Equivariant Graph Network Approximations of High-Degree Polynomials for Force Field Prediction
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: graph neural network, deep learning, molecule
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Equivariant deep models have recently been employed to predict atomic potentials and force fields in molecular dynamics. A key advantage of these models is their ability to learn from data without requiring explicit physical modeling. Nevertheless, use of models obeying underlying physics can not only lead to better performance, but also yield physically interpretable results. In this work, we propose a new equivariant network, known as PACE, to incorporate many-body interactions by making use of the Atomic Cluster Expansion (ACE) mechanism. To provide a solid foundation for our work, we perform theoretical analysis showing that our proposed message passing scheme can approximate any equivariant polynomial functions with constrained degree. By relying physical insights and theoretical foundations, we show that our model achieves state-of-the-art performance on atomic potential and force field prediction tasks on commonly used benchmarks.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1515
Loading