SOAP and Muon optimizers outperform Adam in training machine learning interatomic potentials

Researchers systematically compared matrix-structured optimizers, including SOAP, Muon, and the hybrid SOAP-Muon, against the default Adam optimizer for training NequIP and Allegro machine learning interatomic potentials (MLIPs). The study found that these new optimizers substantially outperform Adam in both convergence speed and final accuracy.

• SOAP and SOAP-Muon emerged as robust methods with consistent performance gains.
• Muon provided only partial improvements relative to Adam.
• Performance advantages were particularly pronounced under conditions of partial force supervision.

The results indicate that the choice of optimizer is a critical, yet often overlooked, design axis for improving the efficiency and accuracy of MLIP training.