Materialist

Summary: This work introduces DeepSciReasoner, a novel deep learning framework for predicting structured scientific spectra, such as eDOS, phDOS, XANES, and Raman spectra, while explicitly incorporating physical constraints. This approach significantly improves prediction accuracy and maintains physical consistency compared to methods that treat spectra as…

Summary: This work introduces FlashSO(2), a novel framework that uses tailored low-precision emulation to significantly accelerate SO(2)-equivariant processing layers in graph neural networks, which are commonly used for machine-learned interatomic potentials and Hamiltonian models. FlashSO(2) achieves substantial speedups (up to 40x throughput) and memory…

Summary: This work introduces a novel "thermodynamic correction layer" for machine learning interatomic potentials, which adjusts phase free energies to match assessed references without requiring retraining or modifying the original potential. This method significantly improves the accuracy of transformation temperature predictions for various elements and…

Summary: MatOS is introduced as an agent operating system for materials research that leverages LLM agents to manage complex workflows, integrating diverse data, knowledge, and simulation tools. It features a scientific asset graph, adaptive workflows, and asset evolution, demonstrating superior performance in various research cases from materials screening…

Summary: This paper addresses the pervasive problem of erroneous data in literature-mined materials datasets, proposing a taxonomy of errors and a layered framework that combines physics-based rules, statistical checks, and model auditing for effective error detection. The framework significantly improves the accuracy of AI models trained on such data, catc…

Summary: This work computationally screens for substitutional solutes to enhance vacancy concentration in Li solid-solution anodes, identifying Ag, Ge, and Cd as promising candidates for improving critical current density in solid-state batteries, with machine learning streamlining the screening process. Why this paper? The paper directly addresses a critic…

Summary: This study combines machine-learning potentials with atomistic simulations and experiments to establish a framework for designing molecular coatings on sulfide solid electrolytes, identifying optimal chemistries and morphologies to enhance moisture stability while preserving ionic conductivity. Why this paper? This paper expertly combines machine-l…

Summary: This work creates a colossal, de-duplicated database of 14.6 million crystals, develops a state-of-the-art graph transformer (CGTNet) to predict properties, and critically re-evaluates the 'novelty' of inverse design versus the 'reach' of forward screening in materials discovery. Why this paper? This paper makes a foundational contribution by integ…

Summary: This study rigorously benchmarks fourteen universal machine-learned interatomic potentials for alloys, demonstrating that discrepancies between different DFT databases set a fundamental limit on measurable potential error and clarifying the true performance of MLIPs against a realistic ground truth. Why this paper? This paper offers a highly releva…

Summary: This paper introduces a modular framework using compositional neural networks and prior-amortized flow matching for multimodal X-ray and neutron scattering inversion. The method allows pretrained single-observation models to be combined at inference time, reducing ambiguity and improving structural characterization of thin films. Why this paper? Th…

Summary: This paper develops eSpinor, an SU(2)-equivariant graph neural network that propagates spinor features to accurately predict molecular magnetism, improving upon existing methods by directly addressing the spinor nature of electronic states. This approach enhances predictions for properties like zero-field-splitting and electron-spin correlations, c…

Summary: This paper proposes a symmetry-based few-shot fine-tuning strategy for a universal machine-learned interatomic potential (MatterSim 1M) to accelerate high-throughput lattice thermal conductivity calculations. The method achieves comparable accuracy to individual compound tuning while being 5-7 times faster for datasets sharing crystal symmetry. Why…