AI Insight
This study compares novel matrix-structured optimizers (SOAP, Muon, and SOAP-Muon) against the standard Adam optimizer for training machine learning interatomic potentials, specifically NequIP and Allegro models. The researchers found that SOAP and SOAP-Muon consistently outperform Adam in both training speed and final model accuracy, with particularly strong improvements when training with partial force supervision. The results demonstrate that optimizer selection, which has been largely overlooked in the field, can significantly impact the performance of machine learning models used for atomic-scale simulations.
Why it matters
Machine learning interatomic potentials are increasingly important for computational materials science and molecular dynamics simulations. By achieving faster training and better accuracy with improved optimizers, researchers can develop more efficient models with less data, potentially accelerating scientific discovery in chemistry, materials science, and drug design.
Understand the Science
arXiv:2607.02499v1 Announce Type: cross
Abstract: Machine learning interatomic potentials (MLIPs) have become a hallmark of AI for scientific simulation. While efforts on new architectures and datasets have led to increasingly accurate and general models, the choice of optimizer for training has largely remained unexplored, defaulting to Adam and its variants in the community. Here, we implement and systematically compare a class of recently proposed matrix-structured optimizers, including Muon, SOAP, and the hybrid SOAP-Muon, for training NequIP and Allegro MLIP models. We find that these optimizers can substantially outperform Adam in both convergence speed and final accuracy. SOAP and SOAP-Muon emerge as robust and consistently strong methods, while Muon only provides partial gains relative to Adam. The improvements are particularly pronounced under partial force supervision. Our results indicate that optimizer choice is an overlooked yet impactful design axis for MLIPs.