A Study on the Fine-Tuning Performance of Universal Machine-Learned Interatomic Potentials (U-MLIPs)

Abstract: Universal machine-learned interatomic potentials (U-MLIPs) have demonstrated effectiveness across diverse atomistic systems but often require fine-tuning for task-specific accuracy. We investigate the fine-tuning of two MACE-based foundation models, MACE-MP-0 and its variant MACE-MP-0b, and identify key insights. Fine-tuning on task-specific datasets enhances accuracy and, in some cases, outperforms models trained from scratch. Additionally, fine-tuned models benefit from faster convergence due to the strong initial predictions provided by the foundation model. The success of fine-tuning also depends on careful dataset selection, which can be optimized through filtering or active learning. We further discuss practical strategies for achieving better fine-tuning foundation models in atomistic simulations and explore future directions for their development and applications.