Materialist

Summary: This work introduces OpenGEM26, a large dataset of organic molecule conformations, and uses it to train GPTFF-mol, a graph neural network potential that outperforms existing models in accuracy for force and energy predictions, particularly for sulfur- and chlorine-containing molecules. This provides a robust tool for efficient molecular simulations…

Summary: This study employs a tabular foundation model with physics-informed sampling to accurately predict the frequency-dependent dielectric function and derived optical properties of multi-component 2D Mo-W-S-Se-Te TMD alloys. The model demonstrates strong zero-shot generalization across various alloy compositions, accelerating optoelectronic materials d…

Summary: MatDiffract is presented as a material-informed automated analysis platform for X-ray powder diffraction (XRPD) data, combining simulated patterns, multi-scale ML features, and Rietveld refinement. It achieves high accuracy in phase identification and quantification, providing an end-to-end solution to accelerate materials characterization in auton…

Summary: This research explores the inverse problem of reconstructing atomic structures from compact, symmetry-based descriptors, showing that accurate geometries can be recovered even from formally incomplete representations. This framework provides insight into descriptor degeneracies and the structural implications of descriptor perturbations. Why this p…

Summary: This work introduces MOF-Sleuth, a tool-grounded LLM agent for explainable and fine-grained auditing of MOF CIFs, using a two-module architecture and reward-guided reinforcement learning. It improves detection and provides evidence-grounded explanations for chemical and structural errors, critical for reliable MOF databases. Why this paper? This pa…

Summary: This paper introduces Rem3Di, a representation-learning framework that repurposes latent features from atomistic foundation models to create transferable, smooth, and chiral 3D molecular descriptors. These descriptors enable accurate property prediction and virtual screening, including distinguishing enantiomers, without relying on experimental lab…

Summary: This work presents a high-throughput computational framework integrating MLIPs, CALPHAD, and phase-field simulations, using analytic derivatives, to design spinodal alloys and predict microstructure evolution in complex composition spaces. The approach is demonstrated by investigating microstructure stability in the Hf-Nb-Ti-V quaternary system. Wh…

Summary: This work uses a machine learning interatomic potential and high-throughput screening to identify structural features governing negative thermal expansion (NTE) in MOFs, creating a database of phonon properties and validating design rules experimentally to enable the precise engineering of NTE materials. Why this paper? This paper offers a highly r…

Summary: ATLAS is a new AI-driven sampler using a diffusion model and equivariant GNN to efficiently generate Boltzmann-distributed amorphous material structures, significantly accelerating the sampling and inverse design of metallic glasses and other amorphous systems, even integrating with LLMs for autonomous material search. Why this paper? This paper pr…

Summary: This study uses machine learning potentials to perform molecular dynamics simulations of MgH2 dehydrogenation, revealing a novel subsurface H2 formation mechanism and identifying electronic descriptors (Miedema electron density) for optimizing dopants to enhance hydrogen release. This provides a robust framework for rational catalyst design for hyd…

Summary: This work demonstrates that engineered disorder in kirigami metamaterials can decouple extension and shear, enabling programmable anisotropic stiffness unavailable with periodic patterns. It achieves this by navigating the design space with a geometry-aware graph neural network coupled to a genetic algorithm for inverse design. Why this paper? This…

Summary: This work introduces STEP (Spin Tensor Equivariant Potential), a novel magnetic machine-learning interatomic potential that treats vector magnetic moments as continuous geometric degrees of freedom and embeds them in an equivariant representation. STEP achieves high data efficiency and accuracy in modeling spin-lattice coupling, magnetic excitation…