Towards exascale reactive molecular dynamics with equivariant and Bayesian force fields

Understanding atomic-level processes in catalysis, batteries and biomolecular systems is complicated by the wide range of time and length scales needed for simulations. Machine learning force fields have recently achieved significant progress in capturing interatomic interactions, closely approaching first principles calculations used as their training data, but at much higher efficiency. We show that it is possible to deploy these models within LAMMPS on parallel GPUs to reach billions of atoms in size or microseconds in time. One such method family is based on equivariant neural networks (NequIP [1] and Allegro [2]) with symmetry-preserving layer architectures that achieves record accuracy and training efficiency for simulating dynamics of biomolecules and ionic conductors. Another method (FLARE [3]) introduces the concept of uncertainty-aware Bayesian force fields based on Gaussian process regression, enabling autonomous selection of the training set using an on-the-fly active learning algorithm during a simulation. We apply these ML-accelerated MD simulations to study surface reconstruction, direct heterogeneous reactions, and nanoparticle shape changes.

[1] S. Batzner et al, Nature Comm. 13 (1), 2453 (2022)
[2] A. Musaelian, S. Batzner et al, Nature Comm. 14, 579 (2023)
[3] J. Vandermause et al, Nature Comm. 13 (1), 5183 (2022)

• Wednesday, 09 Aug 2023
• 09:55 - 10:25 EDT