Model Knowledge Transfer for Ab Initio Molecular Dynamics Simulations

Machine learning based potential energy surface (PES) models have revolutionized ab initio molecular dynamics (AIMD) simulations, providing a much faster alternative to traditional electronic structure methods. Nevertheless, despite their speed advantages, these models incur significant computational overhead and resource inefficiencies when directly used for predicting molecular properties with pre-trained models, which ultimately reduces their overall efficiency. To address this, we propose a novel model knowledge transfer strategy for enhancing ab initio molecular dynamics (KTAIMD) simulations without compromising accuracy. This strategy, applied as a post-training refinement step, significantly reduces computational overhead while retaining the precision of the original model. Extensive experiments conducted on the plane-wave density functional theory (PWDFT) platform validate the KTAIMD strategy, demonstrating a notable reduction in memory usage by an order of magnitude. Consequently, this approach offers substantial practical value for large-scale simulations and establishes a new benchmark for efficiency and scalability in molecular dynamics.