Existing machine learning-assisted alloy design studies often treat models as “black boxes”, lacking interpretability and thus failing to translate predictions into physically meaningful design guidelines. Herein, we propose an integrated strategy combining feature engineering, autonomous model optimization, and interpretability analysis for the efficient design of (CuNiMn)-X alloys, where X denotes Al, Ti, Cr, and Fe alloying elements, with strength-ductility synergy. Through feature cleaning and grid-based hyperparameter optimization, prediction accuracies of 92.6%, 89.2%, and 88.1% are achieved for the microstructure, compressive strength, and fracture strain models, respectively. Experimental validation of the optimized alloy shows errors of -5.0% for compressive strength and -3.2% for fracture strain. SHapley Additive exPlanations analysis reveals the underlying physical mechanisms: valence electron concentration (M-E10) governs phase selection; mean atomic radius (M-A4) and its variance dominate compressive strength via lattice distortion; variances of fusion enthalpy, covalent radius, and shear modulus collectively control compressive strain through triple homogeneity in thermodynamics, structure, and elasticity. This analysis unveils a mirror-image symmetry between high strength and high ductility in feature space, elucidating the strength-ductility trade-off. Guided by predictions, the (CuNiMn)-Al16Cr16Fe16 alloy is experimentally validated, exhibiting a compressive strength of 2,542 MPa and fracture strain of 15.8%. The microstructure consists of Cr-rich body-centered cubic dendrite arms and a Cu-Ni-Fe-Al-rich face-centered cubic interdendritic network, forming a hard-and-tough dual-phase architecture. This closed-loop design strategy provides a new route for the efficient discovery of CuNiMn-based multi-principal element copper alloys and offers guidance for the broader design of complex multi-principal element alloys.

This paper presents an automated strategy for constructing machine-learning potential (MLP) datasets through a physics-strengthened point-sampling scheme called physics-strengthened machine-learning potential (PhyMLP). By integrating heterogeneous data sources - including the Rose equation of state, experimental pressure-volume (P-V) measurements, traditional empirical potentials, and first-principles calculations - this method effectively circumvents computational bottlenecks and reduces the reliance on the exhaustive density functional theory (DFT) computations inherent in conventional training-set construction. The PhyMLP employs an adaptive, physics-guided sampling strategy that leverages intrinsic material responses and requires only a limited number of critical DFT calculations to efficiently characterize the potential energy surface. Using body-centered cubic tungsten as a benchmark system, the moment tensor potential trained on the PhyMLP-generated dataset exhibited exceptional predictive accuracy across a wide spectrum of material properties, ranging from fundamental physical constants to complex defect energetics and kinetic behavior. Ultimately, this study established a systematic and computationally efficient paradigm for developing accurate transferable MLPs, thereby offering a robust framework for large-scale atomistic simulations and advanced material modeling.

To address the reliance on trial-and-error methods and the prolonged development cycles inherent in the composition and process design of novel cast aluminum alloys, this study constructed a multi-source feature system integrating alloy composition, physicochemical properties of elements, testing conditions, and process parameters. Employing a three-step feature selection method, a prediction model for high-temperature ultimate tensile strength (UTS) was established with a test set the coefficient of determination of 0.881 and an mean absolute error of 22.639 MPa. Based on this model, a synergistic design of the alloy composition and heat treatment process was conducted by coupling the model with a genetic algorithm (GA), and four novel cast aluminum alloys were experimentally validated. The experimental results indicate that the designed alloys exhibit enhanced elevated-temperature strength compared with the commercial reference alloys. Notably, the ZL-2 alloy demonstrated optimal performance, achieving a UTS of 214.2 MPa when tested at 300 °C after holding at 300 °C for 1 h. Furthermore, SHapley Additive exPlanations (SHAP) analysis revealed significant nonlinear interactions among alloy composition, testing conditions, and process parameters in determining tensile strength. Multiscale microstructural characterization reveals that the exceptional elevated-temperature strength of ZL-2 stems from the synergistic effects of nanoscale Al₂₀Cu₂Mn₃ and Al₂CuMg precipitates, and micron-scale Al₃Ti-containing intermetallics. This study demonstrates the application potential of data-driven methods in the composition-process synergistic design of heat-resistant cast aluminum alloys.

Extracting reliable knowledge from unstructured materials literature remains a central bottleneck for data-driven and AI-enabled materials discovery. Large language models (LLMs) are reshaping this task by integrating multimodal document parsing, ontology-guided semantic grounding, structured extraction, and agentic verification into increasingly unified workflows. This review analyzes these developments through a Perception–Cognition–Action lens. At the perception layer, we examine how scientific document parsers, multimodal LLMs, table and chart readers, and optical chemical-structure-recognition systems convert visually rich papers into computable evidence. At the cognition layer, we discuss how ontologies and knowledge graphs constrain LLM outputs, support entity alignment, and reduce semantic ambiguity. At the action layer, we compare schema-based extraction, schema-free discovery, and agentic extraction as a control–coverage–autonomy spectrum rather than a simple succession of tools. We further argue that reliability is the decisive criterion for large-scale deployment, and synthesize failure modes, layered defenses, and evaluation protocols that connect source grounding, ontology constraints, physical verification, and human-in-the-loop review. By distinguishing demonstrated extraction capabilities from more speculative AI-scientist and self-driving-laboratory visions, this review provides a comparative and risk-aware account of how LLM-driven systems can produce evidence-linked, physically meaningful, and reusable materials knowledge.

High-entropy alloys (HEAs) exhibit exceptional stability in extreme environments, yet their expansive design space presents a “curse of dimensionality” for traditional discovery methods. While machine learning (ML) offers a data-driven paradigm for material screening, the scarcity of experimental data often results in overfitting and limited physical interpretability. To address these challenges, this study proposes a hybrid physics-informed machine learning (Hybrid PIML) framework for accelerated hardness prediction. By integrating classical solid solution strengthening theory with a residual learning artificial neural network (ANN), the model explicitly embeds the physical coupling of shear modulus and lattice distortion (G·δ_r^2/3) as prior knowledge. This approach ensures predictions adhere to metallurgical principles while significantly outperforming benchmark algorithms, achieving a coefficient of determination (R²) of 0.976 and reducing the root mean square error (RMSE) by approximately 43%. SHapley Additive exPlanations (SHAP) analysis confirms that physics-enhanced features dominate the decision-making process, validating the model’s internalization of strengthening mechanisms. Furthermore, the research elucidates a phase-dependent non-linear correlation between hardness and yield strength, correcting the failure of the classical Tabor formula in work-hardening face-centered cubic (FCC) alloys. Finally, a high-throughput virtual screening funnel based on this framework successfully identified optimized non-equiatomic candidates within the refractory Co-Cr-Ti-Mo-W system. This work establishes a precise, physically consistent pathway for inverse material design under data-constrained conditions.

The electron localization function (ELF) provides a chemically interpretable representation of bonding and interstitial electron localization, but generating volumetric ELF fields with density functional theory (DFT) is computationally expensive for large candidate libraries. We present ELFNet, a structure-to-field model that predicts DFT-derived ELF grids from computationally inexpensive superposed atomic density (SAD) inputs for high-pressure chemical-template screening. To train the model, we constructed a pressure-targeted Perdew–Burke–Ernzerhof (PBE) ELF dataset containing 326,009 converged unary and binary prototype structures spanning 89 elements and covering pressures from approximately -50 to 500 GPa. ELFNet uses periodic full-grid convolutions and a composite objective designed to recover sparse high-ELF features. We evaluated the model on 50,000 metal-only templates derived from known high-pressure superhydride prototypes. Predicted fields captured chemically meaningful differences in the spatial topology of interstitial localization, while global maximum ELF values and maximum-location errors provided complementary measures of agreement with DFT references. On the reported hardware, the full inference-through-write workflow reduced summed per-structure elapsed time by approximately 350-fold relative to fixed-geometry Vienna Ab initio Simulation Package (VASP) ELF generation. These results establish SAD-to-ELF prediction as a low-cost front-end screening method for prioritizing candidate metal templates before downstream first-principles validation.

To realize the exploration of unknown diagrams directly from measurable physical/chemical properties, we improve the property derivatives-based machine learning (ML) framework of phase boundary prediction to overcome prior limitations in signal enhancement, labeling efficiency, uncertainty quantification, and potential applicability to multicomponent composition spaces. An automated phase-region labeling approach combining multiple activity derivatives is introduced to amplify the differentiability signal near phase boundaries, enabling precise identification of distinct phase regions. Three uncertainty sampling (US) strategies, i.e., least confidence (LC), margin sampling (MS), and entropy-based acquisition (EA), are employed to guide query selection in active learning. MS generally promoted a more uniform distribution of queries along phase boundaries and reduced redundant sampling near triple junctions, thereby facilitating efficient convergence. The truncated ML phase boundaries based on multi-layer perceptron + MS achieved Macro-F1 scores of 0.9994 and 0.9814 using only 120 and 186 labeled samples for the Pd-Pt and Au-Ag-Ge systems, respectively. The mean boundary displacements of ML-predicted phase boundaries against the Thermo-Calc-calculated results were 0.0059 for the Pd-Pt system and 0.0778 for the Au-Ag-Ge system.

The rational design of electrocatalysts for multi-electron reaction networks hinges on navigating vast, multidimensional surface-environment spaces defined by coupled variations in composition, coordination topology, and adsorbate configurations. Taking electrochemical carbon dioxide reduction to ethanol as a model challenge, traditional screening methodologies struggle to efficiently represent and search these combinatorial landscapes. Here, we present a reaction-constrained generative materials informatics framework anchored by late-stage oxygenated intermediates (*OCH₂CH₃). A fine-tuned generative Transformer model efficiently sampled 1,000,000 adsorption configurations. To efficiently navigate this massive space, adsorption configurations were compressed into latent topological representations via oxygen-centered local node embeddings derived from the universal models for atoms (UMA) platform. Cluster-guided spatial navigation based on these UMA-derived representations achieved a 15-fold enrichment in positive-candidate discovery over naive representative sampling. By jointly applying multi-adsorbate thermodynamic constraints (*OCH₂CH₃, *CO, and *H) within the CO₂RR network, the workflow down-selected 75 promising material candidates. Density functional theory (DFT) calculations benchmarked representative systems, validating the screening-level fidelity of UMA-predicted binding energies. Furthermore, subgroup discovery extracted interpretable, physically intuitive elemental rules governing candidate enrichment. CO-CO coupling calculations further identified Ga-Pt and Al-Cu-Pd as kinetically favored over Cu. This framework successfully recovers 13 reported CO₂RR-related materials while uncovering 62 unexplored compositional domains, providing a transferable paradigm for reaction-constrained generative surface informatics.