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.
Authors
- Laura Zichi (ORCID: https://orcid.org/0000-0003-3897-3097)
- Anders F. Johansson (ORCID: https://orcid.org/0000-0003-0548-5943)
- Marc Descoteaux (ORCID: https://orcid.org/0000-0001-8504-3429)
- Albert Musaelian (ORCID: https://orcid.org/0000-0003-0475-9236)
- Chuin Wei Tan (ORCID: https://orcid.org/0000-0002-0087-7537)
- Seán R. Kavanagh (ORCID: https://orcid.org/0000-0003-4577-9647)
- Tess Smidt (ORCID: https://orcid.org/0000-0001-5581-5344)
- William C. Witt (ORCID: https://orcid.org/0000-0002-1578-1888)
- Mit Kotak (ORCID: https://orcid.org/0000-0002-6999-1055)
- Aadit Saluja
- Menghang Wang (ORCID: https://orcid.org/0000-0003-2912-1239)
- Gabriel de Miranda Nascimento
- Boris Kozinsky
- Yizhong R. Hu
Institutions
- Harvard University (US)
- Boston Applied Technologies (United States) (US)
- Sandia National Laboratories (US)
- Robert Bosch (United States) (US)
- Massachusetts Institute of Technology (US)
Publication Details
- Journal
- Digital Discovery
- Published
- 2026-01-01
- DOI
- https://doi.org/10.1039/d5dd00423c
- Citations
- 1
- Primary Topic
- Machine Learning in Materials Science
- Type
- article
- Field-Weighted Citation Impact
- 16.08