Fine-Tuning Universal Machine-Learned Interatomic Potentials for Applications in the Science of Steels

Recycling steel at scale is hindered by tramp elements such as Cu and Sn, which degrade material properties. Atomistic simulations using foundational machine-learned interatomic potentials (MLIPs) trained on large databases, such as Materials Project, Alexandria, and OMAT, offer a promising approach to study the effects of these impurities. However, fine-tuning these models to specific systems can lead to catastrophic forgetting–the loss of general chemical knowledge acquired during pretraining. Here, we evaluate forgetting in three foundational MLIPs: CHGNet, SevenNet-O, and MACE, by fine-tuning on a data set of bcc-based structures, with Fe atoms only. When evaluated on a subset of the Materials Project data set with a learning rate of 0.0001, the fine-tuned MLIPs of CHGNet and SevenNet-O exhibited only a minor increase in RMSE of 0.047 and 0.022 eV/atom, respectively, indicating markedly minor forgetting. In contrast, fine-tuned MACE exhibited catastrophic forgetting, despite a range of additional strategies such as layer freezing and data set replay. We attribute the catastrophic forgetting to architectural sensitivity. These results highlight the importance of fine-tuning hyperparameters, model architecture, and data set design, with fine-tuned models of CHGnet and SevenNet-O showing some potential for efficient and transferable modeling of recycled steels.