Beyond Adam: SOAP and Muon for Faster, Label-Efficient Training of Machine Learning 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 systematically implements and compares matrix-structured optimizers—Muon, SOAP, and a hybrid SOAP-Muon—for training machine learning interatomic potentials (MLIPs), specifically the NequIP and Allegro architectures. The study shows that these second-order, matrix-aware optimizers substantially outperform the widely used Adam optimizer in both convergence speed and final accuracy. The gains are especially pronounced under partial force supervision, a label-efficient regime where only a subset of atomic forces is available during training. This work suggests that adopting SOAP- and Muon-style optimizers can significantly reduce the computational cost and data requirements of training MLIPs for molecular and materials simulations. The paper spans machine learning, AI, chemical physics, and computational physics, and is available on arXiv under identifier 2607.02499.