AdaptNTK Improves Active Learning for Molecular Dynamics Simulations

Machine learning interatomic potentials are crucial for bridging the gap between quantum chemical precision and the speed of classical simulations, particularly in molecular dynamics. Improving their reliability often involves active learning, where the model iteratively identifies and adds uncertain, out-of-distribution configurations to its training set. However, existing uncertainty quantification methods often trade off computational cost for reliability or fail to account for redundancy when selecting new data. AdaptNTK addresses these issues by introducing a single-model framework that measures uncertainty as a regularized Mahalanobis distance within the empirical Neural Tangent Kernel (NTK) feature space. A key innovation is that NTK features remain fixed during acquisition, allowing uncertainty to be updated recursively without retraining after each selection, thereby minimizing redundancy within an acquisition batch. On benchmark datasets like rMD17 and Transition-1X, AdaptNTK demonstrates superior performance, achieving the lowest force errors and strong correlations with actual errors. It also provides a 2.6-fold speedup per active learning cycle compared to traditional ensemble methods, making it a highly efficient and accurate tool for data-efficient active learning in computational chemistry.