Materialist

Summary: This study utilizes machine-learned interatomic potentials to reveal that layered halide ordering critically influences miscibility gaps and structural phase transitions in mixed-halide perovskites. These findings provide crucial insights for designing stable perovskite-based devices by guiding compositional choices. Why this paper? This work appli…

Summary: pyeCE is an open-source Python library leveraging machine learning within the embedded cluster expansion (eCE) framework to model complex alloy thermodynamics, particularly for high-entropy alloys. It provides a complete workflow for predicting finite-temperature properties and short-range order, enabling rapid screening of compositions. Why this p…

Summary: uFlowCSP is a novel MeanFlow-based generative model for crystal structure prediction that achieves significant speed improvements (5x-58x faster) while maintaining or exceeding the accuracy of current generative CSP methods. This model drastically reduces the computational cost of generating stable crystal structures, enhancing high-throughput mate…

Summary: This research develops a novel Constitutive State Space (CSS) model for path-dependent plasticity that, unlike traditional recurrent neural networks, is resolution-consistent and parallelizable, providing a more robust and efficient data-driven framework for material behavior modeling. Why this paper? This paper introduces a highly novel Constituti…

Summary: This research demonstrates that Fourier Neural Operators, when combined with wavelet encodings, can accurately learn and predict multiple deformation modes (eigenmodes) of acoustic waves in metamaterials, accelerating design simulations by three orders of magnitude. Why this paper? This paper presents a highly novel approach using wavelet-encoded F…

Summary: This work develops "MLIP Detective," an AI-driven framework that actively searches for and identifies hidden failure modes in machine-learning interatomic potentials by generating and screening physics-informed hypotheses. Why this paper? This paper introduces a novel, physics-informed framework for actively discovering failure modes in machine-lea…

Summary: This work introduces GALAXI, a scalable deep-learning framework that uses independent binary classifiers for robust and automated multiphase identification from powder X-ray diffraction patterns, significantly outperforming existing methods and handling experimental artifacts well. Why this paper? This paper presents GALAXI, a highly novel and scal…

Summary: This work introduces a novel framework where dynamic network materials can learn complex functions in situ through their own physical dynamics by applying forward and time-reversed adjoint drives, effectively acting as programmable matter and physical neural networks. Why this paper? This paper presents a highly novel and potentially transformative…

Summary: This work introduces two Hessian-derived data augmentation schemes, UniAug and ModeAug, that improve the accuracy of machine learning interatomic potentials (MLIPs) for tasks like vibrational analysis and transition state searches. These methods enhance MLIP training by implicitly incorporating Hessian information without requiring architectural mo…

Summary: This study develops a novel E(3)-equivariant graph neural network machine learning potential that incorporates voltage bias for accurate and efficient simulations of electrochemical interfaces under operational, open-system conditions. By decoupling zero-bias and bias-dependent energy contributions, this model overcomes limitations of traditional M…

Summary: This work introduces a generative multi-task deep learning model (GMTDL) to simultaneously optimize multiple, often conflicting, properties in Fe-based soft magnetic alloys, tackling challenges of vast composition space and data imbalance. The GMTDL not only predicts optimal properties but also suggests specific alloy compositions, establishing an…

Summary: This work proposes an orbital-free DFT-assisted machine-learned molecular dynamics method that efficiently simulates electric-field-driven ionic transport by combining a universal machine-learned potential for forces with OFDFT-derived environment-dependent atomic charges. Applied to Li3PS4, it enables cost-effective study of ionic migration and ch…