(159d) Generative AI for Materials Modeling

All data are noisy. This is especially true in condensed matter physics, chemistry, and materials science, as Richard Feynman famously put it, "Everything that living things do can be understood in terms of the jiggling and wiggling of atoms". The main challenge to model big scientific data is that they form high dimensional distributions that are intractable. Inspired by recent advancement in generative AI, we propose to adopt probabilistic approaches to high dimensional probability distributions that are tractable and learnable through statistical learning procedures. The efficacy of our approach is demonstrated through a few distinctive case studies, including a denoiser that identifies both ordered/crystalline and disordered/defect atomic structures with state-of-the-art accuracy, and a generative model for molecular and crystal structures. Finally, we propose score dynamics, a general data-driven probabilistic
framework to accelerate both fine-level atomic and coarse-level mesoscale simulations of materials, and showcase its applications in barrier-crossing reactions of dihedral angle dynamics in organic molecules, as well as nucleation and growth reactions of liquids.