Learning the energy relaxation manifold from unrelaxed structures with RelaxNet

ICLR 2026 Conference Submission21302 Authors

Keywords: neural ODEs, energy minimization, trajectory, relaxation, forcefield, optimization

TL;DR: We developed RelaxNet, a dynamics-aware, equivariant model that leverages neural ordinary differential equations and message passing neural networks for predicting energy relaxation landscapes between initial unrelaxed and final relaxed structures.

Abstract: In efforts to bypass computationally-expensive density functional theory (DFT) calculations for energy minimization and structure relaxation, rapid progress in the development of machine learning force fields (MLFF) and more robust models that adhere to quantum chemistry/physical paradigms and constraints have been realized. However, most research to date involves energy predictions in a static frame only (i.e., given a specific atomic configuration, predict the energy of the current or final instance), which neglects intermediary physical insight-providing contexts. In this work, we developed RelaxNet, a dynamics-aware, equivariant deep learning model that leverages neural ordinary differential equations (neural ODEs) and message passing neural networks (MPNNs) for predicting the energy relaxation landscape between the initial unrelaxed structure and final relaxed structure for the first time. From just the initial structure, which is often the configuration that is fed into DFT simulations, we can accurately recover the energy/forces for the entire trajectory at a competitive prediction accuracy. We further provide extensive insights on the use of implicit vs. explicit latent embedding evolution schemes to offer perspectives on optimal methods for future works integrating expensive graph-based neural networks and neural ODEs.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21302