Universal machine learning interatomic potentials (uMLIPs) represent a significant advancement in interatomic potential modeling, offering remarkable predictive accuracy across a wide range of chemical systems. However, their applications in catalytic reaction simulation are limited by their lack of accuracy in describing reactions, especially in reaction barrier prediction. In this study, we evaluate two established uMLIPs and use fine-tuning strategies to enhance their performance for the prediction of catalytic reaction prediction. We systematically compared the predictive accuracy, data efficiency, and generalization capabilities of two approaches, fine-tuning and training from scratch, using the accuracy of the original pretrained uMLIPs as a baseline. Specifically, we evaluated the applicability of the approaches across a range of tasks, from relatively simple applications such as molecular dynamics (MD) simulations and adsorption energy calculations to more complex challenges such as transition state searches. We also analyzed model performance across varying training set sizes to identify the critical data threshold needed for accurate reaction predictions. Additionally, we assessed the extrapolative generalization of the models by examining improvements in predictive accuracy for unseen elements following fine-tuning across both simple and complex tasks. Our results show that fine-tuning uMLIPs significantly improves the accuracy of reaction energy predictions, reducing the mean absolute error (MAE) to 0.09 eV, compared to 0.38 eV for the original uMLIPs. Notably, the fine-tuned models require only 10%-30% of the data used for training from scratch, yielding a stable and reliable performance. Moreover, the generalization capabilities of the uMLIPs were preserved after fine-tuning. This approach shows significant promise for extending the uMLIPs applicability to diverse catalytic reaction systems.
Ma et al. (2026) studied this question.