High-performance training and inference for deep equivariant interatomic potentials

The NequIP framework is redesigned for scalable distributed training and PyTorch 2.0 compilation. AOT Inductor inference and optimized Allegro kernels accelerate molecular dynamics by factors of 5–18 on practical system sizes.

• NS
  National Science Foundation

  Awards: PHY-2019786, 2118201, DMR-2119351, 2119351, 2019786, DE-AC05-00OR22725, DGE-1745302, 1745302, DE-AC02-05CH11231, DE-NA-0003525, 2011754, DMR-2011754
• UD
  U.S. Department of Energy

  Awards: -AC02-05CH11231, AC05-00OR22725, 0003525, 05CH11231, DE-SC0022199, AC02-05CH11231, DE-AC02, DE-NA-0003525, DE-AC05, 00OR22725, DE-AC02-
• HU
  Harvard University

  Award: DMR-2011754
• RB
  Robert Bosch
• NE
  National Energy Research Scientific Computing Center

  Awards: 05CH11231, AC02-05CH11231, DE-AC05-00OR22725
• OO
  Office of Science

  Awards: DE-AC05-00OR22725, DE-SC0022199, DE-NA-0003525, AC02-05CH11231, -AC02-05CH11231, DE-AC02, AC05-00OR22725
• NN
  National Nuclear Security Administration

  Awards: AC02-05CH11231, DE-AC02-05CH11231, DE-AC05-00OR22725, DE-NA-0003525
• CF
  Center for the Environment, Harvard University
• UN
  U.S. Navy

  Award: N00014-20-1-2418
• MR
  Materials Research Science and Engineering Center, Harvard University

  Awards: DE-AC02-05CH11231, DMR-2011754, 2011754
• OO
  Office of Naval Research

  Awards: DE-AC02-05CH11231, DE-AC05-00OR22725, N00014-20-1-2418, N00014
• DO
  Division of Materials Research

  Awards: DMR-2011754, 2011754, 2118201
• OO
  Office of Advanced Cyberinfrastructure

  Award: 2118201
• BE
  Basic Energy Sciences

  Awards: DE-AC02, DE-AC05-00OR22725, AC02-05CH11231, AC05-00OR22725, DE-SC0022199, DE-AC02-05CH11231, -AC02-05CH11231, DE-NA-0003525
• FD
  FAS Division of Science, Harvard University
• SN
  Sandia National Laboratories

  Awards: DE-AC02-05CH11231, DE-NA-0003525