Beyond Adam: SOAP and Muon Optimizers for Faster, Label-Efficient Training of ML Interatomic Potentials

Summary

A new arXiv paper (2607.02499) by Gil Harari, Yoel Zimmermann, Ola Tangen Kulseng, Laura Zichi, Chuin Wei Tan, Marc L. Descoteaux, and Boris Kozinsky implements and systematically compares matrix-structured optimizers—Muon, SOAP, and a hybrid SOAP-Muon—for training two leading machine learning interatomic potentials (MLIPs), NequIP and Allegro. The study shows that these optimizers substantially outperform Adam (AdamW) in both convergence speed and final accuracy. The gains are especially pronounced under partial force supervision, where force labels are limited or incomplete, indicating that matrix-structured optimizers enable more label-efficient training of MLIPs. This work is relevant to computational chemistry, materials simulation, and machine learning researchers seeking faster training pipelines for atomistic models, as it demonstrates that optimizer choice—beyond architecture and data—can be a key lever for improving MLIP training efficiency and model quality.