Atomic-scale modeling of chemical systems like molecules, catalysts, or solutions offers myriad opportunities for the application of machine learning. A central challenge in such efforts is building model architectures that respect the symmetries of the underlying physics, namely permutation of atoms, rotation, translation, and inversion. This is particularly important in the promising area of machine learning interatomic potentials (MLIPs), which use small numbers of computationally expensive quantum chemistry calculations to enable simulating larger length- and time-scales than those quantum methods can achieve alone.<br/> <br/>This talk discusses Allegro and NequIP, our “E(3)-equivariant” machine learning architectures for MLIPs.  Most existing MLIPs respect the invariance of the true potential energy to symmetry transforms like rotations by operating only on already invariant scalar input encodings such as distances, angles, or even systematic basis expansions. Our models, in contrast, operate directly on 3D geometry: their inputs, internal latent representations, and predictions can contain not only scalars, which are invariant under symmetry transformations of the input, but also “equivariant” geometric vectors and higher-order tensors, which transform correspondingly when the input is transformed. This approach yields remarkable real-world improvements in accuracy, geometric and chemical extrapolation, simulation stability, and data efficiency as demonstrated in small and flexible organic molecules, ionic liquids, catalytic and other materials systems, and biomolecules.  Finally, the Allegro model further enables scaling these desirable traits to simulations at large very length-scales and high speeds.

Accurate, data efficient, and scalable atomic-scale chemical modeling with equivariant deep learning: Allegro and NequIP

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.