We present a framework for generating semantic embeddings of chemical elements to advance alloy inference and discovery. This framework leverages ElementBERT, a domain-specific BERT-based natural language processing model trained on 1.29 million abstracts of alloy-related scientific publications, to capture latent knowledge specific to alloys. These semantic embeddings serve as robust elemental descriptors, consistently outperforming traditional empirical descriptors across multiple downstream tasks, including predicting mechanical and transformation properties, classifying phase structures, and optimizing materials properties via Bayesian optimization. Applications to titanium alloys, high-entropy alloys, and shape memory alloys demonstrate up to 23% improvement in prediction accuracy. Notably, machine learning models based on semantic embeddings exhibit superior performance on newly published alloy data that are not included in the training corpus, indicating promising generalization capabilities for emerging alloy compositions. Our results show that ElementBERT surpasses general-purpose BERT variants by encoding specialized alloy knowledge. By bridging contextual insights from scientific literature with quantitative inference, our framework accelerates the discovery and optimization of advanced alloys, with potential for extension to other material classes.

The heat treatment process of carburized steel is crucial for determining its service performance potential. To overcome the efficiency and accuracy bottlenecks inherent in traditional development methods, this paper proposes a machine-learning-based approach for performance prediction and process optimization. This study constructed a high-dimensional database encompassing chemical composition, physical properties, and multi-stage heat treatment process parameters. With hardness gradient and coefficient of friction (COF) as prediction targets, the DT algorithm was selected through multi-model comparison. Through feature selection, an optimized model was developed, achieving prediction errors of 2.7% and 4.3% for hardness and COF, respectively. Furthermore, the SHAP method was used for model interpretability analysis, identifying critical process parameters such as quenching/tempering temperatures. The optimized process design, based on this approach, was validated through physical experiments: hardness prediction error was below 6%, and the predicted COF trend highly matched the measured results. The study demonstrates that this method can accurately predict performance and guide process optimization, exhibiting excellent potential for engineering applications.

Oxaliplatin (OXA), a critical third-generation platinum chemotherapeutic, is significantly limited by suboptimal loading capacity and encapsulation efficiency in nanoparticle-based delivery systems. To address this, we developed an integrated machine learning (ML) and multi-objective optimization (MOO) framework for the simultaneous prediction and exploration of loading efficiency (LE) and encapsulation efficiency (EE). Ensemble learning models, trained on a curated dataset of 70 experimentally characterized nanocarrier formulations, demonstrated robust predictive performance under stringent leave-one-paper-out (LOPO) cross-validation (R² = 0.87 for LE, R² = 0.84 for EE). The multi-objective exploration identified a Pareto-optimal design space, with predicted performance reaching up to 45.3% LE and 87.2% EE, and pinpointed a balanced knee-point formulation at 40.2% LE and 83.7% EE. Interpretable ML analysis revealed surface area-to-volume ratio, coordination site availability, and zeta potential as the primary physicochemical drivers of OXA loading and retention. Consequently, an optimized nanocarrier profile, characterized by a particle size of 90–110 nm, a negative surface charge, and a carboxylate-rich composition, was derived. This study establishes a predictive, data-driven computational framework that bridges the gap between single-objective prediction and the holistic design of high-performance nanocarriers, providing a rational blueprint for accelerating the development of more effective OXA-based nanotherapies for colorectal cancer.

Poly(lactic-co-glycolic acid) (PLGA) is widely used in biomedical implants due to its excellent biocompatibility and nontoxic degradation by products. Accurately predicting its degradation is crucial for implant stability. Traditional hydrolytic and autocatalytic models offer valuable insights but have limitations in describing real-world degradation kinetics. To address this problem, machine learning (ML) is applied to predict PLGA degradation under physiological conditions. Eight features, including molecular weight, lactide/glycolide ratio, pH, temperature, surface-area-to-volume ratio, geometry shape, degradation time, and lactide stereochemistry type, were used to predict degradation behavior. A dataset with 484 data points from 30 studies trained six ML models: k-nearest neighbors (KNN), random forest regression (RFR), eXtreme Gradient Boosting Regression (XGBR), AdaBoost regression (ABR), Cat Boosting regression (CBR), and gradient boosting regression (GBR). GBR was identified as the optimized model, with a determination coefficient of 0.90. SHapley Additive exPlanations (SHAP) analysis identified degradation time, molecular weight, and lactide/glycolide ratio as the most influential features. This study demonstrates the potential of ML in predicting PLGA degradation behavior, reducing the need for extensive lab tests and providing data support for accurate predictions of in vivo implant degradation.

In the artificial intelligence (AI) era, knowledge-based machine learning (ML) models can accelerate the design and modification of existing materials toward target properties. Therefore, an integrated framework combining ML with solid solution softening knowledge was introduced to design multielement tungsten alloys with strength–ductility synergy. The chemical short-range ordered (CSRO) body-centered cubic (BCC) phases, B32, B2, C11_b, and D0₃, are chosen as training dataset generated from high-throughput first-principle calculations. After meticulous feature engineering and hyperparameter optimization, AdaBoost and Gradient Boosting Regression algorithms are identified as the ML models for describing the thermodynamical and mechanical properties of multielement tungsten alloys. Spanning from binary to quaternary tungsten-based systems, W–Ta–Re is predicted by ML models and displays promising feature in overcoming the inherent strength-ductility trade-off. To validate the ML prediction, a W–10Ta–10Re alloy is synthesized via mechanical alloying and spark plasma sintering. TEM characterization confirms the formation of a uniform solid solution with a strong CSRO tendency. An exceptional compressive strength of 3586.7 MPa and a high work hardening rate are obtained, which surpass current tungsten alloy systems. This study not only establishes a reliable ML pathway for the design of high-performance tungsten alloys but also provides insights into the role of CSRO.

Lattice structures enable tailored mechanical properties for aerospace, biomedical, and energy applications, yet navigating their vast design space while meeting complex performance requirements remains challenging. Here, we introduce DeepSeek-Lattice-KG, an intelligent framework synergistically integrating a domain-adapted 14B-parameter large language models with knowledge graphs for lattice structure design. The model was fine-tuned on 2500 peer-reviewed articles and integrated with a knowledge graph containing 658,623 entities from 50,000 publications, enabling both generative reasoning and graph-based knowledge validation. Evaluation on 2100 expert-curated questions across six technical domains demonstrates 94.8% accuracy, surpassing DeepSeek-R1-670B (88.2%) and conventional fine-tuning approaches, while ensuring data privacy and computational efficiency. Through dynamic knowledge augmentation, the framework maintains 93% accuracy on 2025 emerging topics versus 77% for static systems, enabling real-time updates without retraining. Engineering case studies demonstrate the framework's capability to synthesize knowledge across thousands of publications for design exploration. This work establishes a paradigm that domain-specialized small models combined with structured knowledge provide an efficient alternative to large general-purpose systems for specialized engineering domains.

Materials science research increasingly benefits from the application of machine learning methods, yet encounters fundamental challenges from data scarcity, such as limited dataset sizes and severe distribution imbalance. In this paper, we propose a hybrid framework integrating attention pooling, multi-task learning, auxiliary learning, and classification-corrected regression. Using a 2D materials dataset as a case study, our approach demonstrates significantly enhanced prediction accuracy over the baseline crystal graph convolutional neural networks (CGCNN) method. Specifically, it reduces the mean absolute error for work function prediction from 0.312 to 0.240 eV, and for band gap from 0.301 to 0.230 eV. The framework also proves effective with other graph neural network methods such as atomistic line graph neural network (ALIGNN). These gains stem from the framework's ability to exploit underlying physical correlations between material properties and atomic structures. Through extensive experiments, we demonstrate that attention pooling serves as a generally effective component for diverse property prediction tasks, particularly with small datasets, which also offers the possibility of interpretability analysis through element attention weights. Our architecture enables seamless integration with various graph-based or other end-to-end deep learning models, presenting a computationally efficient and easily implementable solution for constrained datasets in materials science.

Potassium-conducting chalcogenides are attracting increasing attention as solid-state conducting materials for potassium-ion batteries. In this study, Inorganic Crystal Structure Database screening was performed for the first time to select new promising potassium-ion conductors among ternary and quaternary sulfides and selenides. From over 853 compounds studied, using a geometrical-topological analysis of free crystal space, we harvested 541 structures previously unreported as conductive materials in which periodic 1D, 2D, or 3D potassium migration maps were realized. Subsequently, the migration energy of the working cations was calculated for these structures using the bond valence site energy method, which allowed the selection of 37 structures with migration barriers below 1 eV, enabling 2D or 3D migration, and characterized by high crystal chemical stability. Kinetic Monte Carlo modeling then discovered nine structures with ionic conductivity exceeding 10⁻⁶ S⋅cm⁻¹ under standard conditions. For three ordered structures, K₆CdSe₄, K₃GeS₃, and K₄SnSe₄ with the highest ionic conductivities, DFT nudged elastic band simulations revealed 3D conductivity maps with migration barriers below 0.35 eV. According to ab initio molecular dynamics simulations, all three compounds exhibited ionic conductivity exceeding 10⁻⁵ S⋅cm⁻¹ at 300 K. Dependencies between migration barriers and parameters of conduction channels were also revealed in the series of isostructural compounds.

In high-power electronic packaging, Cu/diamond composites are attractive heat-spreader candidates owing to diamond's ultrahigh thermal conductivity. However, their practical thermal performance is often limited by the intrinsically low thermal boundary conductance (TBC) at Cu/diamond interfaces, which can be improved by introducing a manufacturable carbide interlayer. To elucidate the phonon-mediated interfacial thermal transport mechanisms affected by interlayer structures, the present work performs a systematic analysis based on density functional theory (DFT) and the diffuse mismatch model (DMM). The results reveal that TBC is jointly determined by phonon group velocity, PDOS overlap, transmission probability, and the effective participating frequency range, providing qualitative criteria for selecting carbide interlayers. The analysis further indicates that neglecting optical–phonon contributions and restricting phonon branch conversion can substantially underestimate TBC. The interfacial phonon transport is highly sensitive to interlayer thickness and composition: When thickness control is challenging, high-thermal–conductivity carbides (B₄C and WC) better retain interfacial thermal conductance (ITC) at large thickness. Residual B or Cr can provide an additional phonon-coupling pathway and further improve ITC. This work deepens the understanding of phonon-mediated interfacial thermal transport and provides mechanistic guidance for interlayer design in Cu/diamond heat spreaders.

Three-dimensional (3D) braided tungsten–copper (W–Cu) composites are attractive for combining high strength with good electrical performance, yet the high-dimensional design space makes manual optimization impractical. This work integrates high-fidelity finite element modeling (FEM) with an active learning (AL) scheme driven by a Gaussian process regression (GPR) surrogate and an uncertainty-aware acquisition policy. The surrogate provides accurate predictions together with quantified uncertainty, enabling data-efficient exploration of candidate braiding parameters while satisfying a prescribed electrical conductivity requirement. The framework rapidly identifies an optimal braided architecture with a peak yield strength of 1149 MPa, approximately a 30% improvement over the best design in the initial set. SHAP-based interpretability further reveals physically meaningful guidelines, with the in-plane spacing coefficient of the Z-direction fibers emerging as the dominant factor. These results demonstrate a fast interpretable route for optimizing metallic braided architectures and highlight the broader utility of FEM-coupled active learning for materials design.

Elemental segregation at grain boundaries (GBs) is widely recognized to influence the mechanical properties of structural materials. However, the intrinsic descriptors governing GB segregation have not been systematically clarified. Herein, first-principles calculations combined with interpretable machine learning analysis are used to identify the key factors governing elemental segregation at GBs in BCC Fe. A GB segregation database is constructed for 39 solutes commonly present in steels, including 33 metallic and 6 nonmetallic solutes. An interpretable machine learning framework is then developed to rank the intrinsic descriptors that govern segregation behavior. The results indicate that GB segregation of metallic solutes is primarily controlled by geometric features (Voronoi volume), whereas nonmetallic is more strongly governed by electronic features (Chemical bonding). Moreover, an overall contrasting trend is observed in the correlation between Voronoi volume and segregation energy for metallic and nonmetallic solutes. This study provides a new insight into GB engineering and the design of high-performance steels.

Machine-learned interatomic potentials have revolutionized molecular dynamics simulations by providing quantum–mechanical accuracy at empirical–potential speeds. The graphics processing unit molecular dynamics (GPUMD) package, featuring the highly efficient neuroevolution potential (NEP) framework, has emerged as a powerful tool in this domain. However, the complexity of force field development, active learning, and trajectory post-processing often requires extensive manual scripting, imposing a steep learning curve on new users. To address this, we present GPUMDkit, a comprehensive and user-friendly toolkit that streamlines the entire simulation workflow for GPUMD and NEP. GPUMDkit integrates a suite of essential functionalities, including format conversion, structure sampling, property calculation, and data visualization, accessible through both interactive and command-line interfaces. Its modular, extensible architecture ensures accessibility for users of all experience levels while allowing seamless integration of new features. By automating complex tasks and enhancing productivity, GPUMDkit substantially lowers the barrier to using GPUMD and NEP programs. This article describes the program architecture and demonstrates its capabilities through practical applications.

Machine learning and AI assistants are reshaping materials research; however, their day-to-day application in experimental and computational workflows is still inconsistent, limited by insufficient software documentation and software engineering practices, fragmented software engineering ecosystem and brittle integration between computational tools. Especially in the field of materials informatics, the lack of user-friendly tool interfaces has restricted the popularization of data-driven methods in daily scientific research. In this study, we present MatterMind, a user-friendly agentic interface that closes the loop between first-principles computation, generative structure design, and large language model (LLM) analysis. The platform unifies (i) plane-wave first-principles workflows through Vienna Ab initio Simulation Package (VASP), (ii) crystal generation via diffusion models, and (iii) LLM assistance for error diagnosis, result interpretation, and report drafting. With simple “button-click” operations or natural-language prompts, users can: (1) build, launch, and monitor VASP jobs with automated parsing and recovery; (2) sample candidate crystals using modern generative models (MatterGen); and (3) obtain LLM-guided summaries, comparisons, and next-step suggestions for screening decisions. We illustrate end-to-end case studies that couple crystal generation to DFT relaxation and LLM-assisted assessment, reducing scripting overhead and improving transparency and reuse from a software engineering perspective. It provides an intelligent scientific tool with comprehensive software documentation, enhancing the efficiency and reliability of scientific exploration in chemical and materials science research.

The general predictive approach established in our previous work Qiao et al., Materials Genome Engineering Advances. 2025;3(3):e70021. was employed to study the diffusion behavior of interstitial B and N atoms in FCC_CoNiV multi-principal element alloy (MPEA) based on sublattice preference, with comparative C data from prior work, to enrich the diffusion genome database of lightweight interstitial elements. Furthermore, we employed the Kabsch algorithm to describe the lattice distortion of the octahedra containing interstitial atoms quantitatively. The results show that the number of V atoms in the local octahedral environment exerts a different regulatory effect on the diffusion behavior of interstitial atoms B and N; that is, B and C exhibit a higher diffusion barrier when migrating into V-rich sites, whereas N exhibits such higher barrier when leaving these sites. Electron localization function (ELF) analysis shows the difference is due to the diverse bonding strengths between V atoms and interstitial atoms B, N, and C. Nonperiodic diffusion barrier waves and diffusion parameters were quantitatively predicted in detail. The fundamental understanding of interstitial diffusion mechanisms and quantitative characterization of the diffusion parameters of B, N, and C in FCC_CoNiV MPEA provide a benchmark and critical insights for tailoring alloy properties through interstitial engineering.