Truncated automatic sparse differentiation for machine learning interatomic potentials
physics.chem-ph, cond-mat.mtrl-sci, cs.LG
Submitted: 2026-09-17
Updated: 2026-09-17
Comments: 19 pages, 5 figures, 4 tables (7 pages main text). Additional information at https://marcel.science/tasd4mlip
Code: https://github.com/sirmarcel/tasd4mlip-archive
License: http://creativecommons.org/licenses/by/4.0/
The gist: Machine learning interatomic potentials (MLIPs) learn the mapping from atomic positions to potential energy.
Terminology
Abstract
Machine learning interatomic potentials (MLIPs) learn the mapping from atomic positions to potential energy. The forces, the negative gradient of this energy, drive molecular dynamics and are readily obtained using automatic differentiation. Higher-order derivatives, most notably the Hessian, describe collective motion and allow the direct prediction of experimental observables, but are considered computationally inaccessible for large systems. We suggest a solution: in physical systems, interactions decay with distance, and most MLIPs build on this locality through message passing up to a finite receptive field. This implies both sparsity of higher-order derivatives and their decay with distance. This structure can be exploited using automatic sparse differentiation (ASD). We explain how to compute the sparsity pattern for MLIP derivatives and demonstrate that, for multiple foundation MLIPs, ASD computes full Hessians of large porous materials exactly, but with modest speedups at best. The larger gains come from truncated ASD: discarding small, but nonzero, Hessian entries between distant atoms yields order-of-magnitude speedups with negligible impact on predicted observables.
Sources
- Relational inductive biases, deep learning, and graph networks
- Shoot from the HIP: Hessian Interatomic Potentials without derivatives
- High-quality, high-information datasets for universal atomistic machine learning
- Orb-v3: atomistic simulation at scale
- Projected Hessian Learning: Fast Curvature Supervision for Accurate Machine-Learning Interatomic Potentials
- MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures
Related papers
- Transferable Generative Models Bridge Femtosecond to Nanosecond Time-Step Molecular Dynamics
- Accelerated "on-the-fly" coupled-cluster path-integral molecular dynamics: Impact of nuclear quantum effects on an asymmetric proton
- Variational Polaron Theory for Ground States of Strongly Coupled Light-Matter and Electron-Phonon Systems
- Pushing the accuracy of on-top functionals with agent-driven supervised learning
- Scaling Machine Learning Interatomic Potentials with Mixtures of Experts
- Localized intrinsic bond orbitals decode correlated charge migration dynamics