Fast and Accurate Foundation Potentials with Equivariant Graph Neural Networks (NequIP & Allegro)
ORAL
Abstract
Machine learning interatomic potentials, particularly deep equivariant graph neural networks such as NequIP, have demonstrated state-of-the-art accuracy and computational efficiency across diverse atomistic modeling tasks.
Leveraging recent advances in computational performance and extensibility, we present a family of foundation potentials with the NequIP[1,2] and Allegro[3] model frameworks, trained on the largest open-access materials datasets.[4] NequIP foundation models achieve leading inference speeds, scalability and accuracies on community benchmarks.[5,6] Crucially, these achievements challenge recent proposals to remove symmetry constraints (e.g. equivariance) in order to achieve high training and inference speeds with large models.
Lastly, we highlight that the modular and extensible NequIP framework permits customisable fine-tuning and extensions of these models to diverse prediction regimes, with select case studies.
1 S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt and B. Kozinsky, Nat Commun, 2022, 13, 2453.
2 C. W. Tan, M. L. Descoteaux, M. Kotak, G. de M. Nascimento, S. R. Kavanagh, L. Zichi, M. Wang, A. Saluja, Y. R. Hu, T. Smidt, A. Johansson, W. C. Witt, B. Kozinsky and A. Musaelian, arXiv, 2025, preprint, arXiv:arXiv:2504.16068, DOI: 10.48550/arXiv.2504.16068.
3 A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth and B. Kozinsky, Nat Commun, 2023, 14, 579.
4 L. Barroso-Luque, M. Shuaibi, X. Fu, B. M. Wood, M. Dzamba, M. Gao, A. Rizvi, C. L. Zitnick and Z. W. Ulissi, arXiv, 2024, preprint, arXiv:arXiv:2410.12771, DOI: 10.48550/arXiv.2410.12771.
5 J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chiang, B. Deng, G. Ceder, M. Asta, A. A. Lee, A. Jain and K. A. Persson, Nat Mach Intell, 2025, 7, 836–847.
6 B. Póta, P. Ahlawat, G. Csányi and M. Simoncelli, arXiv, 2024, preprint, arXiv:arXiv:2408.00755, DOI: 10.48550/arXiv.2408.00755.
Leveraging recent advances in computational performance and extensibility, we present a family of foundation potentials with the NequIP[1,2] and Allegro[3] model frameworks, trained on the largest open-access materials datasets.[4] NequIP foundation models achieve leading inference speeds, scalability and accuracies on community benchmarks.[5,6] Crucially, these achievements challenge recent proposals to remove symmetry constraints (e.g. equivariance) in order to achieve high training and inference speeds with large models.
Lastly, we highlight that the modular and extensible NequIP framework permits customisable fine-tuning and extensions of these models to diverse prediction regimes, with select case studies.
1 S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt and B. Kozinsky, Nat Commun, 2022, 13, 2453.
2 C. W. Tan, M. L. Descoteaux, M. Kotak, G. de M. Nascimento, S. R. Kavanagh, L. Zichi, M. Wang, A. Saluja, Y. R. Hu, T. Smidt, A. Johansson, W. C. Witt, B. Kozinsky and A. Musaelian, arXiv, 2025, preprint, arXiv:arXiv:2504.16068, DOI: 10.48550/arXiv.2504.16068.
3 A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth and B. Kozinsky, Nat Commun, 2023, 14, 579.
4 L. Barroso-Luque, M. Shuaibi, X. Fu, B. M. Wood, M. Dzamba, M. Gao, A. Rizvi, C. L. Zitnick and Z. W. Ulissi, arXiv, 2024, preprint, arXiv:arXiv:2410.12771, DOI: 10.48550/arXiv.2410.12771.
5 J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chiang, B. Deng, G. Ceder, M. Asta, A. A. Lee, A. Jain and K. A. Persson, Nat Mach Intell, 2025, 7, 836–847.
6 B. Póta, P. Ahlawat, G. Csányi and M. Simoncelli, arXiv, 2024, preprint, arXiv:arXiv:2408.00755, DOI: 10.48550/arXiv.2408.00755.
–
Presenters
-
Seán R Kavanagh
- University of Cambridge
- Cambridge University