Physical Review Materials

X-ray absorption spectroscopy (XAS) is a powerful technique for probing local chemical environments. However, interpreting XAS spectra remains challenging, due to high computational costs and the need for domain expertise. To overcome these barriers, we introduce OmniXAS, a graph neural network framework that leverages transfer learning to directly predict XAS spectra from atomic structures. By capturing the shared spectral trends across the 3d transition metal family, OmniXAS learns a universal model that can be effectively fine-tuned on each specific element. OmniXAS achieves high predictive accuracy as demonstrated on the K-edge spectra of eight 3d transition metals (Ti–Cu), enabling real-time prediction of XAS spectra with minimal computational overhead.

Machine learning (ML) has become widely used in the development of interatomic potentials for molecular dynamics simulations. However, most ML potentials are still much slower than classical interatomic potentials and are usually trained with near equilibrium simulations in mind. Here, the authors have created computationally efficient Gaussian Approximation Potentials (GAP) for large-scale simulations in Cu, Al, and Ni. The models use a selection of low-dimensional descriptors and tabulation (tabGAP), achieving orders-of-magnitude speed up compared to standard GAP. Furthermore, the models include external repulsive pair interactions, and the training databases have been designed with extra attention to far-from equilibrium simulations.

Zeolites, renowned for their unique porous structures, are widely used in applications like gas separation and ion-exchange. However, discovering zeolite-like materials with high potassium ion (K⁺) selectivity has traditionally been time-consuming and expensive. In this study, the authors propose a data-driven paradigm combining Artificial Intelligence and Density Functional Theory to efficiently identify new zeolite-like structures with enhanced K⁺ selectivity. Their framework identifies a novel porous material with the highest K⁺ adsorption capacity to date. This approach paves the way for more efficient materials discovery and the development of advanced ion-selective membranes and energy storage materials.

Interstitial diffusion plays a crucial role in phase stability and irradiation response in concentrated solid-solution alloys (CSAs). Here, the authors integrate machine learning (ML) with kinetic Monte Carlo (kMC) to investigate interstitial-mediated diffusion in Fe–Ni CSAs. Their ML model—trained on migration barriers obtained via nudged elastic band calculations with an EAM potential—predicts barriers on-the-fly during kMC simulations. They demonstrate that sluggish diffusion emerges when reductions in the tracer correlation factor outweigh increases in jump frequency. Barrier differences among correlated migration paths form a ‘route selector’ that favors slower-diffusing species, amplifying correlation effects and suppressing jump-frequency gains. These findings apply broadly to other CSA systems.

Two-dimensional (2D) materials and their heterostructures continue to attract significant research interest owing to their unique physical properties and promising technological applications. The weak van der Waals interactions between individual layers give rise to atomically sharp interfaces and intricate moire patterns, often involving unit cells with thousands of atoms. These structural complexities present major challenges for $ab$ $initio$ simulations. In this study, the authors systematically evaluate the performance of several universal machine learning interatomic potentials (MLIPs), employing a range of structural and electronic similarity metrics tailored to 2D van der Waals heterostructures. The results indicate that the most accurate MLIPs reach a level of precision comparable to the intrinsic uncertainty of density functional theory due to the choice of exchange correlation functional.

This study completes a long-standing effort to understand experimental measurements of self-diffusion in non-magnetic body-centered cubic (bcc) metals and proposes a complete workflow to estimate it. Over the past decade, consensus has emerged attributing anomalous non-Arrhenius behavior to anharmonicity of atomic interactions. This work advances two issues: (i) free-energy contributions from small vacancy clusters are quantitatively accounted for, showing di-vacancy effects to be negligible; and (ii) quantitative agreement with experiment requires interatomic interactions beyond standard exchange-correlation functionals, such as recent meta-Generalized Gradient Approximations (meta-GGAs). The developed workflow enables automated, systematic investigations of diffusion and complex energetic landscapes, providing a foundation for predictive phase stability by connecting atomic vibrations with defect-driven phase diagrams.

Spintronics is next-generation electronics that use electron spin for digital information. A key challenge is finding materials with a strong spin Hall effect to generate spin currents, but conventional screening is slow and impractical at scale. The authors propose a multi-modal transformer that combines material fingerprints from real-space crystal structures and reciprocal-space electronic structures, achieving higher prediction accuracy than single-modal methods. The model captures spin-orbit coupling and crystal symmetry effectively, identifying a material with large spin Hall conductivity. This scalable approach accelerates the discovery of spintronic materials, enabling energy-efficient memory, logic devices, and quantum technologies.

The authors present decoratypes, a new extensible crystal taxonomy that offers a more granular lens for structure–property relationships by classifying materials by hierarchical property-site mappings. For example, the framework generalizes anti-structures into polaritypes, a subclass of decoratype. They demonstrate its utility by building a polaritype-based active learning workflow for discovering ferroelectric and hyperferroelectric materials. Our approach identified six novel candidates, including three strain-activated ferroelectrics and three strain-activated hyperferroelectrics. These findings highlight how decoratypes can provide a novel perspective on the search for functional materials in underexplored chemical spaces.

The authors propose a machine-learning approach that integrates nonequilibrium molecular dynamics under a constant electric field with equivariant neural network models to evaluate ionic conductivity in solid electrolytes. In this approach, Born effective charges are predicted by an equivariant graph neural network and used to describe field-induced forces and current density. These forces are combined with unperturbed forces from an equivariant neural network potential. Applied to the representative solid electrolyte Li₁₀GeP₂S₁₂, the method achieves first-principles accuracy at a fraction of the computational cost. It also captures charge fluctuations in complex materials, providing new physical insights into ionic dynamics.

Quaternary mixed-metal chalcohalides (M(II)${}_2$M(III)Ch${}_2$X${}_3$) combine the beneficial optoelectronic properties of halide perovskites with the stability of chalcogenides, making them as promising photovoltaic absorbers. By integrating density functional theory with random forest regression and Shapley additive explanations, the authors identified compositional trends and design rules across 54 lead-free and lead-containing compounds. The results reveal that electron acceptor sites (Ch and X) are crucial in shaping overall material behaviour, while donor sites (M(II) and M(III)) enable targeted tuning of properties for specific applications. These insights provide guidelines for designing mixed-metal chalcohalides for photovoltaic and optoelectronic applications.

The authors have developed a full-space inverse design strategy powered by machine-learning force fields to accelerate the discovery of superhard materials. Using the D${}^3$REAM framework, they identify two carbon allotropes and three B–C–N compounds with exceptional mechanical performance, including elastic moduli that surpass diamond along certain crystallographic directions. By integrating active learning, global optimization, and first-principles validation, their approach efficiently navigates vast configurational spaces and reveals that incorporating up to 40% B–N enables structurally diverse and superhard phases. This work demonstrates a robust pathway toward data-driven exploration and design of next-generation ultrahard materials.

The authors present the Organic Polymer Energy Conversion Materials (OPECM) Dataset, comprising 3,225 polymers systematically categorized into three key areas: organic semiconductors, photovoltaics, and dielectric materials. Leveraging this dataset, they developed polymer unit-guided regression models to accurately predict essential properties: electron/hole mobility, power conversion efficiency, and dielectric constant, while identifying pivotal polymer units that govern material performance. Moreover, using the SISSO method, the authors constructed interpretable symbolic regression models that uncover critical molecular features and their functional roles in energy conversion. This study provides both data and insights to accelerate the rational design of multi-functional energy materials.

Defects in amorphous materials​ are site inequivalent, and each atomic site​ exhibits multiple metastable configurations, making DFT-based​ studies of defect properties computationally expensive. The authors propose a data-efficient approach based on machine learning interatomic potential (MLIP). The training set is built from locally perturbed defect configurations​ at only a few atomic sites. Despite this small dataset, the trained MLIP successfully predicts defect structures and formation energy distributions for all atomic sites across the amorphous systems. A dual-model cross-validation strategy further identifies and refines inaccurate predictions, improving the overall prediction accuracy. Their approach enables efficient, large-scale statistical analysis of defects in amorphous systems.

The shock loading responses of Sn have attracted significant interest. Although atomistic simulations have been useful for uncovering nano-scale mechanisms behind experimental observations, exiting potentials of Sn lack sufficient accuracy especially for predicting its complex high-pressure phase transitions. To overcome this challenge, the authors construct DP-SCAN-S, an machine learning potential trained on comprehensive DFT data spanning an extensive thermodynamic range from 0–100 GPa pressure and 0–5000 K temperature. It accurately reproduces DFT-derived basic properties, experimental melting curves, solid-solid phase boundaries, and shock Hugoniot results. This demonstrates the model’s potential to bridge ab initio precision with large-scale dynamic simulations.

Machine-learned force fields (MLFFs) can bring first-principles accuracy to finite-temperature molecular dynamics for materials simulations. Here, the authors use MLFFs, trained on-the-fly using only ground-state structures, to predict phase transitions in the prototypical ferroelectrics BaTiO${}_3$, PbTiO${}_3$, LiNbO${}_3$, and BiFeO${}_3$. Order parameter discontinuities, mixed order–disorder and displacive character, and space groups are correctly predicted, while exact transition temperatures are functional-dependent. This demonstrates both the promise and current limitations of MLFFs for simulating the thermal evolution of materials, where long-range electrostatic interactions and collective lattice instabilities are imperative to the physics and phase transitions. Ferroelectrics thus constitute a stringent testbed for MLFFs.

Macroscopic dynamical descriptions are essential for understanding and controlling complex material behavior, yet deriving them from microscopic simulations remains computationally prohibitive for spatially extended stochastic systems. To address this challenge, the authors propose a machine-learning framework that learns large-scale macroscopic dynamics using only small-system simulations. The method uses a partial evolution scheme to generate training data within local patches, a tailored loss to learn the macroscopic dynamics, and a hierarchical upsampling strategy to efficiently construct large-system configurations. Across stochastic PDEs, lattice spin models, and an NbMoTa alloy system, the framework demonstrates high accuracy, robustness, and computational efficiency.

A widely used thermoplastic, poly(ethylene terephthalate) (PET), faces increasing environmental and regulatory pressure, motivating the search for viable alternatives. Here, the authors present an AI-driven polymer design pipeline implemented using the PolymRize. The framework combines virtual forward synthesis with machine learning to generate PET-replacement copolymers. Inspired by the esterification route of PET synthesis, more than 12,000 candidate polymers were systematically constructed from TSCA-listed monomers. ML models predicted glass transition temperature, bandgap, and crystallization tendency to enable multi-objective screening. The approach rediscovered known PET alternatives and identified previously unknown candidates, several of which were synthesized and experimentally validated.

Understanding the atomic evolution from cluster to nanocrystal has long been a challenge in nanoscience. Here, an accurate machine learning potential (MLP) of elemental Pd was developed. The large-scale capacity of this MLP affords long-time simulated annealing for a cross-scale study of Pd${}_n$ nanostructures ($n$=12 - 21856), revealing a continuous transition from discrete clusters to bulk-like nanocrystals and the critical size at which the transition occurs. This study paves the way for studies on other clusters.

Machine learning potentials (MLPs) have been widely used in predicting the thermal transport properties of $\beta$-Ga${}_2$O${}_3$. However, there are few reports on MLPs specifically considering complex defects in $\beta$-Ga${}_2$O${}_3$. In this study, the authors trained a deep neural network MLP model to quantify how different intrinsic point defects suppress the thermal conductivity of $\beta$-Ga${}_2$O${}_3$. These results provide atomic-level insight into thermal transport in defective $\beta$-Ga${}_2$O${}_3$, offering guidance for thermal-management strategies and establishing a general workflow for investigating thermal physics in complex semiconductor materials.

A century after the discovery of superconductivity, the search for new superconductors still relies largely on trial and error, challenging researchers to identify the universal design principles that govern why certain materials superconduct at higher temperatures. Here, the authors introduce an interpretable, data-driven approach that pairs Random Forest screening with SISSO symbolic regression to reveal fundamental “material genes” governing Tc in conventional BCS superconductors. Unlike black-box predictors, this study reveals physically meaningful relationships that provide actionable guideline for high Tc: a near half-filled d-orbital per atom combined with moderate heterogeneity in unfilled orbitals.

Flash events in tetragonal zirconia ceramics trigger unusual mass transport and rapid sintering that cannot be explained by Joule heating alone. To uncover the underlying atomic mechanisms beyond regular machine learning interatomic potentials, the authors developed a machine learning model that also predicts the dynamic response of ions via Born effective charges. These molecular dynamics simulations reveal that an applied electric field significantly enhances the diffusivity of oxygen ions, particularly in the presence of oxygen vacancies. This represents a crucial milestone for understanding defect-mediated ion transport in high-strength ceramics under external electric fields.

In this study, the authors investigate the real-time atomic dynamics and local structural fluctuations in GeTe by coupling density functional theory with large-scale molecular dynamics simulations driven by neuroevolution potentials. Their findings demonstrate a displacive ferroelectric phase transition near the tricritical point. Crucially, the computed longitudinal current correlation function lacks a quasielastic peak, effectively ruling out thermally activated discrete jumps. Furthermore, time-resolved orbital analyses reveal a femtosecond ‘seesaw’ charge transfer driven by the pseudo-Jahn-Teller effect. These results support a ‘macro-ordered yet micro-disordered’ displacive paradigm.

Halide solid electrolytes such as Li₃YCl₆ and Li₃YBr₆ are promising candidates for safer, denser all-solid-state batteries, but their vast alloying space is hard to explore experimentally. Using the universal machine-learning potential PET-MAD, validated against a fine-tuned model, we map how halogen and metal substitution shape structure, phase stability and Li⁺ conductivity in Li₃MX₆ compounds. Halide mixing turns out to act through two competing levers: lattice contraction hinders Li motion, while shorter metal–halide bonds free up diffusion pathways, largely canceling out. Metal-site alloying, in contrast, tunes phase stability and cost with little penalty on conductivity: a practical design principle for optimizing these materials.