Symposium MT04: Next-Generation AI-Catalyzed Scientific Workflow for Digital Materials Discovery

Bowen Deng, University of California, Berkeley

Potential Energy Surface Softening in Universal Machine Learning Interatomic Potentials

Written by Jun Meng

Universal machine learning interatomic potentials (uMLIPs) are revolutionizing atomic-scale simulations by scaling quantum chemical accuracy to complex, large-scale systems. Bowen Deng from the University of California, Berkeley, explored both the promise and limitations of the top-tier uMLIPs, such as M3GNet, CHGNet, and MACE-MP-0, when applied to challenging out-of-distribution (OOD) environments like defects, surfaces, and ion migration barriers.

Deng highlighted a common issue in these models: potential energy surface (PES) softening, where energy and force predictions are systematically underpredicted. Surprisingly, this error can often be corrected by fine-tuning the models with as little as a constant correction factor, demonstrating the efficiency and adaptability of uMLIPs.

This work underscores the importance of systematic error analysis in next-generation materials modeling. By addressing PES softening, uMLIPs need to be optimized for diverse applications.