Journal of Materials Science & Technology 271 (2026) 91–108

Contents lists available at ScienceDirect

Journal of Materials Science & Technology

journal homepage: [Link]/locate/jmst

Review Article

Deep learning-driven innovation in metallic materials: A

comprehensive review on microstructure analysis, property prediction,
and inverse design
Yulin Shengcao a,1, Xu Qin a,b,1, Qinghang Wang a,∗, Huabao Yang c,∗, Yanfu Chai c, Dabiao Xia d,
Bin Jiang e, Hyoung Seop Kim f,g,∗
a
School of Mechanical Engineering, Yangzhou University, Yangzhou 225127, China
b
IMDEA Materials Institute, C/Eric Kandel 2, Getafe, Madrid 28906, Spain
c
School of Mechanical and Electrical Engineering, Shaoxing University, Shaoxing 312000, China
d
Key Laboratory of Functional Materials and Devices for Informatics of Anhui Educational Institutions, Fuyang Normal University, Fuyang 236037, China
e
National Engineering Research Center for Mg Alloys, College of Materials Science and Engineering, Chongqing University, Chongqing 400044, China
f
Graduate Institute of Ferrous and Eco Materials Technology (GIFT), Pohang University of Science and Technology (POSTECH), Pohang 37673, South Korea
g
Department of Materials Science and Engineering, Pohang University of Science and Technology (POSTECH), Pohang 37673, South Korea

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

Article history: The data-driven paradigm is revolutionizing metallic materials research. This review synthesizes
Received 9 December 2025 how Deep Learning (DL) transforms the innovation chain, guided by the Process-Structure-Property-
Revised 24 January 2026
Performance framework. We analyze DL’s role in three interconnected stages: (1) Microstructure anal-
Accepted 10 February 2026
ysis, where DL automates high-throughput quantification and evolves toward generative 2D-to-3D mi-
Available online 12 February 2026
crostructure reconstruction to enable digital twins for simulation; (2) Property prediction, where models
Keywords: advance from single-modal correlations to physics-infused, multimodal frameworks that fuse heteroge-
Metallic materials neous data (images, spectra, text) with mechanistic simulations for predicting mechanical, fatigue, and
Deep learning corrosion behaviors; and (3) Inverse design, which matures from multi-objective-driven optimal com-
Microstructure analysis position search to microstructure generation in the high-dimensional process-structure-property space,
Property prediction generating physically consistent blueprints. Persistent bottlenecks—such as the “black-box” nature, small-
Inverse design
sample generalization, and integration gaps—are critically examined. We conclude with a roadmap toward
physics-informed architectures and integrated autonomous platforms for intelligent, closed-loop materi-
als discovery. This work provides a focused perspective on structural metallic materials, tracing a clear
path from automated characterization to reliable design.
© 2026 Published by Elsevier Ltd on behalf of The editorial office of Journal of Materials Science &
Technology.

1. Introduction [2,3]. This vision has been dramatically accelerated by the Mate-
rials Genome Initiative (MGI) and related efforts such as those by
The pursuit of advanced metallic materials is fundamentally the Center for Hierarchical Materials Design (CHiMaD), which ad-
governed by the Process-Structure-Property-Performance (PSPP) vocate for the convergence of high-throughput computation, exper-
paradigm, a cornerstone concept in materials science that frames imental data, and informatics to compress the materials develop-
design as the establishment of quantitative, invertible links across ment timeline [4,5]. Within this data-driven “fourth paradigm”, Ar-
this chain [1]. Pioneering work by Olson on the computational de- tificial Intelligence (AI) has proven powerful in mapping complex
sign of hierarchically structured materials laid the conceptual foun- PSPP relationships, with demonstrations in material synthesis and
dation for a systems-based approach, integrating thermodynam- property prediction for alloys showcasing its potential [6–11].
ics, kinetics, and mechanistic models into an iterative design loop AI refers to the development of machines and algorithms ca-
pable of mimicking human intelligence, for instance, by optimiz-
∗
ing actions to achieve specific objectives. Machine Learning (ML),
Corresponding authors.
as a branch of AI, enables systems to learn autonomously from
E-mail addresses: wangqinghang@[Link] (Q. Wang), yanghuabao5011@
[Link] (H. Yang), hskim@[Link] (H.S. Kim). given datasets (such as in board games, social network recommen-
1
These authors contributed equally to this work. dations, etc.) without being explicitly programmed. Deep Learning

[Link]
1005-0302/© 2026 Published by Elsevier Ltd on behalf of The editorial office of Journal of Materials Science & Technology.
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 1. Schematic showing an overview of artificial intelligence (AI), machine learning (ML), and deep learning (DL) methods and their applications in materials science and
engineering. DL was considered a part of machine learning, which was contained in the umbrella term artificial intelligence. Reproduced with permission [12]. Copyright
2022, Springer Nature.

(DL), in turn, falls under the domain of machine learning. Inspired tems that encompass process-oriented design and multi-objective
by the structure of biological brains, it employs multi-layer neural optimization, highlighting workflows that navigate the full PSPP
networks to address machine learning tasks. Fig. 1 illustrates the landscape.
hierarchical relationship among AI, ML, and DL, and presents sev- By synthesizing advances across this pipeline, we identify con-
eral key application areas of deep learning in the field of materials vergent trends—such as physics-infused neural networks, cross-
science and engineering [12]. domain transfer learning, and multi-scale data fusion—and persis-
DL offers unprecedented capability in decoding high- tent bottlenecks. We conclude not merely with a summary of algo-
dimensional, multimodal materials data—from micrographs and rithms, but with a forward-looking roadmap that emphasizes the
spectra to atomistic graphs—with minimal human intervention necessity of closing the loop between process control, structural
in feature design. While several excellent reviews have surveyed realization, and performance validation to transition from in silico
DL in materials science broadly [6,12,13], or focused on specific innovation to practical alloy deployment. This review thus seeks
aspects like potentials for phase diagrams [14] or segmentation of to serve as both a systematic technical reference and a strategic
particular alloy systems [15,16], a dedicated and systematic syn- guide for researchers aiming to leverage DL for the targeted design
thesis of DL’s role across the full metal-specific PSPP chain—from of next-generation metallic materials.
automated microstructure analysis, through multimodal property
prediction, to generative inverse design—is notably lacking. Exist- 2. Model paradigms and unique challenges
ing reviews often either encompass all material classes, diluting
the discussion on metallic systems’ unique challenges (e.g., poly- The application of DL to metallic materials is not merely
crystalline anisotropy [17], complex phase transformations [18]), a transfer of generic algorithms, but a domain-specific adapta-
or concentrate on a single subtask. This gap is particularly salient tion driven by the unique characteristics of metallic materials:
as the field enters a phase of exponential growth, moving from multi-scale hierarchical structures (atoms, precipitates, grains, and
proof-of-concept studies toward integrated, physics-aware, and phases), complex non-linear physics (plasticity, phase transforma-
application-ready frameworks. tions, and damage), and frequently data-limited and multimodal
This review aims to fill this gap by providing a focused, criti- experimental landscapes. This section synthesizes the core DL ar-
cal examination of DL applications specifically for structural metal- chitectures that have shown particular promise for metallic mate-
lic materials. We explicitly limit our scope to metallic materials rials, framing them not in abstract terms but through the lens of
to deeply address domain-specific challenges, such as handling the PSPP challenges they address. We distill their functions into a
equiaxed vs. lamellar microstructures, predicting fatigue crack ini- unified summary (see Supplementary materials) and focus here on
tiation in polycrystals, or designing for strength-ductility trade-offs. the specific adaptations and challenges encountered when apply-
Our discussion is structured to mirror the PSPP design logic: (1) ing these models to structural metallic materials.
Microstructure analysis: We explore how DL advances beyond tra-
ditional image analysis to achieve high-throughput, robust quan- 2.1. Bridging the PSPP chain with DL paradigms
tification of grains, boundaries, and phases, with particular atten-
tion to emerging techniques for 2D-to-3D microstructure genera- DL models offer distinct pathways to encode and navigate the
tion and reconstruction. (2) Property prediction: We examine the PSPP relationship for metallic materials:
evolution from single-modal models to multimodal frameworks From structure to descriptors: Convolutional Neural Networks
that fuse images, spectra, and text, emphasizing the integration of (CNNs) act as automated, high-throughput “digital microscopes”,
micromechanical principles and Finite Element Method (FEM) sim- transforming raw OM/SEM/EBSD images into quantitative mi-
ulations to enhance the prediction of mechanical, fatigue, and cor- crostructural fingerprints (grain size, phase fraction, and texture)
rosion properties. (3) Inverse design: We review the shift from di- [19–23]. The primary challenge extends beyond segmentation it-
rect property-composition mapping to generative, closed-loop sys- self, focusing instead on achieving robust generalization across the

92
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

varying contrasts, noise levels, and length scales characteristic of the system. The following sections delve into how these adapted
different imaging techniques and alloy systems. paradigms are concretely applied to the tasks of microstructure
Encoding non-euclidean relationships: Graph Neural Networks analysis (Section 3), property prediction (Section 4), and inverse
(GNNs) provide a native framework for representing metallic design (Section 5).
microstructures as graphs (atoms as nodes in crystal graphs;
grains as nodes in polycrystalline graphs) [24–28]. This is crucial 3. DL in microstructure analysis: from image segmentation to
for capturing topology-dependent properties like fracture tough- quantitative 3D descriptors
ness or intergranular corrosion susceptibility, where the con-
nectivity and attributes of grain boundaries are as important Accurate and efficient quantification of microstructure—the “S”
as the grain interiors—a feature often lost in pixel-based CNN in the PSPP chain—is the foundational step for establishing robust
analyses. process-structure and structure-property linkages. Traditional anal-
Modeling temporal and sequential processes: Recurrent Neu- ysis of micrographs relies on manual thresholding or rule-based
ral Networks (RNNs), Long Short-Term Memory Networks (LSTMs), algorithms, which are subjective, low-throughput, and often fail
and Transformers are essential for incorporating the “Process” in with complex, low-contrast features prevalent in advanced metal-
PSPP. They encode time-series data from heat treatment profiles, lic systems [56]. DL has fundamentally transformed this domain
rolling schedules, or in-situ deformation sequences, learning to by providing automated, generalizable tools that convert raw imag-
predict the ensuing microstructure evolution or property outcome ing data (from OM, SEM, EBSD, STEM, etc.) into quantitative, com-
[29–32]. The challenge is data scarcity for long-term processes and putable structural descriptors. This section moves beyond a mere
integrating physics-based rate laws to ensure predictions are ther- catalog of segmentation techniques to critically examine how DL
modynamically and kinetically plausible. addresses the unique challenges of metallic microstructure analy-
Inverting the PSPP map: Generative architectures—specifically sis (e.g., low-contrast multiphase microstructures and topological
Variational Auto-Encoders (VAEs), Generative Adversarial Networks analysis of grain boundary networks), with a particular emphasis
(GANs), and the emerging Denoising Diffusion Probabilistic Mod- on the pivotal trend of reconstructing 3D microstructural reality
els (DDPMs)—are empowering the inverse design of alloy composi- from 2D data—a critical bridge to physics-based simulation and re-
tions and microstructures from target properties [33–37]. The fron- liable property prediction.
tier here involves moving beyond generating statistically realistic
2D images to synthesizing physically consistent 3D microstructures 3.1. Automated segmentation of microstructural features
and enforcing processability constraints (e.g., suppressing unrealis-
tic phase formations under given cooling conditions). The core task of segmenting atoms, grain boundaries, and
phases has evolved from feature engineering to end-to-end DL
2.2. Challenges in applying DL to metallic materials models, primarily based on CNNs and their encoder-decoder vari-
ants (e.g., U-Net). The key advancement lies in their ability to learn
The effective deployment of these models must contend with hierarchical, context-aware features directly from data, overcoming
several domain-specific hurdles: challenges such as uneven illumination, subtle phase contrast, and
Data bottleneck & multimodal fusion: High-fidelity experimen- noisy backgrounds.
tal data (especially 3D-EBSD, in-situ TEM) is scarce and costly. A Atomic image segmentation: DL has made significant progress
major focus is on developing physics-informed data augmentation, in processing electron microscopy images. The AtomAI framework,
synthetic data generation via simulation (e.g., phase-field method, an open-source Python toolkit, was developed to enable deep
Density Functional Theory (DFT)) [38–43], and multimodal fusion learning-based segmentation of atomic images and convert them
techniques that can jointly reason over EBSD maps, XRD spectra, into local descriptors for subsequent statistical and graph analy-
and chemical compositions to build robust models from limited ex- ses [57]. Building on this foundation, the TEMImageNet training
amples [44–47]. library and AtomSegNet model were further developed for high-
In tegrating Physical and Domain Knowledge: Pure data-driven precision atom segmentation, localization, denoising, and super-
models risk producing “unphysical” predictions. The field is in- resolution processing. These advancements outperformed tradi-
creasingly focused on physics-infused DL, which hardcodes con- tional methods, especially in complex contrast conditions [58].
servation laws, thermodynamic rules, or crystal symmetry into the However, limitations in data augmentation and model training
model architecture or loss functions [48–53]. This improves inter- within the AtomAI framework were identified, and improvements
pretability, generalization to unseen conditions, and reduces the were made by introducing early stopping mechanisms and opti-
required training data. mizing classification thresholds, which significantly enhanced the
From 2D to 3D microstructural realism: Most DL work oper- model’s generalization ability and performance [59,60], as shown
ates on 2D microsections, but properties are governed by 3D mi- in Fig. 2(a). Meanwhile, an unsupervised learning method based
crostructure. A significant trend is using generative models (like on one-class support vector machines (OCSVM) was proposed for
SliceGAN) for 2D-to-3D microstructure expansion [54,55], learning detecting crystallographic defects in atomic-resolution images [61].
to reconstruct statistically equivalent 3D volumes from sparse 2D By using the Patterson function for feature extraction to ensure
slices, thereby bridging a critical gap between characterization and translational invariance, this method successfully detected point,
simulation. line, and planar defects, demonstrating extremely high detection
Computational cost vs. Fidelity trade-off: While DL inference accuracy.
is fast, training advanced models on large 3D datasets (e.g., from Grain boundary segmentation: Accurate grain boundary net-
high-energy X-ray tomography) is computationally intensive. Re- work extraction is vital for predicting properties like intergranu-
search is needed into efficient graph representations of 3D mi- lar corrosion and fracture. DL approaches have progressed from
crostructures and multi-fidelity learning that blends high-cost ac- pixel-classification CNNs to instance-aware segmentation models
curate data with abundant low-cost approximate data. like Mask R-CNN, which can separate touching grains of similar
In summary, the selection and adaptation of a DL model for contrast [62–66]. For example, Saleh et al. [65] delivered a seven-
a metallic materials problem must be a deliberate choice guided stage renderer that outputs photorealistic synthetic TEM images
by the specific PSPP link being targeted, the nature and vol- and a diffusion-based grain-boundary extractor (Fig. 2(b)), attain-
ume of available data, and the fundamental physical constraints of ing 97.23 % accuracy on nanoscale alloys. A significant challenge

93
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 2. (a) Deep neural network model constructed for quantification analysis of ADF-STEM image. (b) Full architecture of the presented model showing the three main
stages and the In/Out data type for each stage. An example from a TEM image was given to show how grain boundaries are usually visible as borders between regions with
different brightness in microscopy images due to the different scattering behavior each crystal orientation has on incoming beams. (c) Schematic outline of the main steps
of the workflow on phase segmentation through SEM images using U-Net with energy dispersive X-ray spectroscopy-guided labeling. (a) Reproduced with permission [60].
Copyright 2021, Wiley. (b) Reproduced with permission [65]. Copyright 2025, Elsevier. (c) Reproduced with permission [65]. Copyright 2025, Elsevier.

in metallic materials is the occlusion of boundaries by secondary metallic materials where exhaustive manual labeling of complex
phases or particles. Advanced frameworks now employ multi-task multi-phase microstructures is prohibitively costly.
learning or generative inpainting to simultaneously segment parti-
cles and restore the obscured boundary network, preserving topo- 3.2. The critical leap: from 2D analysis to 3D microstructural
logical accuracy [67,68]. reconstruction
Phase Segmentation: For metallic alloys with distinct phases
(e.g., α /β in Ti alloys [69], ferrite/martensite in dual-phase steels While 2D analysis is invaluable, most mechanical and physical
[70]), DL models now achieve accuracy rivaling human experts. As properties are governed by the true 3D microstructure. Traditional
highlighted in Refs. [71–73] focused on complex steels, the effec- 3D characterization (e.g., serial sectioning, X-ray tomography) re-
tiveness of these models hinges on high-quality, domain-adapted mains time-intensive and resource-heavy. Here, DL offers a trans-
training data. Recent innovations address the fundamental data formative solution: inferring or generating statistically equivalent
bottleneck: weakly-supervised and semi-supervised learning lever- 3D microstructures from one or a few 2D sections.
ages sparse annotations (e.g., from SEM/EBSD maps [74–76], as This approach, often termed 2D-to-3D microstructure genera-
shown in Fig. 2(c)), while synthetic data generation using phase- tion or dimensionality expansion, has emerged as a major trend.
field or other simulation tools creates unlimited, perfectly labeled Pioneering work by Kench and Cooper [54] demonstrated the use
training samples for data augmentation [77,78]. This is crucial for of a GAN to synthesize realistic 3D microstructures from a single

94
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 3. (a) DREAM.3D, (b) SliceGAN, and (c) integrated SliceGAN-DREAM.3D platform. Comparison of SEVM microstructural characteristics with those from EBSD maps: (d)
Prior particle 2-point correlation S2 function of the sum of 5 equally spaced and normalized slices of each SEVM in each orthogonal plane (XY (top) and YZ (bottom)), and
the corresponding extremum of the central ellipse measured for a two-point correlation value of S2 = 0.25 across each axis; (e) PDFs of the major grain features viz. grain
size, aspect ratio, and misorientation for each SEVM measured as 2D slices in the YZ plane; (f) texture (pole figures). (a–f) Reproduced with permission [81]. Copyright 2024,
Springer Nature.

2D slice, learning the underlying spatial statistics. This method- models have also been companied to learn the probability dis-
ology has been successfully adapted to metallic systems [55,79– tribution of high-dimensional raw data and successfully recon-
82]. For instance, Watanabe et al. [55] performed a compara- struct the 3D microstructures [83,84]. The latest report from Phan
tive study between experimentally observed and GAN-generated et al. [85] pointed out an improvement of up to three times
3D microstructures of dual-phase steels, validating that key mor- in the Frechet Inception Distance (FID) score and enhanced ac-
phological metrics (phase connectivity, size distribution) could curacy in generating 3D images by diffusion-GAN compared to
be accurately reproduced. Furthermore, hybrid approaches have SliceGAN.
been developed to handle complex microstructures. Murgas et al. The implications are profound:
[80,81] introduced an approach, strategically integrating Slice-
GAN for producing bimodal cold spray-formed AA7050 alloy mi- (1) Data augmentation for 3D simulations: Generated 3D volumes
crostructures, with the synthetic microstructure builder Dream provide abundant, synthetic input for computational microme-
3D for packing prior particles with coarse grains having sta- chanics (e.g., Crystal Plasticity FEM, and CPFEM), enabling sta-
tistically equivalent morphological and crystallographic descrip- tistical studies of property variability without exhaustive exper-
tors to EBSD maps (Fig. 3(a–f)). Besides GAN, diffusion-based imental 3D imaging.

95
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

(2) Bridging characterization scales: It helps correlate easier-to- predictive models clearly illustrates a complete evolutionary path:
acquire 2D metallography with more informative but sparse 3D from statistical correlation based on apparent features, to system-
datasets. atic modeling through multimodal information fusion, and further
(3) Inverse design feedstock: Generated 3D structures serve as a di- to an in-depth exploration pursuing mechanistic interpretability.
rect starting point for 3D-aware inverse design pipelines. Initial attempts focused on building regression models using
the most readily available single data types, such as predicting al-
3.3. Toward standardized microstructural descriptors loy hardness directly from chemical composition or strength from
a single process parameter (e.g., annealing temperature) [86,87].
The ultimate goal of microstructure analysis is not just segmen- These models were severely limited by data quantity and qual-
tation, but the extraction of quantitative descriptors that serve as ity. To break this bottleneck, researchers employed strategies on
inputs for downstream property prediction models. DL is pivotal two fronts. One is data augmentation techniques [88–91]; for in-
in automating this descriptor extraction, moving from basic met- stance, Shen et al. [91] used a Wasserstein GAN (WGAN) to gen-
rics (e.g., volume fraction and grain size) to sophisticated, high- erate synthetic data to augment a small-sample dataset of magne-
dimensional microstructural descriptors. sium alloys. They employed an 80/20 train-test split and repeated
These descriptors can be: the training with five-fold cross-validation to ensure robustness.
The other is developing more powerful network architectures [92–
(1) Latent vector embedding: Using the bottleneck layer of an
95]; for example, Yin et al. [95] proposed a novel alloy design
autoencoder compresses a microstructure image into a low-
strategy that integrates a Gated Recurrent Unit (GRU) DL model
dimensional vector that captures its essential, statistically rel-
to predict the yield strengths (YSs) of Cu-Ni-Si alloys. The model
evant features. This latent space enables efficient similarity
was evaluated using 10-fold cross-validation, with an independent
search, clustering, and serves as a compact input for property
hold-out test set (20 % of data) to report final accuracy (Mean
prediction models.
Squared Error Loss (MSELoss) < 0.01), the attention weights func-
(2) Physics-informed descriptors: DL models can be trained to out-
tion analogously to gradient-based interpretability methods like
put descriptors aligned with physical intuition, such as contin-
Gradient-weighted Class Activation Mapping (Grad-CAM), enabling
uous size distributions, shape anisotropy, or local texture gra-
the model to automatically visualize and focus on key contributing
dients, which bridge the gap between “black-box” features and
factors. This enhances interpretability by providing insight into the
meaningful physical variables, offering higher interpretability.
model’s decision-making process.
In summary, DL has revolutionized metallic microstructure Recognizing the one-sidedness of single data sources, multi-
analysis by automating the transformation of images into quanti- modal data fusion has become key to enhancing prediction ro-
tative, multi-scale descriptors. The field is now moving beyond 2D bustness. This requires models to collaboratively process differ-
quantification toward predictive 3D reconstruction, thereby creat- ent types and scales of information [96–102]. A typical case is
ing a crucial digital twin of the material’s structure. This enriched, the work by Ren et al. [98], who predicted the ultimate strengths
3D representation of “Structure” forms a far more robust foun- (UTSs) and elongation-to-failures (ELs) of dual-phase steels fusing
dation for establishing the “Structure-Property” relationships dis- three Mn grades with a co-registered EBSD triad—band-contrast
cussed in the next section, and for enabling the generative design map, phase map, and KAM map; the triple-image plus chemical
of microstructures with targeted characteristics. compositions stack yielded high coefficient of determination values
(R2 = 0.962 for UTS and R2 = 0.953 for EL). Additionally, Han et al.
4. DL in property prediction: the intelligent evolution from [99] constructed a CP-guided DL model (CP-CNN) through the mul-
correlative learning to mechanism fusion timodal fusion of image data and numerical data (e.g., composition,
processing parameters, and annealing temperature). Using the dot
Property prediction serves as the core relation connecting “S” product method, normalized numerical vectors are fused with the
and “P” within the PSPP chain. For structural metallic materials, pixel matrix of stress nephograms to form a multi-channel input.
their key properties—such as mechanical strength, fatigue life, and This enables the model to simultaneously learn the material’s mi-
corrosion resistance—are fundamentally the macroscopic response croscopic mechanical fields and macroscopic process-composition
of the microstructure to external fields, governed by multi-scale, features, achieving accurate and efficient generalized predictions
nonlinear physicochemical processes. DL, by deciphering high- of the mechanical properties of dual-phase steels across different
dimensional and complex “Structure-Property” mappings, provides systems without the need for repeated parameter fitting (Fig. 4(a–
a novel pathway to overcome the limitations of traditional empir- c)). More complex information fusion is exemplified by research
ical models and the bottlenecks of high-cost physical simulations. like that of Furushima et al. [100], Cao et al. [101], and Li et al.
This section first elaborates on the predictive advances and appli- [102], which combined XRD patterns, Continuous Cooling Transfor-
cation examples of DL in the three core areas: mechanical prop- mation (CCT) diagrams, and periodic-table image (PTR) with SEM
erties, fatigue properties, and corrosion properties, demonstrat- micrographs, enabling the model to understand the influence of
ing its evolution from single-data correlation to multimodal fu- the phase transformation process on final mechanical properties.
sion. Building upon this, we systematically analyze a fundamen- These works suggest that fusing composition, images, and even
tal paradigm shift: predictive models are evolving from purely spectra curves can construct a unique, comprehensive “digital fin-
data-driven systems into “mechanism-enhanced hybrid intelligent gerprint” of the material state, which is the necessary path toward
agents” through deep integration with physical simulations (e.g., high-precision prediction.
FEM, CPFEM, phase-field method) and multi-scale modeling. This Cutting-edge research is no longer satisfied with high-accuracy
transformation is key to enhancing prediction reliability, inter- “black-box” prediction; instead, it pursues deep mechanism fusion
pretability, and extrapolation capability, laying a solid foundation and model interpretability, actively embedding physical knowledge
for closed-loop inverse design. into models and explaining predictions [103–108]. For example,
Salam et al. [103] reported that the Transformer-Based Physics-
4.1. Mechanical property prediction Informed Neural Network (T-PINN) integrated physical constraints
such as phase transition laws in metallurgy into the loss func-
Mechanical properties are the most fundamental evaluation tion, achieving high accuracy in predicting the hardness of high-
metrics for structural metallic materials. The development of their entropy alloys and quantifying the contribution of key features

96
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 4. DL framework guided by the CP theory: (a) the dataset generated by CP simulation; (b) integration of multi-modal information regarding material composition,
processing, and annealing temperature; (c) the framework of the CP-CNN model. (a–c) Reproduced with permission [99]. Copyright 2025, Elsevier.

through SHapley Additive exPlanations (SHAP) analysis. Song et al. Directly predicting fatigue life from micrographs is the most
[104] constructed a cross-modal knowledge graph, numerical rea- intuitive application of DL. Kishore et al. [109], targeting addi-
soning knowledge graph (NR-KG) shown in Fig. 5, bridging the tively manufactured Ti-6Al-4V, extracted statistical descriptors like
gap between numerical and semantic data via projection predic- the two-point autocorrelation function from SEM images and com-
tion loss and contrastive learning loss, and achieved significant bined them with a Gaussian Process Regression model to achieve
performance improvement in high-entropy alloy hardness predic- low-error fatigue life prediction (Mean Absolute Percentage Error,
tion. These studies indicate that the core trend of modern intelli- MAPE = 6.20 %). Their validation protocol included a 70/30 train-
gent material prediction lies in: constraining model structures with test split, with 10 % of the training data reserved for hyperpa-
physical mechanisms, decoding prediction logic with knowledge rameter tuning. Zhou et al. [110] went a step further by analyz-
graphs, breaking data bottlenecks through cross-modal fusion, and ing in-situ high-temperature fatigue experiment videos of a Ni-
realizing intelligence encompassing “high-precision prediction and based single-crystal superalloy. They used YOLOv8 to automatically
interpretable mechanisms”. identify and quantify crack propagation, then employed an LSTM
to predict its future damage state, achieving dynamic, visualizable
early warning of the fatigue failure process. Managing the com-
4.2. Fatigue property prediction plexity of such sequential models (like LSTMs) on long video data
often requires careful architecture design and dataset curation.
Fatigue failure is the primary mode of destruction for metal Fatigue crack initiation strongly depends on topological infor-
components under cyclic loading. Its prediction is extremely chal- mation such as grain boundary characteristics and misorientation
lenging due to the complex processes of crack initiation and prop- between neighboring grains. GNNs, naturally suited for handling
agation. DL is driving this field away from reliance on macroscopic relational data, have become a breakthrough tool in this field.
empirical formulas and toward a new paradigm of prediction that Thomas et al. [111] transformed EBSD data of ferritic steel into a
directly correlates with microstructural topology and mesoscopic grain graph (nodes = grain features, edges = grain boundary fea-
damage mechanisms, with a growing emphasis on interpretable tures) and used a GNN to successfully predict “high-risk grains”
and computationally efficient models. prone to fatigue crack initiation (Fig. 6(a)). Interpretability anal-

97
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 5. The NR-KG framework includes KG construction and the numerical reasoning process. First, cross-modal KG ςm was constructed with numerical and semantic data.
Then, proxy nodes represented by material numerical features vfm |em , semantic nodes ei , and edges r in ςm were projected into a canonical KG ςc using CMP. ςc was
supervised by CLL and PPL. GNN Regression Prediction facilitated information exchange and predicted material properties v . , Num-Proj. Layer, Sem-Dict, Margin Func, and
fermat() referred to the join operation, Numerical Projection Layer, Semantic Dictionary, Margin function, and high-dimensional generalized F-point calculation. Reproduced
with permission [104]. Copyright 2025, Elsevier.

ysis of this GNN, potentially using graph-specific feature attribu- loy [131,132,142], titanium alloy [132,143], nickel-based superalloy
tion methods, indicated that grains with high Young’s modulus and [132,143], structural steel [130,133], and Additive Manufacturing
high Schmid factor were at higher risk (Fig. 6(b)), which aligned (AM)-specific materials [129,132,142]. The fatigue types cover full
perfectly with classical fatigue mechanisms. The benchmark study scenarios such as uniaxial/multiaxial fatigue [128,143], high-low
by Dai et al. [112] confirmed that for large 3D microstructures con- cycle combined fatigue [133], corrosion fatigue [131,141], multi-
taining thousands of grains, the computational efficiency of GNNs defect fatigue [130,142], low cycle-creep fatigue [133,142], multi-
far surpasses that of voxel-based CNNs, clearing the obstacle for scale crack fatigue [143], and probabilistic fatigue [129,142]. Model
high-throughput fatigue property screening and highlighting a fa- innovation has formed three main routes: (1) Deeply integrat-
vorable complexity-to-accuracy trade-off. ing physical constraints such as critical plane models and frac-
The high cost of high-quality fatigue experiments leads to data ture mechanics theories to ensure physical consistency; (2) Opti-
scarcity. Cross-domain transfer learning provides an effective solu- mizing the network architecture, integrate temporal/image feature
tion to this small-sample problem and reduces the need to train extraction, enhancing small-sample generalization ability through
overly complex models from scratch [113–122]. Chen et al. [118] meta-learning and transfer learning, and addressing data scarcity
constructed a “tensile-to-fatigue” transfer framework via two steps, with multi-fidelity frameworks; (3) Upgrading optimization strate-
including pre-straining and then fine-tuning, significantly improv- gies, adopting genetic algorithms and Bayesian optimization to ad-
ing prediction accuracy. Zhang et al. [119] proposed a discrimina- just physical loss weights. PINN shows outstanding advantages in
tive multi-source transfer strategy that dynamically evaluated and physical interpretability, small-sample adaptability, and engineer-
fused knowledge from different alloy systems (e.g., carbon steel, ing scalability.
stainless steel), maintaining excellent cross-material prediction ac-
curacy even with minimal target data (Fig. 6(c, d)). 4.3. Corrosion property prediction
The recent works aim to deeply integrate data-driven mod-
els with mechanistic simulation tools [123–127]. Tang et al. [127] Metal corrosion is a dynamic process involving coupled elec-
generated a large number of synthetic 3D microstructures via trochemistry, mass transfer, and mechanics. DL is pushing the re-
Voronoi tessellation, used CPFEM/Cellular Automaton (CA) to cal- search paradigm in this field from qualitative assessment of static
culate their Fatigue Indicator Parameters (FIPs), and then trained morphology toward quantitative early warning of dynamic pro-
a 3D CNN model to establish an interpretable predictive chain of cesses and cross-scale mechanistic analysis, with recent studies in-
“microstructure-fatigue damage driving force-life” (Fig. 7(a)). This creasingly incorporating interpretability to uncover critical factors.
hybrid paradigm of “simulation-generated data + DL-built surro- Early research focused on using the initial microstructure be-
gate model” not only addresses data scarcity but also grounds pre- fore corrosion onset (e.g., phase composition [145–147], grain size
dictions on clear physical mechanisms. On the other hand, PINNs [148,149], crystal orientation [150–152], high-density twins [153])
have become the core technology for predicting material fatigue to predict material corrosion tendency. DL models can classify or
performance [128–144]. Through the fusion paradigm of “data- regress pitting susceptibility or uniform corrosion rate. Post-hoc
driven + physical constraints”, they break through the limitations interpretability techniques like SHAP can be applied to such mod-
of traditional models and fully cover multi-material, multi-fatigue els to identify which microstructural features (e.g., specific phase
type, and complex engineering scenarios (Fig. 7(b) [132]). The ma- boundaries, inclusion types) are most decisive for corrosion initia-
terial system includes stainless steel [128,130,133], aluminum al- tion, moving beyond pure prediction to mechanistic insight.

98
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 6. (a) Example protrusion instances and the conversion of a grain neighborhood into graph and tabular representations. (b) Importance of a pair of features (the
maximum generalized Schmid factor in the grain and Young’s modulus of the grain in the axial direction to affect fatigue damage prediction). (c) Multi-source transfer
learning framework for cross-material fatigue life prediction. (d) Comparison of cross-material generalization performance across ten random initializations. (a, b) Reproduced
with permission [111]. Copyright 2023, Springer Nature. (c, d) Reproduced with permission [119]. Copyright 2025, Elsevier.

More advanced models are dedicated to capturing and predict- cessfully predict the subsequent propagation trend and severity of
ing the dynamic evolution of corrosion [154–158]. The work by localized corrosion, with accuracy far exceeding traditional linear
Sun et al. [156] is exemplary: they combined in-situ Electrochemi- regression methods, providing a powerful tool for real-time corro-
cal Atomic Force Microscopy (EC-AFM) with a Convolutional LSTM sion monitoring and early intervention.
to continuously image and model the surface morphology changes In corrosion prediction tasks, small sample sizes have always
of advanced high-strength steel in a corrosive solution (Fig. 8(a– been a core challenge restricting model performance and general-
c)). Using only the early-stage morphological evolution sequence, ization ability. This issue is further exacerbated when hybrid mod-
these models (like GRU [157], YOLOv5-Shape-IoU [158]) could suc- els incorporate DL components such as CNNs and LSTMs, as these

99
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 7. (a) Schematic representation of the integrated multiscale modeling and deep learning framework employed for fatigue strength prediction. (b) Framework of proposed
PINN for low-cycle fatigue life prediction of copper alloy. (a) Reproduced with permission [127]. Copyright 2026, Elsevier. (b) Reproduced with permission [132]. Copyright
2026, Elsevier.

models demand massive volumes of training data and powerful In addition, a series of optimization strategies at the data and
computational resources, which amplifies the problems caused by algorithm levels can also help tackle the small sample challenge.
data scarcity [159,160]. Compounded by issues like bias and inad- For instance, design algorithms that are insensitive to outliers and
equate prediction reliability in corrosion data collected from real shifts in data distribution (e.g., Bayesian Ridge Regression) [162],
marine environments, the scalability of such models faces severe generate artificial data to achieve balanced processing of datasets
tests. To address this pain point, transfer learning has emerged [163], conduct refined data cleaning [164], and perform data aug-
as an effective solution. This approach leverages models, which mentation on image-based corrosion datasets [165]. These strate-
are pre-trained on large-scale general-purpose datasets like Ima- gies remedy the deficiencies of small sample data from various di-
geNet [161], to transfer the general knowledge they have learned to mensions, providing crucial support for enhancing the stability and
corrosion prediction scenarios. These pre-trained models are then reliability of corrosion prediction models.
fine-tuned on smaller, scenario-specific corrosion datasets. This not Current research frontiers are committed to integrating physics
only effectively alleviates the problems of scarce or biased corro- principles to achieve interpretable prediction. One direction is
sion data but also accelerates the inference process for corrosion the deep integration of DL with corrosion mechanistic simula-
test data. tion, such as FEM [166], phase-field method [167], and digital

100
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 8. Schematic of the workflow for local corrosion behavior using deep learning: (a) in-situ EC-AFM, where the local corrosion behavior was measured within the exposed
region inside the O-ring; (b) image processing involved drift correction and height calibration to get the corrosion dataset; (c) extraction of training samples from the
corrosion dataset to train the ConvLSTM model to predict pixel-wise corrosion behavior, which was later constructed as an image of the time step t. (a–c) Reproduced with
permission [156]. Copyright 2024, Elsevier.

twin technology [168]. For example, Ma et al. [167] combined The evolution of corrosion property prediction clearly reflects
phase-field method-simulated corrosion interface evolution data a deepening from “post-facto observation” to “process early warn-
with experimental macroscopic parameters to train a CNN model, ing”, and further to “mechanistic design”. Through deep integra-
achieving rapid and accurate prediction of the long-term failure tion with dynamic modeling, physical simulation, and composition
behavior of protective coatings, with computational efficiency design—and augmented by interpretability methods that explain
improved by a factor of one million compared to pure phase-field model decisions—DL not only predicts “whether and how corrosion
simulation (Fig. 9(a)). Another more disruptive direction is devel- will occur” but also begins to answer “how to design for corrosion
oping a high-precision, high-efficiency, and physically consistent resistance”, significantly enhancing the scientific depth and engi-
surrogate model to achieve rapid prediction of corrosion behavior neering value of this research direction.
[169–171]. For instance, Venkatraman et al. [171] utilized LSTM
and Stochastic Variational Inference (SVI) to capture the temporal 4.4. Toward a mechanism-enhanced hybrid intelligent predictive
evolution of charges, incorporates physical constraints such as framework
charge neutrality, and was supplemented by active learning to
optimize data in key regions, ultimately achieving rapid prediction The progress in mechanical, fatigue, and corrosion prediction
of partial charges related to corrosion—with a speed increased outlined above reveals a common evolutionary path and a cutting-
by two orders of magnitude compared to Molecular Dynamics edge paradigm: DL is deeply integrating with physical simulation
(MD) simulations (Fig. 9(b)). Additionally, language models have and multi-scale modeling to form a “mechanism-enhanced hybrid
been applied to predict corrosion behaviors [172–174]. Busch et al. intelligent predictive framework”. This paradigm aims to combine
[174] demonstrated the significant potential of Large Language data-driven learning capability, the constraining power of physical
Models (LLMs) in predicting the corrosion inhibition efficiency of laws, and the predictive power of cross-scale correlations, primar-
small organic molecules for the ZE41 alloy (Fig. 9(c)). Compared to ily manifested in three core fusion modes:
traditional machine learning methods (like Multilayer Perceptrons (1) Simulation data-driven surrogate models: Utilizing high-fidelity
(MLPs)) and specialized chemical LLMs, next-generation LLMs like simulation tools like CPFEM, phase-field method, and macro-
GPT-4o can leverage pre-trained knowledge to make accurate pre- scopic FEM to generate “virtual experimental” data spanning
dictions using only molecular names and SMILES strings (Fig. 9(d)), atomic, micro-, meso-, and even macro-scales, which is then
without the need for complex feature engineering. Furthermore, used to train lightweight, high-efficiency DL surrogate mod-
well-designed prompting strategies and the reasoning capabilities els [123–127,166–168]. This fundamentally breaks the bottle-
of LLMs are the key factors enabling their superior performance necks of cost, throughput, and scale associated with experimen-
over other methods on small datasets. tal data.

101
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 9. (a) Prediction performance with corrosion time for the in vitro pitting corrosion behavior of coated magnesium alloys via phase-field method assisted CNN-based
DL framework. (b) Workflow used to estimate and refine physics-aware predictions of partial charge. (c) Overview of the prompting strategies. With some exceptions, the
prompting works like a chat starting at the top, where previous prompts and answers were also appended before the current prompt as chat history. This graphic showed
the prompting schemes for the approaches without descriptors. When the descriptors were added to the input data, the pre-analysis prompt was expanded to a set of
analyses with different focuses. (d) Overview of the best results from each group using different evaluation metrics. (a) Reproduced with permission [167]. Copyright 2025,
Elsevier. (b) Reproduced with permission [171]. Copyright 2025, Springer Nature. (c, d) Reproduced with permission [174]. Copyright 2025, Elsevier.

(2) Physics-constrained neural network architectures: Encoding 5. DL in inverse design: multi-objective-driven optimal
physical laws derived from different scales (e.g., electronic composition search and microstructure generation
structure information at the atomic level, constitutive relations
at the meso-scale, conservation laws at the macro-scale) as soft Inverse design represents the ultimate realization of the
or hard constraints into the neural network architecture design, “materials-by-design” vision. Its core lies in reversing the causal
loss function, or training process [103–108,128–144,169–171]. relationship within the PSPP chain. It inversely solves for the
This ensures the model’s predictions are physically consistent material composition, microstructure, and even process routes
across scales and significantly enhances their extrapolation ca- that meet the requirements. Traditional “trial-and-error” meth-
pability under unknown conditions. ods or rule-based searches are inefficient in the vast, high-
(3) Multi-scale modeling information fusion: It focuses on con- dimensional materials space. This section aims to demonstrate
structing and utilizing cross-scale correlations themselves. DL that DL-driven inverse design is evolving from a “property-to-
models are used to learn and establish efficient mapping re- composition/process/structure” mapping tool into a loop decision-
lationships between models at different scales, achieving loss- making system that simultaneously considers process feasibility,
less or high-fidelity transfer of information between scales. multi-objective trade-offs, and physical consistency.
For example, using DL models rapidly predict the mesoscale
initial microstructure from macroscopic process parameters
5.1. From composition inverse mapping to full-chain PSPP
[154–158], or upscale the statistical information of local re-
optimization
sponses from mesoscopic data to macroscopic properties [96–
101,127,166–168], thereby replacing traditional empirical ho-
Generative models and inverse design strategies: Generative
mogenization methods and enabling tight coupling and accel-
models serve as the core tool for achieving inverse design in al-
eration of multi-scale simulations.
loys, capable of directly generating candidate alloys that meet tar-
This “mechanism and multi-scale fusion” paradigm fundamen- get properties from high-dimensional composition space, thereby
tally elevates the depth and breadth of property prediction. It overcoming the limitations of traditional forward prediction. In
transforms predictive models from mere data-fitting “black boxes” wire-arc additive manufacturing steel inverse design [175], the
into “mechanism-enhanced partners” capable of fusing first princi- DDPM framework can serve as a powerful generative design tool
ples, traversing key scales, revealing the intrinsic cross-scale causal to synthesize reasonable alloy compositions with target prop-
chain of “Structure-Property,” and performing reliable extrapola- erties (martensite start temperature, Ms ) under strict feedstock
tion in data-scarce regions. constraints. For the interpretable design of high-entropy alloys

102
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 10. Flowchart of the generative machine learning framework for the design of eutectic compositionally complex alloys: (a) The structure of the initial database, the data
descriptors or features, and the data label; (b) the pre-processing of the data; (c) the architecture of the artificial neural network; (d) the architecture of the variational
autoencoder; (e) the distribution of the eutectic forming likelihood (EFL) values of the generated compositions; (f) the experimental validation of a selected composition.
(a–f) Reproduced with permission [177]. Copyright 2024, Springer Nature.

[176,177], Chen et al. [177] deployed a Conditional VAE (CVAE) for attention-weight visualization function clarifies the model’s prior-
generative P2C mapping in eutectic alloys (Fig. 10(a–f)). After in- ity focus on physical properties like atomic weight and melting
tegrating thermodynamic screens (eutectic-point constraint), the point, enhancing the physical reasonableness of predictions. Duct-
model proposed a latent space of Al-Si-Cu compositions with a GPT [182], based on the GPT-2 Transformer decoders, integrated
melting point below 577°C and exceeding 15 % ductility. The se- physical descriptors like Fermi-level density of states and elas-
lected alloy yielded beyond 90 % eutectic fraction experimentally, tic constants to rapidly predict the ductility of refractory MPEAs
extending inverse design from deterministic prediction to genera- through text interaction. This model successfully screened two
tive exploration. In the aluminum alloy design [178], the ACGANs- novel refractory alloy systems, NbTa-Ti-V and W-Ti-V, demonstrat-
SA framework was proposed, combining the category controllabil- ing that ductility improves from below 6 % to exceeding 18 %
ity of auxiliary classifier generative adversarial networks (ACGANs) when Ti content exceeds 50 at.%. The core advantage of LLMs lies
with the global optimization capability of simulated annealing in breaking down barriers between specialized software and non-
(SA). This framework was used to design a novel aluminum alloy expert users, enabling property prediction and composition screen-
grade, achieving directional optimization for either density reduc- ing through natural language interaction. Simultaneously, LLMs can
tion or strength enhancement. For hydrogen storage alloys [179], integrate multi-source heterogeneous data (such as literature text,
a generative model based on VAE was constructed. The encoder experimental reports, and spectral data), providing a new path for
learned the latent distribution of compositions, and the decoder cross-platform data fusion.
generated binary to decenary hydrogen storage Multi-Principal El- Physics-data fusion modeling: It has become a key direction
ement Alloys (MPEAs). By introducing a CVAE that used hydride for enhancing model generalizability and reliability. Research em-
formation enthalpy as a condition variable, the model could direc- beds thermodynamic principles [177,185,186], atomic-scale fea-
tionally generate alloy compositions targeting specific thermody- tures [187], and physical descriptors [183,188] to construct model
namic parameters. architectures that balance prediction accuracy with physical in-
LLMs-empowering alloy design: Leveraging their powerful text terpretability. For instance, a physics-informed transfer learning
interaction and feature encoding capabilities, LLMs have lowered framework was developed [185]. It trained neural networks us-
the technical barrier for alloy design, enabling a convenient “natu- ing low-fidelity empirical force field data to capture the essence
ral language description to property prediction” design paradigm. of plastic deformation and embeds high-fidelity parameters from
Two typical LLM application frameworks show great promise: first-principles calculations. This enabled chemically accurate pre-
one is a property prediction model based on Transformer en- diction of the Peierls stress in metal alloys, allowing for the screen-
coders [180,181], and another is a generative platform support- ing of high-strength metastable alloys with only single-point high-
ing interaction and screening [182–184]. AlloyBERT [180], based fidelity calculations. For designing B2-phase compounds [188], a
on RoBERTa/BERT encoders, trains an alloy-specific tokenizer us- random sublattice-based descriptor was proposed, integrating 18
ing Byte Pair Encoding (BPE). It inputs alloy composition, process physical parameters, such as atomic size mismatch, mixing en-
parameters, and elemental physical properties in text form to pre- thalpy, and electronegativity. This descriptor more accurately dis-
dict key properties such as elastic modulus and yield strength. Its tinguished B2-phase from non-B2-phase structures compared to

103
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

traditional random mixing descriptors. Embedding physical infor- tion in the performance-driven inverse design of dual-phase steel
mation not only enhances model generalization in data-scarce sce- microstructures.
narios but also ensures predictions align more closely with fun- In summary, DL-driven inverse design is charting a clear evolu-
damental principles of materials science, avoiding the generation tionary path: from finding a single optimal composition point, to
of chemically infeasible candidate compositions and providing reli- achieving multi-objective, constrained global optimization in the
able assurance for experimental validation. high-dimensional “Process-Structure-Property” space; from gener-
ating seemingly plausible 2D images, to synthesizing physically
5.2. From statistically realistic to physically realizable microstructure consistent 3D microstructures directly usable for simulation. The
generation destination of this path is a closed-loop system tightly integrated
with high-throughput computation and automated experimenta-
Hybrid generative models: Single generative models (e.g., GAN tion. This system can not only “design the material” but also “de-
[189], VAE [190,191], and DDPM [192]) struggle to simultane- sign the process route to manufacture it”, and preliminarily fore-
ously meet the dual demands of “reconstruction quality” and “op- see its service performance [200], thereby truly achieving digital
timization efficiency”, making hybrid architectures the dominant integration and intelligent decision-making across the entire PSPP
direction in current research. The VAE-CDGM model innovatively chain.
adopted a two-stage training strategy [193]: first, a VAE reduced
the dimensionality of microstructures and generated a blurred re-
construction map; then, using this blurred map as a condition, a 6. Prospects: toward closed-loop materials intelligence
DDPM was trained for detail refinement. This successfully resolved integrating physics, data, and autonomous experimentation
the conflict between the blurriness of traditional VAEs and the
inefficiency of DDPM optimization. GANs enhance microstructure Although DL has demonstrated transformative potential in
generation accuracy through adversarial training between the gen- various stages of metallic materials development—including mi-
erator and discriminator, but their latent space often lacks inter- crostructure analysis, property prediction, and inverse design—
pretability [54,55,189]. Zhang et al. [194] proposed a DA-VEGAN translating these isolated successes into an engineering-ready,
model that introduced a β -VAE to construct an interpretable la- high-reliability, and complete PSPP solution still faces systemic
tent space (with the β parameter balancing reconstruction qual- bottlenecks. These bottlenecks include insufficient model inter-
ity and latent space smoothness) and integrated differentiable data pretability and physical consistency, weak generalization under
augmentation techniques. This enabled high-fidelity reconstruction small-sample and cross-domain conditions, difficulties in multi-
with a minimal dataset of only 16 samples, achieving a reconstruc- modal and cross-scale data fusion, and the gap between “vir-
tion error of ≤ 1 %, making it suitable for the microstructure design tual design” and “experimental validation”. Future breakthroughs
of 2D dual-phase materials. Building upon the Deep Convolutional will no longer be confined to improvements in individual algo-
GAN framework, the CCDCGAN model incorporates formation en- rithms but will focus on constructing an intelligent system featur-
ergy constraints or backpropagation optimization [195]. It has been ing closed-loop interaction among data, physics, and experimenta-
successfully applied to the inverse design of crystal structures in tion. This section proposes a multi-level roadmap, aiming to pro-
the Bi-Se binary system, reproducing 11 experimentally accessi- pel the field from the application of discrete tools toward a new
ble phases and supporting extension to multi-component systems, paradigm of system-level materials research and development.
providing an effective tool for the design of stable phases in inor-
ganic crystalline materials. In addition, Attari et al. [196] showed
a microstructure-centered materials design workflow using gener- 6.1. Roadmap level 1: building a foundation of physics-informed
ative phase-field modeling and deep VAE, where high-throughput perception and cognition
phase-field modeling was used to generate a large dataset of syn-
thetic microstructures and VAE to generate new microstructures (1) Developing physics-constrained model architectures: Incorpo-
with desired thermal conductivity. rating key governing equations or variational principles from
Bidirectional prediction-inverse design frameworks: Bidirec- phase-field simulations and CP theory as regularization terms
tional frameworks significantly improve design efficiency by us- into the loss function can ensure model predictions (e.g., gen-
ing a forward model for rapid property prediction from mi- erated microstructures, predicted stress fields) adhere to basic
crostructures and a reverse model for mapping target properties physical laws.
back to microstructures. For the design of impact response in (2) Establishing universal strategies for small-sample and cross-
metastable high-entropy alloys [197], a CNN served as the for- domain transfer: It is necessary to develop a “pre-training and
ward model, rapidly predicting stress-strain curves and marten- fine-tuning” paradigm for foundational models. Pre-training on
site volume fraction. A GAN served as the reverse model, in- large-scale, cross-alloy, cross-property databases allows models
ferring microscopic texture and temperature parameters based to learn universal material representations, which can then be
on target responses, enabling precise control over impact perfor- adapted to new systems with minimal data. Concurrently, gen-
mance. The 3D inverse design pipeline proposed by Shang et al. erative models (e.g., diffusion models) can be used to synthe-
[198], which fused DL with a genetic algorithm for Ti-6Al-4V, size high-quality virtual data guided by physical rules, com-
could not only generate 3D dual-phase structures optimized for bined with active learning strategies to intelligently guide key
different application scenarios (e.g., high strength for aerospace, experiments to fill knowledge gaps.
low modulus for biomedical use) but also predicted their macro- (3) Achieving unified multimodal and cross-scale representation:
scopic stress-strain response via a 3D CNN and obtain local stress Developing unified encoders capable of mapping multi-source,
fields via a 3D conditional GAN (Fig. 11(a–d)). Kusampudi and heterogeneous information—such as composition, 2D/3D im-
Diehl [199] pioneered an innovative generative model framework ages, spectral data, and textual reports—into a shared semantic
based on VAE and Bayesian optimization. This framework auto- space. This will provide the foundation for cross-scale correl-
matically extracted low-dimensional descriptors of microstructures ative analysis from atomic arrangement to macroscopic prop-
and established microstructure-property relationships (specifi- erties and enable the quantification of prediction uncertainty
cally for yield stress and susceptibility to ductile damage), arising from multi-source information fusion using Bayesian
with experimental validation confirming its successful applica- methods.

104
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

Fig. 11. Inverse exploration results: (a) Illustrative genetic algorithm (GA) optimization loop. (b) Convergence plot for two optimization objectives OPTaero and OPTmed . (c)
Evolution of GA solutions over 40 optimization generations visualized in a reduced dimension space. Solutions from the first and last ten generations are circled and labelled,
and the trajectory of solution evolution is marked by a black arrow. (d) Comparison of σy , E, and Kt between inverse exploration results and the original FEA dataset. The
optimal microstructures and the stress fields for corresponding optimization tasks are visualized. The X, Y, and Z axis are elastic modulus (E), yield strength (σy ), and stress
concentration factor (Kt ), respectively. (a–d) Reproduced with permission [198]. Copyright 2023, Elsevier.

6.2. Roadmap level 2: achieving closed-loop optimization across the tomated characterization (e.g., robotic electron microscopes),
entire PSPP Chain and rapid property testing platforms to form a “AI Design -
Automated Experimentation - Data Feedback - Model Update”
(1) Strengthening property-oriented inverse design: Directly incor- closed loop. This not only enables rapid validation of AI predic-
porating target properties as designable variables into inverse tions but also allows for continuous iterative model optimiza-
generative models. Through real-time interaction with ther- tion using experimental data, fostering self-evolving R&D capa-
modynamic (CALPHAD) and kinetic simulations, ensure that bility.
inversely recommended compositions, processing, and struc-
tures are thermodynamically feasible and kinetically attainable, 6.3. Roadmap level 3: incubating autonomous agents for materials
thereby directly outputting manufacturable “recipes”. discovery and innovation
(2) Developing multi-objective optimization frameworks for service
performance: Property prediction models need to expand from (1) Creating platforms for 3d dynamic microstructure generation
single static properties to encompass time-varying and coupled and simulation: Integrating high-resolution in-situ 3D char-
service performance indicators like fatigue, creep, corrosion, acterization techniques with 3D generative networks to cre-
and wear. The inverse design system must then integrate multi- ate dynamic microstructure digital twins that reflect tempo-
objective optimization algorithms to find Pareto-optimal solu- ral evolution (e.g., recrystallization, phase transformation) and
tion sets under multiple constraints, providing decision-makers load-response history. This will provide unprecedented insight
with a trade-off space. into manufacturing processes and service failure, enabling inte-
(3) Constructing computational-experimental closed-loop valida- grated inverse design of “manufacturing process-final structure-
tion platforms: Connecting the DL design system with high- service performance” for complex processes like additive man-
throughput synthesis (e.g., combinatorial material chips), au- ufacturing.

105
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

(2) Exploring autonomous exploration of new material spaces: An Data curation, Conceptualization. Yanfu Chai: Writing – review &
intelligent agent integrating all the above capabilities could, editing, Funding acquisition, Formal analysis, Data curation. Dabiao
driven by high-level human-defined goals, autonomously plan Xia: Funding acquisition, Formal analysis, Data curation. Bin Jiang:
experiments, interpret results, formulate hypotheses, and initi- Writing – review & editing, Formal analysis. Hyoung Seop Kim:
ate new rounds of exploration. This will ultimately free human Writing – review & editing, Methodology, Investigation, Formal
scientists from repetitive trial-and-error, allowing them to focus analysis, Data curation.
on higher-level scientific questions and innovative concepts.
Acknowledgments
7. Conclusions
This research was funded by the National Natural Science Foun-
This review systematically summarizes the advances in apply-
dation of China (52204407 and 52304398), the Natural Science
ing deep learning to the research and development of metallic
Foundation of Jiangsu Province (BK20220595), and the National Re-
materials. Following the inherent logic of PSPP, it elaborates on
search Foundation of Korea (NRF-2022R1A5A1030054).
the paradigm shifts triggered by DL across three key stages: mi-
crostructure analysis, property prediction, and inverse design. The
main conclusions are as follows: Supplementary materials

(1) In microstructure analysis, DL has achieved the leap from man- Supplementary material associated with this article can be
ual thresholding to automated, high-throughput quantitative found, in the online version, at doi:10.1016/[Link].2026.02.007.
analysis. It is now building 3D digital twins connecting char-
acterization and simulation through 2D-to-3D generation tech- References
niques, laying a solid foundation for accurate “structure” de-
scription. [1] G.B. Olson, Science 277 (1997) 1237–1242.
(2) In property prediction, models have evolved from regression [2] G.B. Olson, Science 288 (20 0 0) 993–998.
[3] W. Xiong, G.B. Olson, NPJ Comput. Mater. 2 (2016) 15009.
based on single features to hybrid intelligent frameworks that [4] T. Onishi, T. Kadohira, I. Watanabe, Sci. Technol. Adv. Mater. 19 (2018)
fuse multimodal data, embed physical mechanisms, and couple 649–659.
with multi-scale simulations. This not only improves prediction [5] A. Agrawal, A. Choudhary, Int. J. Fatigue 113 (2018) 389–400.
[6] X. Jiang, D. Xue, Y. Bai, W.Y. Wang, J. Liu, M. Yang, Y. Su, Rev. Mater. Res. 1
accuracy but also endows the models with interpretability and
(2025) 10 0 010.
the ability to reveal complex physical mechanisms, solidifying [7] Y. Lan, G.S. Zheng, R.W. Song, J.N. Hao, J.L. Liu, C.L. Shen, J.Y. Zhu, S. Cao,
the core correlations within the PSPP chain. J.L. Zhao, Q. Lou, C.X. Shan, Nat. Commun. 16 (2025) 8167.
(3) In inverse design, the research focus has shifted from finding [8] Y.L. Wu, C.F. Wang, M.G. Ju, Q.Q. Jia, Q.H. Zhou, S.H. Lu, X.Y. Gao, Y. Zhang,
J.L. Wang, Nat. Commun. 15 (2024) 138.
an optimal composition point to achieving multi-objective, con- [9] G.L.W. Hart, T. Mueller, C. Toher, S. Curtarolo, Nat. Rev. Mater. 6 (2021)
strained global optimization in the high-dimensional “process- 730–755.
structure-property” space. Powered by generative models, the [10] L.F. Zhu, F. Körmann, Q. Chen, M. Selleby, J. Neugebauer, B. Grabowski, npj
Comput. Mater. 10 (2024) 274.
inverse design is moving toward synthesizing physically con- [11] G. Peteani, M.T.D. Huynh, G. Gerebtzoff, R. Rodríguez-Pérez, Nat. Commun. 15
sistent and manufacturable microstructure blueprints, demon- (2024) 5764.
strating the capability to effectively close the PSPP loop. [12] K. Choudhary, B. DeCost, C. Chen, A. Jain, F. Tavazza, R. Cohn, C.W. Park,
A. Choudhary, A. Agrawal, S.J.L. Billinge, E. Holm, S.P. Ong, C. Wolverton, npj
In summary, the true power of DL lies not in replacing tradi- Comput. Mater. 8 (2022) 59.
[13] M. Ragone, R. Shahabazian-Yassar, F. Mashayek, V. Yurkiv, Prog. Mater. Sci.
tional materials science knowledge but in serving as a powerful 138 (2023) 101165.
“accelerator” and “connector,” deeply integrated with classical the- [14] X.Y. Zheng, X.B. Zhang, T.-T. Chen, I. Watanabe, Adv. Mater. 35 (2023)
ory, computational simulation, and high-throughput experimenta- 2302530.
[15] K.Y. Zheng, Z.P. He, L. Che, H. Cheng, M. Ge, T.Y. Si, X.T. Xu, Mater. Sci. Semi-
tion. Future development will inevitably move beyond the pursuit cond. Process. 179 (2024) 108514.
of performance in single models toward constructing a closed-loop [16] B. Ahmadikia, O. Paraskevas, W.V. Hyning, J. Hestroffer, I.J. Beyerlein,
intelligent system driven by data, constrained by physics, and vali- C. Thrampoulidis, Acta Mater. 265 (2024) 119585.
[17] S.L. Liu, V. Kaban, O. Gutowski, T. Volkmann, K. Nielsch, I. Kaban, Acta Mater.
dated by autonomous experiment. This system will span the entire
292 (2025) 121017.
materials lifecycle—from atomic-scale design to engineering com- [18] Z. Wu, Z. Wang, J. Zeng, M.L. Zhang, K. Yao, X.Y. Chen, Q.A Li, B.S. Liu, J.
ponent service—ultimately realizing the grand vision of “design- Magnes. Alloy. 13 (12) (2025) 6100–6120.
[19] M.K. Padhan, A. Rai, M. Mitra, Comput. Mater. Sci. 229 (2023) 112416.
on-demand” and “manufacture-with-precision”, thereby marking
[20] V. Vakharia, J. Vora, S. Khanna, R. Chaudhari, M. Shah, D.Y. Pimenov, K. Giasin,
the definitive entry of metallic materials research into the era of P. Prajapati, S. Wojciechowski, J. Mater. Res. Technol. 18 (2022) 325–337.
AI-driven intelligence. [21] C. Yang, Y. Kim, S. Ryu, G.X. Gu, Mater. Des. 189 (2020) 108509.
[22] F. Gapsari, F. Utaminingrum, C.W. Lai, K. Anam, A.M. Sulaiman, M.F. Haidar,
T.S. Julian, E.E. Ebenso, J. Mater. Res. Technol. 30 (2024) 1116–1127.
Declaration of competing interest [23] X. Qin, Q.H. Wang, X.Q. Zhao, S.X. Xia, L. Wang, J.B. Long, Y.H. Zhang, Y.F. Chai,
D.L. Chen, J. Magnes. Alloy. (2025), doi:10.1016/[Link].2025.06.005.
The authors declare that they have no known competing finan- [24] K. Choudhary, B. DeCost, npj Comput. Mater. 7 (2021) 185.
[25] V. Fung, J.X. Zhang, E. Juarez, B.G. Sumpter, npj Comput. Mater. 7 (2021) 84.
cial interests or personal relationships that could have appeared to [26] J. Wang, H. Kwon, S.H. Oh, H.S. Kim, B.J. Lee, Acta Mater. 286 (2025) 120716.
influence the work reported in this paper. [27] X.Q. Zhao, X. Qin, S.X. Xia, J.B. Long, D.B. Xia, H.B. Yang, D. Zhao, Q.H. Wang,
D.L. Chen, J. Magnes. Alloy. (2026), doi:10.1016/[Link].2025.101964.
[28] G.-J. Sim, M.-G. Lee, M.I. Latypov, Scr. Mater. 255 (2025) 116407.
CRediT authorship contribution statement
[29] S. Kumar, R. Jaafreh, N. Singh, K. Hamad, D.H. Yoon, J. Magnes. Alloy. 12
(2024) 3216–3228.
Yulin Shengcao: Writing – original draft, Methodology, Investi- [30] P. Wang, Y. Jiang, W.J. Liao, R. Wang, M.J. Lai, H.C. Kou, X.B. Liang, J.S. Li,
T. Lookman, R.H. Yuan, Acta Mater. 296 (2025) 121275.
gation, Formal analysis, Data curation, Conceptualization. Xu Qin:
[31] G. Antonesi, T. Cioara, I. Anghel, V. Michalakopoulos, E. Sarmas, L. Toderean,
Writing – original draft, Methodology, Investigation, Formal anal- Appl. Energy 401 Part: A (2025) 126670.
ysis, Data curation, Conceptualization. Qinghang Wang: Writing – [32] A. Sherstinsky, Physica D 404 (2020) 132306.
original draft, Methodology, Investigation, Funding acquisition, For- [33] S. Byun, J. Yu, S. Cheon, S.H. Lee, S.H. Park, T. Lee, J. Magnes. Alloy. 12 (2024)
186–196.
mal analysis, Data curation, Conceptualization. Huabao Yang: Writ- [34] Y.B. Dan, Y. Zhao, X. Li, S.B. Li, M. Hu, J.J. Hu, npj Comput. Mater. 6 (2020) 84.
ing – review & editing, Methodology, Investigation, Formal analysis, [35] E. Azqadan, H. Jahed, A. Arami, Acta Mater. 261 (2023) 119406.

106
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

[36] C. Zeni, R. Pinsler, D. Zügner, A. Fowler, M. Horton, X. Fu, Z. Wang, [86] F.M. Bai, Y.Y. Huang, J.L. Wang, H.W. Zhou, X.T. Wu, H.L. Wei, L.Q. Zhang,
A. Shysheya, J. Crabbe, S. Ueda, R. Sordillo, L.X. Sun, J. Smith, B. Nguyn, P.K. Liaw, Mater. Chem. Phys. 343 (2025) 131068.
H. Schulz, S. Lewis, C.W. Huang, Z.H. Lu, Y.C. Zhou, H. Yang, H.X. Hao, J.L. Li, [87] L.J. Li, G.J. Liu, H.R. Yu, Z.Y. Chen, H.J. Kang, Z.N. Chen, E.Y. Guo, J.C. Jie,
C.L. Yang, W.J. Li, R. Tomioka, T. Xie, Nature 639 (2025) 624–632. T.M. Wang, J. Mater. Sci. Technol. 267 (2026) 184–197.
[37] C. Cui, F.F. Chamasemani, M. Reisinger, P. Imrich, W. Hartner, W. Robl, R. Brun- [88] Z.Y. Yang, S. Li, S. Li, J. Yang, D.D. Liu, Comput. Mater. Sci. 220 (2023) 112064.
ner, npj Mater. Degrad. 9 (2025) 55. [89] A.H. Joy, S. Zoha, S. Akhter, Mater. Today Commun 46 (2025) 112288.
[38] Y.W. Mao, S. Keshavarz, M.N.T. Kilic, K.W. Wang, Y.J Li, A.C.E. Reid, W.K. Liao, [90] Q.H. Wang, X. Qin, S.X. Xia, L. Wang, W.Q. Wang, W.Y. Huang, Y. Song,
A. Choudhary, A. Agrawal, Scr. Mater. 254 (2025) 116315. W.N. Tang, D.L. Chen, J. Magnes. Alloy. 13 (2025) 2866–2883.
[39] S. Keshavarz, Y.W. Mao, A.C.E. Reid, A. Agrawal, Int. J. Plast. 185 (2025) [91] Y. Shen, J. Du, Y. He, H. Wang, T. Hao, X. Niu, Q. Jiang, Mater. Today Commun.
104221. 47 (2025) 113198.
[40] J. You, C. Wang, H. Ju, S.-Y. Hu, Y.-Z. Wang, H.-Y. Wang, J. Mater. Sci. Technol. [92] J. Jung, J. Na, H.K. Park, J.M. Park, G. Kim, S. Lee, H.S. Kim, npj Comput. Mater.
238 (2025) 24–35. 7 (2021) 96.
[41] P. Wu, H.W. Huang, Q.C. Yang, B. Qian, Y.X. Gao, Y.G. Yang, H.R. Zhang, [93] E. Azqadan, A. Arami, H. Jahed, J. Magnes. Alloy. 13 (2025) 4231–4244.
Q. Zhen, Comput. Mater. Sci. 256 (2025) 113883. [94] Y. Murakami, R. Furushima, K. Shiga, T. Miyajima, N. Omura, Acta Mater. 286
[42] K. Kano, K. Koga, Comput. Mater. Sci. 244 (2024) 113211. (2025) 120683.
[43] N.N. Nyangiwe, Next Mater. 8 (2025) 100683. [95] J. Yin, Q. Lei, X. Li, X.Y. Zhang, X.P. Meng, Y.B. Jiang, L. Tian, S. Zhou, Z. Li, Acta
[44] J.W. Leng, J.J. Wang, L.H. Zhou, R.L. Zhao, C. Chen, D. Zhang, S. Zheng, Q. Liu, Mater. 243 (2023) 118420.
W.M. Shen, Adv. Eng. Inform. 70 (2026) 104117. [96] P.L. Zhao, Y.W. Wang, B.Y. Jiang, H.M. Zhang, X.W. Cheng, Q.B. Fan, Mater. Sci.
[45] H.Y. Chen, Y.B. Liu, Y. Liang, Y. Chen, W.B. Lin, C.J. Fu, C.S. Luo, Adv. Eng. In- Eng. A 879 (2023) 145202.
form. 69 (2026) 104029. [97] A. Gollapalli, A.K. Singh, Comput. Mater. Sci. 253 (2025) 113854.
[46] M. Lavrskyi, N. Gey, B. Gueraud, A. Couturier, M. Salib, L. Germain, Mater. [98] D. Ren, C.C. Wang, X.L. Wei, Q.Q. Lai, W. Xu, Acta Mater. 252 (2023) 118954.
Charact. 230 (2025) 115718. [99] S.Y. Han, C.C. Wang, Q.Q. Lai, L.Y. Wang, W. Xu, H.S. Di, Acta Mater. 289 (2025)
[47] K. Kano, T. Segi, H. Ozono, K. Koga, J. Power Sources 655 (2025) 237944. 120936.
[48] G.E. Kamiadakis, I.G. Kevrekidis, L. Lu, P. Perdikaris, S.F. Wang, L. Yang, Nat. [100] R. Furushima, Int. J. Refract. Met. Hard Mater. 134 (2026) 107492.
Rev. Phys. 3 (2021) 422–440. [101] Y. Cao, S.W. Wu, S. Tang, G.M. Cao, C.D. Zhang, P. Hedstrom, X.G. Zhou,
[49] L. Ma, X.M. Duan, K.Y. Chen, C.Y. Hou, H.Q. Li, J.Y. Ye, Y.Z. Wu, J. Rong, X.H. Yu, Z.Y. Liu, Metall. Mater. Trans. A 56 (2025) 168–179.
Y. Wei, J. Feng, Chem. Eng. Sci. 321 (2026) 122793. [102] S. Li, J. Yang, S. Li, D.R. Liu, M.Y. Zhang, Trans. Nonferrous Met. Soc. China 35
[50] C.C. Wang, K.Y. Zhu, P. Hedström, Y. Li, W. Xu, J. Mater. Sci. Technol. 128 (2025) 1855–1874.
(2022) 31–43. [103] M.Y.A. Salam, E.N. Ogunmuyiwa, V.K. Manisa, A. Yahya, I.A. Badruddin, J.
[51] G.Y. Hu, M.I. Latypov, Comput. Mater. Sci. 243 (2024) 113121. Mater. Res. Technol. 38 (2025) 1934–1946.
[52] Y.Z. Hu, G.W. Zhou, M.G. Lee, P.D. Wu, D.Y. Li, Int. J. Plast. 179 (2024) [104] G.X. Song, D.M. Fu, Z.W. Qiu, Z.J. Yang, J.X. Dai, L.W. Ma, D.W Zhang, Appl.
104017. Soft Comput. 185 (2025) 113776.
[53] B.Y. Zheng, J.R. Wen, Y.S. Choy, C.W. Fei, Expert Syst. Appl. 298 (2026) 129667. [105] R. Kushwaha, A. Chauhan, Eng. Fract. Mech. 332 (2026) 111782.
[54] S. Kench, S.J. Cooper, Nat. Mach. Intell. 3 (2021) 299–305. [106] M. Madani, V. Lacivita, Y. Shin, A. Tarakanova, npj. Comput. Mater. 11 (2025)
[55] I. Watanabe, K. Sugiura, T.T. Chen, T. Ogawa, Y. Adachi, Sci. Technol. Adv. 15.
Mater. 25 (2024) 2388501. [107] V. Venugopal, E. Olivetti, Sci. Data 11 (2024) 217.
[56] Y. Li, Y. Wei, A. Saxena, M. Kühbach, C. Freysoldt, B. Gault, Prog. Mater. Sci. [108] F. Therrien, E.H. Sargent, O. Voznyy, Nat. Commun. 16 (2025) 4301.
156 (2026) 101561. [109] P. Kishore, A. Mondal, A. Trivedi, P. Singh, A. Alankar, Int. J. Fatigue 192 (2025)
[57] M. Ziatdinov, A. Ghosh, C.Y. Wong, S.V. Kalinin, Nat. Mach. Intell. 4 (2022) 108724.
1101–1112. [110] J.L. Zhou, Y.X. Zhang, N. Wang, W.J. Gao, L.E. Liu, L. Tang, J. Wang, J.X. Lu,
[58] R.Q. Lin, R. Zhang, C.Y. Wang, X.Q. Yang, H.L.L. Xin, Sci. Rep. 11 (2021) 5386. Y.F. Zhang, Z. Zhang, Int. J. Fatigue 190 (2025) 108637.
[59] P. Vashishtha, H.G. Kattamuri, N. Thawari, M. Amirthalingam, R. Batra, Nat. [111] A. Thomas, A.R. Durmaz, M. Alam, P. Gumbsch, H. Sack, C. Eberl, Sci. Rep. 13
Mach. Intell. 7 (2025) 79–84. (2023) 12562.
[60] S.H. Yang, W. Choi, B.W. Cho, F.O.T. Agyapong-Fordjour, S. Park, S.J. Yun, [112] M.Y. Dai, M.F. Demirel, Y.Y. Liang, J.M. Hu, npj Comput. Mater. 7 (2021) 103.
H.J. Kim, Y.K. Han, Y.H. Lee, K.K. Kim, Y.M. Kim, Adv. Sci. 8 (2021) 2101099. [113] X.L. Wei, S.V.D. Zwaag, Z.X. Jia, C.C. Wang, W. Xu, Acta Mater. 235 (2022)
[61] Y.M. Guo, S.V. Kalinin, H. Cai, K. Xiao, S. Krylyuk, A.V. Davydov, Q.Y. Guo, 118103.
A.R. Lupini, npj Comput. Mater. 7 (2021) 180. [114] X.L. Wei, C. Zhang, S.Y. Han, Z.X. Jia, C.C. Wang, W. Xu, Int. J. Fatigue 163
[62] K. Wijesinghe, J. Wanni, N.K. Banerjee, S. Banerjee, A. Achuthan, Mater. Des. (2022) 107050.
208 (2021) 109926. [115] L. Xiao, X.Y. Xue, N. Wang, Q.B. Ren, J.M. Hua, F. Wang, Constr. Build. Mater.
[63] F.X. Lin, H.X. Fang, H. Liu, Y.B. Zhang, D.J. Jensen, E. Hovad, Mater. Charact. 400 (2023) 132679.
219 (2025) 114576. [116] B.H. Chen, J.F. Zhang, S.C. Zhou, G.P. Zhang, F. Xu, J. Mater. Res. Technol. 32
[64] C. Gorynski, M. Frei, F.E. Kruis, M. Winterer, Acta Mater. 256 (2023) 119106. (2024) 2767–2779.
[65] A.S. Saleh, K. Croes, H. Ceric, I.D. Wolf, H. Zahedmanesh, Comput. Mater. Sci. [117] M.M. Yu, X. Xie, Eng. Appl. Artif. Intell. 136 (2024) 109054.
251 (2025) 113723. [118] Z.Y. Chen, Z.J. Han, S.B. Xia, Z.X. Li, Q.L. Huang, W.W. Xu, Eng. Fract. Mech.
[66] X.Y. Zhou, Y. Wei, M. Kühbach, H. Zhao, F. Vogel, R.D. Kamachali, G.B. Thomp- 321 (2025) 111087.
son, D. Raabe, B. Gault, Acta Mater. 226 (2022) 117633. [119] P. Zhang, A.B. Wang, K.K. Tang, Eng. Fract. Mech. 326 (2025) 111372.
[67] B.Y. Ma, X.Y. Wei, C. Liu, X.J. Ban, H.Y. Huang, H. Wang, W.H. Xue, S. Wu, [120] S. Anjum, M. Khan, C. Nnaji, A. Aryal, A. Koh, Autom. Constr. 181 (2026)
M.F. Gao, Q. Shen, M. Mukeshimana, A.O. Abuassba, H.K Shen, Y.J Su, npj 106680.
Comput. Mater. 6 (2020) 125. [121] G. Minerva, M. Awd, A. Koch, F. Walther, S. Beretta, Int. J. Fatigue 195 (2025)
[68] J. Na, J. Lee, S.-H. Kang, S.J. Kim, S. Lee, Mater. Charact. 206 (2023) 113410. 108852.
[69] K.C. Zhang, C. Lv, K. Wang, Y.Y. Ma, W.L. Wu, Z.L. Wang, W.L. Xiao, M.H. Cai, [122] R. Mostakim, R.T. Batley, S. Saha, Int. J. Fatigue 202 (2026) 109185.
L.C. Qi, X.Q. Zhao, J. Mater. Sci. Technol. 243 (2026) 237–244. [123] Z.X. Zhan, M.W. Zhang, X.F. He, X.F. Li, Z.H. Wang, X.L. Chen, B. Han, W.P. Hu,
[70] M.H. Liu, C.W. Du, T. Wang, C. Hai, X.G. Li, Mater. Sci. Eng. A 945 (2025) Q.C. Meng, H. Li, Int. J. Struct. Integr. 16 (2025) 1255–1316.
148993. [124] Y.D. Zhao, Y.J. Xiang, K.K. Tang, Eng. Fract. Mech. 303 (2024) 110106.
[71] M. Müller, M. Stiefel, B.-I. Bachmann, D. Britz, F. Mücklich, Metals 14 (2024) [125] A. Al-Hagri, H. Stang, J.P. Waldbjørn, A. Kolios, E. Katsanos, Eng. Fract. Mech.
553. 332 (2026) 111800.
[72] L.T.H. Nguyen, K. Edalati, V. Kshitij, M. Dadfarnia, A. Staykov, T. Tsuchiyama, [126] Z.H. Wang, Z.X. Zhan, Q.Y. Xia, Y.J. Zhang, Q. Qin, X.Y. Li, W.P. Hu, Q.C. Meng,
M. Kubota, H. Matsunaga, J.A. Krogstad, P. Sofronis, Mater. Sci. Eng. A 953 H. Li, Eng. Fract. Mech. 318 (2025) 110915.
(2026) 149714. [127] W.Q. Tang, Z.X. Luo, D.Y. Li, Y.H. Peng, Addit. Manuf. 115 (2026) 105047.
[73] S.H. Tian, X. Jiang, W.R. Wang, Z.H. Jing, C. Zhang, C. Zhang, T. Lookman, [128] L.H. Tian, Y. Gong, H. Liu, S.Z. Hao, J.Y. Zhang, L.B. Zhao, N. Hu, Int. J. Fatigue
Y.J. Su, Acta Mater. 285 (2025) 120663. 206 (2026) 109442.
[74] K. Dia, T. David, L. Guetaz, Z. Saghi, G. Daniel, Mater. Des. 260 (2025) 115200. [129] L.F. Jiang, H.L. Luo, Y.X. Liu, Y.N. Hu, H. Li, X.J. Shao, X. Zhang, Q.H. Kan,
[75] M. Lavrskyi, N. Gey, B. Gueraud, A. Couturier, M. Salib, L. Germain, Mater. G.Z. Kang, Int. J. Fatigue 204 (2026) 109367.
Charact. 230 (2025) 115718. [130] S.S. Yang, Z.H. Zhang, R. Guo, Z.X. Zhan, Int. J. Fatigue 203 (2026) 109296.
[76] F.L. Zheng, L.F. Xie, Q.S. Bai, Y.Y. Zhu, M. Yin, Y.H. Zhang, K.Y. Niu, Addit. [131] E. Akbari, T.N. Chakherlou, H. Tabrizchi, A. Mosavi, Comput. Model. Eng. Sci.
Manuf. 97 (2025) 104628. 145 (1) (2025) 305–325.
[77] D.M.D.O. Zapiain, J.A. Stewart, R. Dingreville, npj Comput. Mater. 7 (2021) 3. [132] Y.J. Guo, H. Meng, Z.H. Zhu, J.Y. Wang, P.Y. Wang, T. Zuo, S.H. Huo, Z.F. Yue,
[78] K.G. Zhang, W. Chen, W.K. Liu, L.C. Brinson, D.W. Apley, Acta Mater. 301 Int. J. Fatigue 204 (2026) 109347.
(2025) 121588. [133] S.Y. Li, J.Y. Xie, W.Y. Zhang, J. Build. Eng. 117 (2026) 114859.
[79] S. Kench, I. Squires, A. Dahari, S.J. Cooper, Sci. Data 9 (2022) 645. [134] Y.W. Wang, Y. Zeng, J.R. Tang, H.M. Tan, T.Y. Sun, C. Wu, Theor. Appl. Fract.
[80] B. Murgas, J. Stickel, S. Ghosh, npj Comput. Mater. 10 (2024) 32. Mech. 143 (2026) 105456.
[81] B. Murgas, J. Stickel, L. Brewer, S. Ghosh, Nat. Commun. 15 (2024) 9441. [135] L.W. Dang, X.F. He, D.C. Tang, H. Xin, B. Wu, Int. J. Struct. Integr. 16 (2025)
[82] T. Zhang, N.J. Bian, X. Li, Comput. Aided Des. 176 (2024) 103760. 327–354.
[83] X.R. Lyu, X.D. Ren, Sci. Rep. 14 (2024) 5041. [136] S.W. Zhou, M. Henrich, Z.C. Wei, F. Feng, B. Yang, S. Münstermann, Eng. Fract.
[84] K.H. Lee, G.J. Yun, npj Comput. Mater. 10 (2024) 99. Mech. 322 (2025) 111136.
[85] J. Phan, M. Sarmad, L. Ruspini, G. Kiss, F. Lindseth, Sci. Rep. 14 (2024) 6498. [137] J.K. Shu, Z.Q. Fu, X.C. Li, B.H. Ji, Q.D. Wang, Int. J. Fatigue 199 (2025) 109061.

107
Y. Shengcao, X. Qin, Q. Wang et al. Journal of Materials Science & Technology 271 (2026) 91–108

[138] H.H. Liu, X. Ding, J.H. Liu, Y.L. Zhang, B. Zhang, E.L. Li, Z.X. Guo, Mech. Mater. [170] D. Raabe, J.R. Mianroodi, J. Neugebauer, Nat. Comput. Sci. 3 (2023) 198–209.
209 (2025) 105429. [171] A. Venkatraman, M.A. Wilson, D. Montes de Oca Zapiain, npj Comput. Mater.
[139] L.Y. Wang, S.-P. Zhu, B.R. Wu, Z.J. Xu, C.Q. Luo, Q.Y. Wang, Comput. Methods 11 (2025) 24.
Appl. Mech. Eng. 439 (2025) 117924. [172] K.N. Sasidhar, N.H. Siboni, J.R. Mianroodi, J. Neugebauer, D. Raabe, Sci. Adv. 9
[140] Q.Z. Tian, L. Chang, J.W. Li, C.Y. Zhou, Theor. Appl. Fract. Mech. 139 (2025) (2023) 32.
105118. [173] S.Y. Jin, Z.H. Fu, G.X. Song, J.M. Pei, H.Y. Gong, L.W. Ma, D.W. Zhang, Corros.
[141] H.Y. Liao, J. Pan, X.H. Su, X.Y. Sun, X. Chen, Int. J. Fatigue 193 (2025) 108799. Sci. 256 (2025) 113217.
[142] F. Feng, T. Zhu, B. Yang, Z. Zhang, S.W. Zhou, S.N. Xiao, Expert Syst. Appl. 275 [174] M. Busch, M. Tacke, S.V. Lamaka, M.L. Zheludkevich, K. Linka, C.J. Cyron,
(2025) 127098. C. Feiler, R.C. Aydin, Corros. Sci. 255 (2025) 113080.
[143] Y.X. Dong, X.F. Yang, D.D. Chang, Q. Li, Int. J. Fatigue 190 (2025) 108626. [175] P. Fernandez-Zelaia, S. Thapliyal, R. Kannan, P. Nandwana, Y. Yamamoto,
[144] T.H. Huang, X.Y. Li, Y.Z. Zhang, L.J Yao, T. Zhang, Eng. Fail. Anal. 174 (2025) A. Nycz, V. Paquit, M.M. Kirka, Addit. Manuf. 94 (2024) 104478.
109521. [176] C.Y. Lee, C.Y. Jui, A.C. Yeh, Y.J. Chang, W.J. Lee, J. Alloy. Compd. 976 (2024)
[145] A. Shahriari, S. Shakerin, N. Birbilis, M. Mohammadi, Corros. Sci. 257 (2025) 173144.
113343. [177] Z.Q. Chen, Y.H. Shang, X.D. Liu, Y. Yang, npj Comput. Mater. 10 (2024) 204.
[146] Y.X. Lou, Z. Shen, Z.Y. Dong, K. Zhang, J.Q. Li, Y.H. Wu, Z.F. Jiang, H. Wang, [178] Z.S. Zhang, D.Y. Liang, G.B. Tang, Y.B. Hua, B.W. Dong, Z.W. Liang, Eng. Appl.
Y. Ding, X.T. Zhang, Y. Wang, K. Chen, X.Q. Zeng, Corros. Sci. 257 (2025) Artif. Intell. 165 (2026) 113474.
113354. [179] C.L. Qin, S.Y. Ma, L. Zhao, J.D. Liu, G. Jiang, Int. J. Hydrog. Energy 172 (2025)
[147] S.Q. Jian, Z. Wang, Z.L. Yang, J.X. Ma, C. Yang, J. Wang, L. Zhang, Corros. Sci. 151363.
257 (2025) 113277. [180] A. Chaudhari, C. Guntuboina, H.S. Huang, A.B. Farimani, Comput. Mater. Sci.
[148] V.K. Beura, Y. Karanth, K. Darling, K. Solanki, npj Mater. Degrad. 6 (2022) 62. 244 (2024) 113256.
[149] A. Bahmani, S. Moradi, M. Lotfpour, H.T. Jeong, W.J. Kim, Corros. Sci. 230 [181] S.F. Nabavi, H. Garmestani, F. Fekri, J. Manuf. Process. 156 (2025) 86–120.
(2024) 111892. [182] S.P.R. Guduru, M.O. Kekung, R.T. Ott, S. Roy, P. Singh, Acta Mater. 304 (2026)
[150] C. Evered, C. Falconer, J.L. Liu, M.R. Tonks, K. Sridharan, A. Couet, Corros. Sci. 121763.
258 (2026) 113385. [183] F. Tan, Z.X. Zhao, Y.B. Jiang, W.C. Zhang, T. Xie, W. Chen, M.Z. Ma, Y.F. Liu,
[151] L. Wang, Y.H. Peng, T. Nakata, C. Xu, J. Sun, Z. Zheng, Y. Li, F.Y. Cao, L. Geng, Y.P. Ye, Z. Xiao, Q. Lei, G.F. Xu, J. Ren, Y.Y. Zhao, Z. Li, Acta Mater. 306 (2026)
S. Kamado, Corros. Sci. 255 (2025) 113151. 121886.
[152] K. Hagihara, M. Okubo, M. Yamasaki, T. Nakano, Corros. Sci. 109 (2016) 68–85. [184] Z.R. Pei, J.Q. Yin, P.K. Liaw, D. Raabe, Nat. Commun. 14 (2023) 54.
[153] C.J. Yan, Y.C. Xin, X.B. Chen, D.K. Xu, P.K. Chu, C.Q. Liu, B. Guan, X.X. Huang, [185] D. Raabe, J.R. Mianroodi, J. Neugebauer, Nat. Comput. Sci. 3 (2023) 198–209.
Q. Liu, Nat. Commun. 12 (2021) 4616. [186] P.J. Xu, K.M. Hu, Y.C. Wu, H.D. Zhang, Z.M. Lv, Z.Y. Zhang, Y.Q. An, Q. Sun,
[154] M.E.A.S. Ben, T.T. Truong, C. Feiler, D. Höche, Results Eng. 25 (2025) 104417. X. He, J. Manuf. Syst. 83 (2025) 134–156.
[155] Y. Yılmaz, S. Nayır, Ş. Erdoğdu, Mater. Today Commun. 44 (2025) 112050. [187] F.X. Wang, A.G. Iwanicki, A.T. Sose, L.A. Pressley, T.M. McQueen, S.A. Desh-
[156] C. Sun, P. Sriboriboon, J. Han, S.J. Ko, S.Y. Lee, Y. Heo, J.H. Shim, S. Yang, mukh, npj Comput. Mater. 11 (2025) 124.
J.G. Kim, Y. Kim, Corros. Sci. 240 (2024) 112431. [188] W.J. Zhao, Z.Q. Chen, Y.H. Shang, Q. Wang, L. Wang, B. Liu, Y. Liu, Y. Yang, npj
[157] J.L. Zhang, Y.K. Zeng, B. Starly, SN Appl. Sci. 3 (2021) 442. Comput. Mater. 11 (2025) 278.
[158] Q.F. Yu, Y.D. Han, Y. Han, X.J. Gao, L.Y. Zheng, J. Mar. Sci. Eng. 12 (2024) 2295. [189] X. Qin, Q.H. Wang, X.Q. Zhao, S.X. Xia, L. Wang, Y.H. Zhang, C. He, D.L. Chen,
[159] Y.F. Wang, D.B. Wang, J. Gao, Z.X. Meng, J.X. Hou, J.Q. Wang, X.P. Wang, Corros. B. Jiang, Scr. Mater. 265 (2025) 116762.
Sci. 260 (2026) 113567. [190] Y. Kim, H.K. Park, J. Jung, P. Asghari-Rad, S. Lee, J.Y. Kim, H.G. Jung, H.S. Kim,
[160] C. Özkan, L. Sahlmann, T. Würger, C. Feiler, S. Lamaka, M. Zheludkevich, Mater. Des. 202 (2021) 109544.
P. Taheri, A. Mol, npj Mater. Degrad. 9 (2025) 157. [191] Z.R. Pei, K.A. Rozman, Ö.N. Dŏgan, Y.H. Wen, N. Gao, E.A. Holm, J.A. Hawk,
[161] S.Y. Xu, F. Li, C.C. Du, D.C. Ju, Y.Q. Hou, X.Y. Li, npj Mater. Degrad. 9 (2025) D.E. Alman, M.C. Gao, Adv. Sci. 8 (2021) 2101207.
131. [192] A. Ikeda, R. Higuchi, T. Yokozeki, K. Endo, Y. Kojima, M. Suzuki, M. Mura-
[162] B.V.S. Reddy, A.M. Shaik, C.C. Sastry, J. Krishnaiah, S. Patil, C.P. Nikhare, J. matsu, Sci. Rep. 15 (2025) 37147.
Manuf. Process. 137 (2025) 320–341. [193] X.R. Lyu, X.D. Ren, Mater. Today Commun. 48 (2025) 113087.
[163] X.R. Wang, W. Zhang, Z.F. Lin, H.J. Li, Y.Q. Zhang, W.Y. Quan, Z.W. Chen, [194] Y.C. Zhang, P. Seibert, A. Otto, A. Raßloff, M. Ambati, M. Kästner, Comput.
X.Q. You, Y. Zeng, G. Wang, B.L. Luo, Z.H. Yu, Coatings 14 (2024) 1051. Mater. Sci. 232 (2024) 112661.
[164] A. Ucar, M. Karakose, N. Kırımça, Appl. Sci. 14 (2024) 898. [195] T. Long, N.M. Fortunato, I. Opahle, Y.X. Zhang, I. Samathrakis, C. Shen, O. Gut-
[165] M. Ghorbani, Z. Li, A. Pasquini, N. Birbilis, npj Mater. Degrad. 9 (2025) 155. fleisch, H.B. Zhang, npj Comput. Mater. 7 (2021) 66.
[166] X.K. Qi, A.B. Liu, Y.Y. Li, L.X. Dang, J.L. Zhang, L. Wang, Mater. Today Commun. [196] V. Attari, D. Khatamsaz, D. Allaire, R. Arroyave, Acta Mater. 259 (2023)
44 (2025) 111915. 119204.
[167] S.Y. Ma, D.W. Zhang, P.L. Zhang, B. Markert, Comput. Mater. Sci. 247 (2025) [197] C.-Z. Liu, W. An, Q. Zhou, Q.-L. Xiong, Mater. Des. 260 (2025) 115159.
113546. [198] X. Shang, Z.Y. Liu, J.H. Zhang, T.Y. Lyu, Y. Zou, Mater. Today 70 (2023) 71–81.
[168] W.S. Peng, J. Hou, W.M. Guo, J.Z. Li, T.G. Duan, npj Mater. Degrad. 9 (2025) [199] N. Kusampudi, M. Diehl, Int. J. Plast. 171 (2023) 103776.
125. [200] Y.H. Zang, P.S. Koutsourelakis, Acta Mater. 284 (2025) 120600.
[169] S.W. Feng, X.Y. Sun, G. Chen, X. Chen, Eng. Appl. Artif. Intell. 141 (2025)
109807.

108