"Machine learning in chemistry and beyond" (ChE-651) seminar by Simon Batzner "Equivariant Interatomic Potentials"

Representations of atomistic structures for machine learning must transform predictably under the geometric transformations of 3D space, in particular rotation, translation, and reflection as well as permutation of atoms of the same species. This requirement has traditionally been satisfied by operating on descriptors of the atomistic geometry that remain invariant under the actions of the Euclidean group E(3), starting from the Lennard-Jones potential all the way to today's machine learning approaches. In this talk, I will discuss our efforts on generalizing the concept of invariant representations to the broader class of equivariant representations and will demonstrate how this leads to order-of-magnitude improvements in accuracy, sample efficiency and generalization. The talk will discuss in particular the Neural Equivariant Interatomic Potential (NequIP) [1] and show applications to a variety of molecular and materials systems including small molecules, water, a catalytic surface reaction, glass formation of a lithium phosphate, and Li diffusion in a superionic conductor. I will discuss then the discovery of a remarkably enhanced power-law behaviour of the loss of equivariant interatomic potentials. Finally, I will show our recent efforts to scale equivariant neural networks to large-scale systems through the strictly local equivariant interatomic potential Allegro [2] and show how this allows us to simulate a 100-million-atom Molecular Dynamics simulation. I will conclude by outlining our current theoretical understanding of equivariant interatomic potentials as well as open research  questions. 

 

[1] In print, Nature Communications, Preprint: arxiv.org/abs/2101.03164

[2] Preprint: arxiv.org/abs/2204.05249