TAIP: Test-Time Adaptation Interatomic Potential for Molecular Dynamics

We propose a novel online test-time adaptation strategy for interatomic potentials (TAIP) to improve the accuracy and generalizability of machine learning interatomic potentials (MLIPs)

Our work introduces a novel online test-time adaptation strategy for improving the performance of interatomic potentials on unseen molecular structures. The method is applied to two widely used invariant and equivariant machine learning interatomic potentials, namely SchNet and PaiNN, respectively. We demonstrate the superior performance of our method on multiple datasets, including small organic molecules, isomers of C7O2H10, liquid water, ice, and electrolyte solutions of different concentrations. The test data from these datasets are designed to have a larger distribution shift from the training data, and are therefore used to investigate the improvement in generalizability achieved by our method. We evaluate the performance of the models on fixed datasets and in molecular dynamics simulations. The results show that our method can enhance the accuracy of interatomic potentials on unseen molecular structures, and it is compatible with complex systems featuring periodic boundary conditions. Moreover, we explore the impact of our method on feature distributions through dimensionality reductions, which provides insights into the mechanism behind the improvement from the perspective of feature embedding.