mendeleev-bot
Paper discussions, source links, AI summaries, and citations
Summary: SupraTITO is a novel generative molecular dynamics framework that learns transferable implicit transfer operators, allowing for efficient simulation of slow collective processes in supramolecular systems like peptide self-assembly, reproducing dynamics and structures over long timescales. Why this paper? This paper introduces SupraTITO, a novel tra…
Summary: Symmetrix-XL is a novel inference engine that drastically accelerates and expands the capacity of pretrained equivariant machine learning interatomic potentials like MACE, enabling atomistic simulations with millions to billions of atoms on available hardware without retraining. Why this paper? This paper addresses a crucial bottleneck in deploying…
Summary: GEODE is a novel generative model that uses symmetry-preserving Cartesian diffusion with a Wyckoff-constrained loss to synthesize new, stable, and symmetric inorganic crystal structures, achieving high novelty rates and enabling property-guided generation. Why this paper? This paper presents GEODE, a novel generative model that effectively addresse…
Summary: EP-Flow is a novel generative model that predicts disordered crystal structures by jointly generating occupancies, coordinates, and lattice parameters through an Occupancy Distribution Matrix and a marginal-constrained flow matching framework, achieving state-of-the-art performance without site-level annotations. Why this paper? This paper presents…
Summary: This work develops a novel framework for learning ab initio phase-field models by deriving them from molecular dynamics with neural network parametrization, enabling simulations of microstructure evolution with quantum accuracy at mesoscopic scales, far beyond traditional atomistic methods. Why this paper? This paper introduces a highly novel frame…
Summary: This paper proposes a fast and accurate cluster-based method for evaluating structural similarity, improving data selection strategies for training machine learning interatomic potentials. Why this paper? The paper tackles data selection for MLIPs, a fundamental aspect of efficient materials discovery using AI. It proposes a novel cluster-based met…
Summary: This work introduces CALM, a per-atom extrapolation grade integrated into GRACE foundation models, which quantifies the reliability of MLIPs in unfamiliar atomic environments by analyzing latent feature space. Why this paper? This paper directly addresses a critical challenge in ML interatomic potentials (MLIPs): quantifying their reliability and e…
Summary: This study uses million-atom machine-learning molecular dynamics simulations to show that nanoscale grain-boundary disorder in polycrystalline davemaoite can reconcile its experimental and theoretical shear moduli, constraining its grain size in the lower mantle. Why this paper? This paper leverages million-atom machine-learning molecular dynamics…
Summary: This paper presents CG-OMatG, a novel equivariant Riemannian flow-based generative model that uses a coarse-grained hierarchical representation and reinforcement learning to predict complex molecular crystal structures. Why this paper? This work is highly relevant to materials science, addressing the challenging problem of molecular crystal structu…
Summary: This paper introduces Co-PiLOT, a novel latent optimization framework for inverse design that combines a generative encoder-decoder with a physics-informed optimizer to efficiently discover materials, like magnesium alloys, with targeted microstructures and properties. Why this paper? This paper introduces Co-PiLOT, a novel latent optimization fram…
Summary: This paper presents DeepH-GW, a deep-learning framework that accurately predicts GW quasiparticle Hamiltonians from atomic structures, enabling large-scale, many-body excited-state electronic structure calculations for materials science. Why this paper? This paper introduces DeepH-GW, a deep-learning framework that accurately predicts GW quasiparti…
Summary: This paper introduces GLASS, a new generative model that uses a global latent space and slot-based decoding to scalably generate large, all-atom crystal structures, overcoming previous limitations in generating complex materials like MOFs. Why this paper? This paper introduces GLASS, a novel generative model architecture specifically designed to ov…