6 (2019) 1901395. T.D. Morrissey, T.H. Haley, O. Proskurin, M.B. Rooney, N. Taherimakhsousi,
[60] R. Yuan, D. Xue, D. Xue, J. Li, X. Ding, J. Sun, T. Lookman, Knowledge-based D.J. Dvorak, H.N. Chiu, C.E.B. Waizenegger, K. Ocean, M. Mokhtari,
descriptor for the compositional dependence of the phase transition in BaTiO3- C.P. Berlinguette, A self-driving laboratory advances the Pareto front for material
based ferroelectrics, ACS Appl. Mater. Interfaces 12 (2020) 44970–44980. properties, Nat. Commun. 13 (2022) 995.
[61] Y. Zhang, C. Wen, C. Wang, S. Antonov, D. Xue, Y. Bai, Y. Su, Phase prediction in [90] B. Burger, P.M. Maffettone, V. V Gusev, C.M. Aitchison, Y. Bai, X. Wang, X. Li,
high entropy alloys with a rational selection of materials descriptors and machine B.M. Alston, B. Li, R. Clowes, N. Rankin, B. Harris, R.S. Sprick, A.I. Copper,
learning models, Acta Mater. 185 (2020) 528–539. A mobile robotic chemist, Nature 583 (2020) 237–241.
[62] Y. Zhang, C. Wen, P. Dang, T. Lookman, D. Xue, Y. Su, Toward ultra-high strength [91] W. Zhang, M. Yu, T. Liu, M. Cong, X. Liu, H. Yang, Y. Bai, Q. Zhu, S. Zhang, H. Gu,
high entropy alloys via feature engineering, J. Mater. Sci. Technol. 200 (2024) X. Wu, Z. Zhang, Y. Wu, H. Tian, X. Li, W. Zhu, A.I. Copper, Accelerated discovery
243–252. of molecular nanojunction photocatalysts for hydrogen evolution by using
[63] R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, L.M. Ghiringhelli, SISSO: a automated screening and flow synthesis, Nat. Synth. 3 (2024) 595–605.
compressed-sensing method for identifying the best low-dimensional descriptor in [92] Q. Zhu, Y. Huang, D. Zhou, L. Zhao, L. Guo, R. Yang, Z. Sun, M. Luo, F. Zhang,
an immensity of offered candidates, Phys. Rev. Mater. 2 (2018) 83802. H. Xiao, X. Tang, X. Zhang, T. Song, X. Li, B. Chong, J. Zhou, Y. Zhang, B. Zhang,
[64] C.J. Bartel, C. Sutton, B.R. Goldsmith, R. Ouyang, C.B. Musgrave, J. Cao, G. Zhang, S. Wang, G. Ye, W. Zhang, H. Zhao, S. Cong, H. Li, L. Ling,
L.M. Ghiringhelli, M. Scheffler, New tolerance factor to predict the stability of Z. Zhang, W. Shang, J. Jiang, Y. Luo, Automated synthesis of oxygen-producing
perovskite oxides and halides, Sci. Adv. 5 (2019) eaav0693. catalysts from Martian meteorites by a robotic AI chemist, Nat. Synth. 3 (2024)
[65] T. Wang, J. Hu, R. Ouyang, Y. Wang, Y. Huang, S. Hu, W.-X. Li, Nature of metal- 319–328.
support interaction for metal catalysts on oxide supports, Science 386 (2024) [93] R.W. Epps, M.S. Bowen, A.A. Volk, K. Abdel-Latif, S. Han, K.G. Reyes, A. Amassian,
915–920. M. Abolhasani, Artificial chemist: an autonomous quantum dot synthesis bot, Adv.
[66] C. Shen, C. Wang, X. Wei, Y. Li, S. van der Zwaag, W. Xu, Physical metallurgy- Mater. 32 (2020) 2001626.
guided machine learning and artificial intelligent design of ultrahigh-strength [94] A. Slattery, Z. Wen, P. Tenblad, J. Sanjos
e-Orduna, D. Pintossi, T. den Hartog,
stainless steel, Acta Mater. 179 (2019) 201–214. T. No€el, Automated self-optimization, intensification, and scale-up of
[67] L. Jiang, H. Fu, H. Zhang, J. Xie, Physical mechanism interpretation of photocatalysis in flow, Science 383 (2024) eadj1817.
polycrystalline metals' yield strength via a data-driven method: a novel Hall–Petch [95] A.E. Gongora, B. Xu, W. Perry, C. Okoye, P. Riley, K.G. Reyes, E.F. Morgan,
relationship, Acta Mater. 231 (2022) 117868. K.A. Brown, A Bayesian experimental autonomous researcher for mechanical
[68] S. Liu, T. Wen, A.S.L.S. Pattamatta, D.J. Srolovitz, A prompt-engineered large design, Sci. Adv. 6 (2020) eaaz1708.
language model, deep learning workflow for materials classification, Mater. Today [96] P. Schwaller, B. Hoover, J.-L. Reymond, H. Strobelt, T. Laino, Extraction of organic
80 (2024) 240–249. chemistry grammar from unsupervised learning of chemical reactions, Sci. Adv. 7
(2021) eabe4166.

22
X. Jiang et al. Review of Materials Research 1 (2025) 100010

[97] N.J. Szymanski, B. Rendy, Y. Fei, R.E. Kumar, T. He, D. Milsted, M.J. McDermott, [123] M. Umehara, H.S. Stein, D. Guevarra, P.F. Newhouse, D.A. Boyd, J.M. Gregoire,
M. Gallant, E.D. Cubuk, A. Merchant, An autonomous laboratory for the Analyzing machine learning models to accelerate generation of fundamental
accelerated synthesis of novel materials, Nature 624 (2023) 86–91. materials insights, npj Comput. Mater. 5 (2019) 34.
[98] B. Mortazavi, X. Zhuang, T. Rabczuk, A. V Shapeev, Atomistic modeling of the [124] P.L. Craig, Modeling Software for Materials 4.0, 2021.
mechanical properties: the rise of machine learning interatomic potentials, Mater. [125] W.-Y. Wang, J. Yin, Z. Chai, X. Chen, W. Zhao, J. Lu, F. Sun, Q. Jia, X. Gao,
Horiz. 10 (2023) 1956–1968. B. Tang, X. Hui, H. Song, F. Xue, Z. Liu, J. Li, Big data-assisted digital twins for the
[99] K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for smart design and manufacturing of advanced materials: from atoms to products,
molecular and materials science, Nature 559 (2018) 547–555. J. Mater. Inf. 2 (2022) 1–27, [Link]
[100] J. Behler, M. Parrinello, Generalized neural-network representation of high- [126] Y. Yi, Y. Yan, X. Liu, Z. Ni, J. Feng, J. Liu, Digital twin-based smart assembly
dimensional potential-energy surfaces, Phys. Rev. Lett. 98 (2007) 146401. process design and application framework for complex products and its case study,
[101] H. Wang, L. Zhang, J. Han, E. Weinan, DeePMD-kit: a deep learning package for J. Manuf. Syst. 58 (2021) 94–107.
many-body potential energy representation and molecular dynamics, Comput. [127] Y. Xu, X. Liu, X. Cao, C. Huang, E. Liu, S. Qian, X. Liu, Y. Wu, F. Dong, C.-W. Qiu,
Phys. Commun. 228 (2018) 178–184. J. Qiu, K. Hua, W. Su, J. Wu, H. Xu, Y. Han, C. Fu, Z. Yin, M. Liu, R. Roepman,
[102] W. Jia, H. Wang, M. Chen, D. Lu, L. Lin, R. Car, E. Weinan, L. Zhang, Pushing the S. Dietmann, M. Virta, F. Kengara, Z. Zhang, L. Zhang, T. Zhao, J. Dai, J. Yang,
limit of molecular dynamics with ab initio accuracy to 100 million atoms with L. Lan, M. Luo, Z. Liu, T. An, B. Zhang, X. He, S. Cong, X. Liu, W. Zhang, J.P. Lewis,
machine learning, in: International Conference for High Performance Computing, J.M. Tiedje, Q. Wang, Z. An, F. Wang, L. Zhang, T. Huang, C. Lu, Z. Cai, F. Wang,
Networking, Storage and Analysis, 2020, pp. 1–14, [Link] J. Zhang, Artificial intelligence: a powerful paradigm for scientific research,
SC41405.2020.00009. Atlanta, GA, USA. Innovation 2 (2021) 100179, [Link]
[103] Linfeng Zhang, (n.d.). [Link] [128] N. Nicola, Can artificial intelligence create the next wonder material, Nature 533
[104] Z. Guo, D. Lu, Y. Yan, S. Hu, R. Liu, G. Tan, N. Sun, W. Jiang, L. Liu, Y. Chen, (2016) 22–25.
L. Zhang, M. Chen, H. Wang, W. Jia, Extending the limit of molecular dynamics [129] A.R. Oganov, C.J. Pickard, Q. Zhu, R.J. Needs, Structure prediction drives
with ab initio accuracy to 10 billion atoms, in: Proc. 27th ACM SIGPLAN Symp. materials discovery, Nat. Rev. Mater. 4 (2019) 331–348.
Princ. Pract. Parallel Program, 2022, pp. 205–218. [130] X. Zhang, L. Wang, J. Helwig, Y. Luo, C. Fu, Y. Xie, M. Liu, Y. Lin, Z. Xu, K. Yan,
[105] Z. Fan, Y. Wang, P. Ying, K. Song, J. Wang, Y. Wang, Z. Zeng, K. Xu, E. Lindgren, K. Adams, M. Weiler, X. Li, T. Fu, Y. Wang, A. Strasser, H. Yu, Y. Xie, X. Fu, S. Xu,
J.M. Rahm, A.J. Gabourie, J. Liu, H. Dong, J. Wu, Y. Chen, Z. Zhong, J. Sun, Y. Liu, Y. Du, A. Saxton, H. Ling, H. Lawrence, H. St€ark, S. Gui, C. Edwards, N. Gao,
P. Erhart, Y. Su, T. Ala-Nissila, GPUMD: a package for constructing accurate A. Ladera, T. Wu, E.F. Hofgard, A.M. Tehrani, R. Wang, A. Daigavane, M. Bohde,
machine-learned potentials and performing highly efficient atomistic simulations, J. Kurtin, Q. Huang, T. Phung, M. Xu, C.K. Joshi, S.V. Mathis, K. Azizzadenesheli,
J. Chem. Phys. 157 (2022) 114801. A. Fang, A. Aspuru-Guzik, e. Bekkers, M. Bronstein, M. Zitnik, A. Anandkurmar,
[106] Y. Li, Z. Tang, Z. Chen, M. Sun, B. Zhao, H. Li, H. Tao, Z. Yuan, W. Duan, Y. Xu, S. Ermon, P. Lio, R. Yu, S. Günnemann, J. Leskovec, H. Ji, J. Sun, R. Barzilay,
Neural-network density functional theory based on variational energy T. Jaakkola, C.W. Coley, X. Qian, X. Qian, T. Smidt, S. Ji, Artificial Intelligence for
minimization, Phys. Rev. Lett. 133 (2024) 76401. Science in Quantum, Atomistic, and Continuum Systems, 2023. ArXiv Prepr.
[107] A.M. Kiirikki, H.S. Antila, L.S. Bort, P. Buslaev, F. Favela-Rosales, T.M. Ferreira, ArXiv2307.08423.
P.F.J. Fuchs, R. Garcia-Fandino, I. Gushchin, B. Kav, N. Kucerka, P. Kula, M. Kurki, [131] H. Wang, T. Fu, Y. Du, W. Gao, K. Huang, Z. Liu, P. Chandak, S. Liu, P. Van
A. Kuzmin, A. Lalitha, F. Lolicato, J.J. Madsen, M.S. Miettinen, C. Mingham, Katwyk, A. Deac, A. Anandkumar, K. Bergen, C.P. Gomes, S. Ho, P. Kohli,
L. Monticelli, R. Nencini, A.M. Nesterenko, T.J. Piggot, A.
Pi~ neiro, N. Reuter, J. Lasenby, J. Leskovec, T. Liu, A. Manrai, D. Marks, B. Ramsundar, L. Song, J. Sun,
S. Samantray, F. Su
arez-Lest

on, R. Talandashti, O.H.S. Ollila, Overlay databank J. Tang, P. Velickovi
c, M. Welling, L. Zhang, C.W. Coley, Y. Vengio, M. Zitnik,
unlocks data-driven analyses of biomolecules for all, Nat. Commun. 15 (2024) Scientific discovery in the age of artificial intelligence, Nature 620 (2023) 47–60.
1136. [132] National Science and Technology Council, Materials Genome Initiative Strategic
[108] N. Yao, X. Chen, Z.-H. Fu, Q. Zhang, Applying classical, ab initio, and machine- Plan, 2021.
learning molecular dynamics simulations to the liquid electrolyte for rechargeable [133] X. Gao, X. Lin, Q. Yan, Z. Wang, X. Yu, Y. Zhou, Y. Hu, W. Huang, Effect of Cu
batteries, Chem. Rev. 122 (2022) 10970–11021. content on microstructure and mechanical properties of in-situ $β$ phases
[109] Q. Zhang, Y. Yuan, J. Zhang, P. Fang, J. Pan, H. Zhang, T. Zhou, Q. Yu, X. Zou, reinforced Ti/Zr-based bulk metallic glass matrix composite by selective laser
Z. Sun, F. Yan, Machine learning-aided design of highly conductive anion melting (SLM), J. Mater. Sci. Technol. 67 (2021) 174–185.
exchange membranes for fuel cells and water electrolyzers, Adv. Mater. 36 (2024) [134] M.Y. Khalid, Z.U. Arif, A. Tariq, M. Hossain, R. Umer, M. Bodaghi, 3D printing of
2404981. active mechanical metamaterials: a critical review, Mater. Des. (2024) 113305.
[110] D. Xu, Q. Zhang, X. Huo, Y. Wang, M. Yang, Advances in data-assisted high- [135] T.M. Smith, C.A. Kantzos, N.A. Zarkevich, B.J. Harder, M. Heczko, P.R. Gradl,
throughput computations for material design, Mater. Genome Eng. Adv. 1 (2023) A.C. Thompson, M.J. Mills, T.P. Gabb, J.W. Lawson, A 3D printable alloy designed
e11, [Link] for extreme environments, Nature 617 (2023) 513–518.
[111] A. Merchant, S. Batzner, S.S. Schoenholz, M. Aykol, G. Cheon, E.D. Cubuk, Scaling [136] S. Yu, H. Chai, Y. Xiong, M. Kang, C. Geng, Y. Liu, Y. Chen, Y. Zhang, Q. Zhang,
deep learning for materials discovery, Nature 624 (2023) 80–85. C. Li, H. Wei, Y. Zhao, F. Yu, A. Lu, Studying complex evolution of hyperelastic
[112] G. Wang, L. Peng, K. Li, L. Zhu, J. Zhou, N. Miao, Z. Sun, ALKEMIE: an intelligent materials under external field stimuli using artificial neural networks with
computational platform for accelerating materials discovery and design, Comput. spatiotemporal features in a small-scale dataset, Adv. Mater. 34 (2022) 2200908.
Mater. Sci. 186 (2021) 110064. [137] Z. Liu, M. Ding, J. Xie, Advancements in digital manufacturing for metal 3D
[113] X. Yang, Z. Wang, X. Zhao, J. Song, M. Zhang, H. Liu, MatCloud: a high-throughput printing, Acta. Met. Sin. 60 (2023) 569–584.
computational infrastructure for integrated management of materials simulation, [138] E. Liao, Y. Zhang, Y. Su, C. Zhang, C. Geng, C. Li, X. Liu, Y. Liu, A. Lu, Active
data and resources, Comput. Mater. Sci. 146 (2018) 319–333. mixing of two-component viscoelastic silicone ink at molecular level for
[114] W.Y. Wang, S. Zhang, G. Li, J. Lu, Y. Ren, X. Wang, X. Gao, Y. Su, H. Song, J. Li, spatiotemporally controlled 3D/4D printing of cellular silicones, Polymer (Guildf).
Artificial intelligence enabled smart design and manufacturing of advanced 302 (2024) 127053.
materials: the endless Frontier in AIþ era, Mater. Genome Eng. Adv. 2 (2024) e56. [139] F. Zhang, Z. Li, M. Xu, S. Wang, N. Li, J. Yang, A review of 3D printed porous
[115] W.Y. Wang, J. Li, W. Liu, Z.-K. Liu, Integrated computational materials ceramics, J. Eur. Ceram. Soc. 42 (2022) 3351–3373.
engineering for advanced materials: a brief review, Comput. Mater. Sci. 158 [140] T. Ma, Y. Zhang, K. Ruan, H. Guo, M. He, X. Shi, Y. Guo, J. Kong, J. Gu, Advances
(2019) 42–48. in 3D printing for polymer composites: a review, InfoMat 6 (2024) e12568.
[116] S.M. Hashemi, S. Parvizi, H. Baghbanijavid, A.T.L. Tan, M. Nematollahi, [141] Z. Ma, J. Wang, L. Qin, A. Chortos, 3D printing in biofabrication: from surface
A. Ramazani, N.X. Fang, M. Elahinia, Computational modelling of process– textures to biological engineering, Chem. Eng. J. (2024) 156477.
structure–property–performance relationships in metal additive manufacturing: a [142] Z. Zhu, D.W.H. Ng, H.S. Park, M.C. McAlpine, 3D-printed multifunctional
review, Int. Mater. Rev. 67 (2022) 1–46. materials enabled by artificial-intelligence-assisted fabrication technologies, Nat.
[117] X. Wang, W. Xiong, Uncertainty quantification and composition optimization for Rev. Mater. 6 (2021) 27–47.
alloy additive manufacturing through a CALPHAD-based ICME framework, npj [143] Z. Ren, L. Gao, S.J. Clark, K. Fezzaa, P. Shevchenko, A. Choi, W. Everhart,
Comput. Mater. 6 (2020) 188. A.D. Rollett, L. Chen, T. Sun, Machine learning–aided real-time detection of
[118] J. Qi, X. Fan, D.I. Hoyos, M. Widom, P.K. Liaw, J. Poon, Integrated design of keyhole pore generation in laser powder bed fusion, Science 379 (2023) 89–94.
aluminum-enriched high-entropy refractory B2 alloys with synergy of high [144] J.A. Lee, J. Park, M.J. Sagong, S.Y. Ahn, J.-W. Cho, S. Lee, H.S. Kim, Active
strength and ductility, Sci. Adv. 10 (2024) eadq0083. learning framework to optimize process parameters for additive-manufactured Ti-
[119] Y. Du, R. Schmid-Fetzer, J. Wang, S. Liu, J. Wang, Z. Jin, Computational Design of 6Al-4V with high strength and ductility, Nat. Commun. 16 (2025) 931.
Engineering Materials: Fundamentals and Case Studies, Cambridge University [145] M. Chen, Q. Du, R. Shi, H. Fu, Z. Liu, J. Xie, Phase field simulation of
Press, 2023. microstructure evolution and process optimization during homogenization of
[120] W.Y. Wang, B. Tang, D. Lin, C. Zou, Y. Zhang, S.-L. Shang, Q. Guan, J. Gao, L. Fan, additively manufactured Inconel 718 alloy, Front. Mater. 9 (2022) 1043249.
H. Kou, H. Song, J. Ma, X. Hui, M.C. Gao, Z. Liu, J. Li, A brief review of data-driven [146] B. Zong, J. Li, C. Zhou, P. Wang, B. Tang, R. Yuan, Enhancing creep rupture life
ICME for intelligently discovering advanced structural metal materials: insight prediction of high-temperature titanium alloys using convolutional neural
into atomic and electronic building blocks, J. Mater. Res. 35 (2020) 872–889. networks, Mater, Genome Eng. Adv. 2 (2024) e68.
[121] V. Attari, D. Khatamsaz, D. Allaire, R. Arroyave, Towards inverse microstructure- [147] F. Xiang, S. Zhou, Y. Zuo, F. Tao, Digital twin driven end-face defect control
centered materials design using generative phase-field modeling and deep method for hot-rolled coil with cloud-edge collaboration, IEEE Trans. Ind. Inf. 19
variational autoencoders, Acta Mater. 259 (2023) 119204. (2022) 1674–1682, [Link]
[122] W.Y. Wang, P. Li, D. Lin, B. Tang, J. Wang, Q. Guan, Q. Ye, H. Dai, J. Gao, X. Fan, [148] W. Luo, T. Hu, W. Zhu, F. Tao, Digital twin modeling method for CNC machine
H. Kou, H. Song, F. Zhou, J. Ma, Z. Liu, J. Li, W. Liu, DID code: a bridge connecting tool, in: 2018 IEEE 15th Int. Conf. Networking, Sens. Control, 2018, pp. 1–4.
the materials genome engineering database with inheritable integrated intelligent [149] S. Yu, N. Ran, J. Liu, Large-language models: the game-changers for materials
manufacturing, Engineering 6 (2020) 612–620. science research, Artif. Intell. Chem. 2 (2024) 100076.

23
X. Jiang et al. Review of Materials Research 1 (2025) 100010

[150] T. Song, M. Luo, X. Zhang, L. Chen, Y. Huang, J. Cao, Q. Zhu, D. Liu, B. Zhang, development platform powered by large language models, Nat. Commun. 15
G. Zou, G. Zhang, F. Zhang, W. Shang, Y. Fu, J. Jiang, Y. Luo, A multiagent-driven (2024) 10160.
robotic AI chemist enabling autonomous chemical research on demand, J. Am. [153] S. Zhang, W.Y. Wang, X. Wang, G. Li, Y. Ren, X. Gao, F. Sun, B. Tang, H. Song,
Chem. Soc. (2024), [Link] J. Li, Large language models enabled intelligent microstructure optimization and
[151] M.H. Prince, H. Chan, A. Vriza, T. Zhou, V.K. Sastry, Y. Luo, M.T. Dearing, defects classification of welded titanium alloys, J. Mater. Inforatics 4 (2024) 34,
R.J. Harder, R.K. Vasudevan, M.J. Cherukara, Opportunities for retrieval and tool [Link]
augmented large language models in scientific facilities, npj Comput. Mater. 10 [154] K.T. Butler, K. Choudhary, G. Csanyi, A.M. Ganose, S. V Kalinin, D. Morgan,
(2024) 251. Setting standards for data driven materials science, npj Comput. Mater. 10 (2024)
[152] Y. Ruan, C. Lu, N. Xu, Y. He, Y. Chen, J. Zhang, J. Xuan, J. Pan, Q. Fang, H. Gao, 231.
X. Shen, N. Ye, Q. Zhang, Y. Mo, An automatic end-to-end chemical synthesis

24