Skip to yearly menu bar Skip to main content


Poster
in
Workshop: Geometry-grounded Representation Learning and Generative Modeling

An Equivariant Flow Matching Framework for Learning Molecular Crystallization

Shengchao Liu · Divin Yan · Hongyu Guo · Anima Anandkumar

Keywords: [ Geometry ] [ SDE ] [ crystallization ] [ flow matching ] [ physics ] [ Molecule ]


Abstract:

Molecular crystallization is the transformation of molecules from weakly-correlated structures to strongly-correlated structures, \textit{e.g.}, liquid water freezing into solid ice. It is essential in determining the functionalities of compounds across various fields, from pharmaceuticals to materials science. However, existing simulation methods for crystallization, which primarily rely on numerical techniques, can be exceedingly time-consuming. In this work, we build up a novel ML paradigm for solving the crystallization problem, \textit{i.e.}, learning a distribution mapping from weakly-correlated structures to strongly-correlated structures. First, we construct two datasets, coined COD-Cluster17, for benchmarking and also design two packing matching metrics for crystallization evaluation. Second, we propose CrystalFlow, an SE(3)-equivariant Flow Matching framework, for modeling the crystallization trajectories of molecular clusters.

Chat is not available.