Speaker
Description
This work demonstrates that systematic fine-tuning transforms modern foundation machine-learned interatomic potentials (MLIPs) into consistently accurate, near-ab initio models across diverse architectures. Benchmarking five leading frameworks-MACE, GRACE, SevenNet, MatterSim, and ORB-we show that fine-tuning improves force and energy predictions by up to one and three orders of magnitude, respectively, while largely eliminating architecture-dependent performance differences.
Beyond benchmarking, we systematically explore the roles of training strategy and data composition by comparing fine-tuning with training-from-scratch across varying dataset sizes. In particular, we investigate the impact of data generated via short ab initio molecular dynamics trajectories alongside structures obtained from foundation-model-driven simulations and subsequently recalculated with density functional theory. This unified perspective reveals how dataset origin and size jointly control accuracy, efficiency, and transferability.
Focusing on rare-event systems involving bond breaking, formation, and high energy barrier transitions, we develop simple, robust, and systematically improvable protocols that remain effective even in data-limited regimes. We further demonstrate an iterative fine-tuning approach, where model-driven sampling and retraining progressively refine accuracy, enabling efficient exploration of challenging regions of configuration space.
These strategies yield harmonized performance across architectures while maintaining computational efficiency and enabling realistic simulations of degradation reactions in fuel cells under operating conditions. To facilitate adoption, we introduce the aMACEing Toolkit, providing unified and reproducible workflows for both fine-tuning and training-from-scratch.