Assessing stability remains a fundamental prerequisite for deploying materials across a wide range of applications, including batteries, catalysts, and photovoltaics. However, first-principles stability checks such as phonon dispersion and energy above hull calculations typically require days to weeks of computing time per composition, creating a critical bottleneck for truly high-throughput discovery. In this Perspective, we highlight the underutilized potential of geometric tolerance factors (T_f) as lightweight yet informative indicators for rapid stability assessment. First, we review the T_f developed for representative materials systems, including perovskites, spinels, and garnets, and analyze recent cases where such indicators have been integrated into AI-driven materials discovery. Then, we identify key open challenges in designing T_f that are both accurate and generalizable, as well as in effectively incorporating them into AI frameworks. The potential solutions, including active learning for multi-composition structure, electron density profile-based learning for ionic radii estimation, and diffusion model for thermodynamic and kinetic stability, are proposed to address these challenges. The synergy between T_f-based heuristics and advanced AI models has the potential to triage vast compositional spaces before committing to expensive first-principles stability validation, thereby enabling broader innovations in materials design and deployment.

Graph neural networks (GNNs) have become a transformative modeling paradigm in materials science, offering a data-efficient and structure-aware approach for learning from complex material systems. This review focuses on the recent progress of GNNs in alloy design and property prediction. We begin by introducing the foundational concepts of graph representations and the general architecture of GNNs, including node embeddings, message passing, and pooling strategies. The review then categorizes major types of GNNs, including supervised and unsupervised learning, with a focus on the achievements and applications of GNNs in materials modeling, and discusses their strengths and inherent limitations in the context of materials modeling. Particular emphasis is placed on the application of GNNs in the alloy domain, covering a diverse range of data types, from atomic structures and compositions to microstructural images, and target properties, such as mechanical strength, thermal stability, and phase stability. We highlight how GNNs are integrated into alloy composition optimization, multi-property prediction, and frontier research workflows. The review concludes with a summary of multi-model and multiscale approaches and outlines key challenges and future directions for constructing generalizable, physics-informed GNN frameworks for alloy discovery.

Phase-change materials (PCMs) are among the most promising candidates for next-generation non-volatile memory and neuromorphic computing technologies. However, their photonic applications are hindered by a trade-off between refractive index contrast and optical absorption losses. Artificial intelligence-assisted computational approaches are essential for fundamental understanding and device modeling of PCMs. In this work, we systematically investigate structural and optical properties of crystalline and amorphous Ge₂Sb₂Se_xTe_5-x (x = 0 to 4) alloys using density functional theory (DFT), and then use the DFT-computed optical parameters for modeling and optimization of photonic computing devices via the finite-difference time-domain method. Among the investigated compositions, we identify a promising candidate, i.e., Ge₂Sb₂Se₃Te₂ for all-optical switching on a silicon-on-insulator (SOI) platform. Finally, we design a dual-disk PCM waveguide structure on SOI with an enhanced switching contrast and a low optical loss for scalable photonic neural network application.

Uncertainty is crucial when the available data for building a predictor are insufficient, which is ubiquitous in machine-learning-driven materials studies. However, the impact of uncertainty estimation on predictor selection and materials optimization remains incompletely understood. Here, we demonstrate that in active learning, uncertainty estimation significantly influences predictor selection, as well as that the calibration of uncertainty estimation can improve the optimization. The idea is validated on three alloy datasets (Ni-based, Fe-based, and Ti-based) using three commonly used algorithms - support vector regression (SVR), neural networks (NN), and extreme gradient boosting (XGBoost) - which yield comparable predictive accuracy. It is shown that XGBoost presents more reliable uncertainty estimation than SVR and NN. Using the directly estimated uncertainty for the three predictors with similar accuracy, we find that the optimization is quite different. This suggests that uncertainty estimation plays a role in predictor selection. The uncertainty estimation is then calibrated to improve reliability, and its effect on optimization is compared with the uncalibrated case. Among the nine cases considered (three models and three datasets), eight show improved optimization when calibrated uncertainty estimation is used. This work suggests that uncertainty estimation and its calibration deserve greater attention in active learning-driven materials discovery.

Co-based superalloys exhibit exceptional high-temperature properties, granting them broad application prospects in the superalloy domain. However, constrained by the exorbitant trial-and-error costs and protracted research cycles inherent in their development, machine learning (ML) has emerged as the most pivotal research direction in this field. This review systematically examines ML-driven approaches for Co-based superalloys, progressing from fundamental regression models for property prediction to advanced multi-model, multi-scale computational paradigms-structured according to model sophistication and problem complexity. Furthermore, we discuss current challenges and future prospects in applying ML to Co-based superalloys, with particular emphasis on addressing data scarcity through the integration of high-throughput experimentation. This synergistic approach enables efficient establishment of standardized superalloy databases, accelerating research progress to meet evolving demands in aerospace applications.

Rare-earth (RE) zirconates and tantalates are promising candidates for next-generation thermal barrier coatings (TBCs) due to their high-temperature stability and low thermal conductivity. However, the substantial compositional complexity introduced by multiple RE element substitutions poses significant challenges for systematic property optimization. To address these challenges, a high-throughput, data-driven computational framework was employed to systematically investigate and compare structural stability, thermodynamic properties, lattice thermal conductivity (κ_L) and fracture toughness (K_IC) of RE₂Zr₂O₇ and RE₃TaO₇ oxides (RE = Sc, Y, La ~ Lu) in their pyrochlore and Weberite-type structures, respectively. κ_L and intrinsic K_IC were systematically evaluated using phonon-scattering and Griffith-based models. The results reveal that RE₃TaO₇ exhibits consistently lower κ_L than RE₂Zr₂O₇ due to its low symmetry, heavier atomic masses and higher structural disorder. Interestingly, theoretical predictions indicate slightly higher intrinsic K_IC in RE₂Zr₂O₇, which is attributed to its ordered vacancy sublattice and symmetric bonding. In contrast, experimental data often report superior K_IC for RE₃TaO₇, likely due to extrinsic microstructural effects not captured in idealized calculations. Correlation and SHapley Additive exPlanations analyses further reveal that bond energy, charge disorder and bond-length heterogeneity are key descriptors governing κ_L and K_IC. These findings provide mechanistic insight into structure–property relationships and offer a predictive framework for the rational design of RE oxide TBC materials.

Highlights
• Integrating high-throughput first-principles calculations, lattice-level descriptor engineering and interpretable machine learning to design RE₂Zr₂O₇ and RE₃TaO₇ (RE = Sc, Y, La ~ Lu) oxide-based thermal barrier materials.
• Data-driven selection and classification of key physical descriptors (bond energy, charge disorder, bond-length heterogeneity) enable predictive modeling of κ_L and K_IC across 17 rare-earth elements.
• Combining thermodynamic stability analysis, phonon-based transport models and SHapley Additive exPlanations interpretability to establish structure–property relationships and guide rational oxide design.

Carbon-based nanomaterials, particularly carbon dots (CDs), have attracted growing attention due to their unique optical properties and cost-effective synthesis. Despite their promise, challenges remain in elucidating luminescence mechanisms and achieving controlled synthesis. Traditional trial-and-error approaches are inefficient, while machine learning (ML) offers powerful tools to accelerate materials discovery by capturing complex relationships. This review summarizes recent progress in applying ML to CDs, focusing on three key areas: enhancing the regulation of intrinsic properties, improving detection sensitivity and multicomponent recognition through the analysis of high-dimensional spectral data, and uncovering correlations between molecular features, experimental parameters, and CD performance with explainable ML. These advances enable more rational and efficient design of multifunctional CDs. Finally, we discuss future directions for CD informatics, including the development of structured data resources, the integration of large language models, interpretable ML techniques, and automated experimental platforms. These trends are expected to provide new insights and drive continued innovation in the multifunctional applications of CDs.

Aerospace alloys often operate under extreme conditions. Accurate defect segmentation in images of aerospace components is the key to quantifying the defects and evaluating their impact for part lifespan. The components usually have complex free-form surfaces, leading to uneven light distribution in images. The variable image presentations pose a great challenge for accurate segmentation, especially with limited data. Generative adversarial networks and other training-based methods are commonly used for image generation, but they still rely on sufficient high-quality training data. In this paper, a physical-based image generation method is proposed to create any possible scratches according to physical laws to improve the scratch segmentation capability with limited data. First, an efficient scratched blade surface image generation pipeline is developed. Then, a systematic strategy to maximize the effect of physical synthetic scratch images is presented. The experiments show that the segmentation intersection-over-union could be improved from 0.66 to 0.83 with only 20 real images for training, and reveal the influences of network structure, image and label quality, data fusion strategy on segmentation performance.

Casting blowhole defects seriously affect product quality and performance. Accurate detection, segmentation, and measurement of these defects are essential for quality control. To solve problems such as the varying sizes of blowholes in castings, segmentation uncertainty caused by texture overlap, and the subjectivity of manual rating, this paper proposes a rating strategy for casting blowhole defects based on image instance segmentation results. In the preprocessing stage, contrast-limited adaptive histogram equalization (CLAHE) is applied to enhance defect features. You Look Only Once version 8 (YOLOv8), YOLOv11, YOLOv13, and semantic segmentation models are compared, and YOLOv13 is chosen as the main model for segmentation. Its mean Average Precision (mAP) at an IoU threshold of 0.5 (mAP50) reaches 0.964, showing the best performance. Based on the segmentation results, the pixel area and percentage of the segmented regions are calculated. The actual defect size is then converted using the practical sampling area, and rating is performed according to the GB/T 11346-2018 standard. Validation through manual measurement in Photoshop and physical sectioning confirms that the proposed strategy reduces the maximum error by 17.1% compared with traditional manual rating. The method significantly enhances the automation and accuracy of blowhole defect rating and provides reliable technical support for casting quality control.

Extracting reliable, tuple-level information from materials texts is essential for data-driven materials design, yet multi-tuple sentences remain difficult due to intertwined semantics, syntactic complexity, and sparse supervision in higher-density cases. In this study, we address these challenges by formulating information extraction as an integrated process that couples entity extraction with tuple allocation. The framework combines an entity extraction module based on bidirectional encoder representations from transformers (MatSciBERT) with pointer networks and an allocation module that models inter- and intra-entity attention to enforce tuple coherence. Using the mechanical properties of multi-principal element alloys as a case study, we define the target schema and evaluate exact match tuple accuracy. Our experiments demonstrate F1 scores of 0.96, 0.95, 0.85, and 0.75 on datasets containing one to four tuples per sentence, and 0.85 on a randomly curated set. Ablation studies show that the allocation module is most critical, with inter-entity attention contributing more than intra-entity attention. Error analysis attributes the density-related performance decline mainly to semantic overlap and syntactic complexity, with upstream extraction errors more prominent under sparse supervision and allocation errors concentrated in structurally complex templates. This approach delivers precise, structured outputs suitable for downstream analysis and offers a domain-adaptable alternative to prompt-based large models when strict correctness is required.

Data-driven methods are transforming materials design by accelerating the discovery of new compounds and the optimization of existing systems. However, the progress of such approaches is often constrained by the scarcity of high-fidelity data from experiments and advanced simulations. Multi-fidelity (MF) learning has emerged as a powerful strategy to address this challenge by integrating information from diverse data sources that vary in accuracy and cost. In this review, we provide a systematic overview of the major methodologies for MF learning, including statistical and parametric models, machine learning models with fidelity features, correction-based models such as co-kriging, deep learning frameworks, and active learning frameworks. We discuss the strengths, limitations, and typical applications of each method in materials science, with illustrative examples spanning electronic structure modeling, alloy design, and interatomic potential development. Cross-cutting issues are also examined, including the bias-variance trade-off, data requirements for nested vs. non-nested designs, and computational scalability. Finally, we highlight outstanding challenges and outline emerging opportunities, such as physics-informed and generative MF models, standardized datasets, and integration with autonomous laboratories. Together, these perspectives define a roadmap for advancing MF learning as a core enabler of next-generation materials discovery.

Image recognition, classification, and analysis of large sets of high-resolution molecular images are time-consuming and labor-intensive, even for human experts, due to the lack of standardized approaches. In recent years, machine learning has emerged as a powerful tool for automating image data analysis in materials science. In this work, we developed a computer vision program for efficient object detection and instance segmentation, offering a fast alternative to manual molecular image analysis. By integrating You Only Look Once version 9 (YOLOv9) with an incremental learning strategy and hyperparameter optimization, the system enables accurate detection, classification, and segmentation of molecular species across diverse scanning tunneling microscopy datasets. Our results demonstrate robust performance and minimal forgetting rates across multiple molecular categories, enabling scalable and updatable surface image analysis workflows. We anticipate that computer vision methods will see increasing applications in image data analysis within the field of on-surface chemistry.