ABSTRACT
Accurate modeling of atomic-scale processes, such as protein dynamics and catalysis, is a central challenge in computational structural biology, chemistry, and materials science. While machine learning force fields (MLFFs) have emerged as powerful tools, approaching quantum mechanical accuracy with promising generalisation capabilities, their application is hindered by prohibitive inference times, particularly for long timescale simulations of large systems required for many biological applications. In this work, we introduce BoostMD, a MLFF surrogate architecture, designed to mitigate this computational bottleneck. BoostMD leverages node features from previous molecular dynamics time steps to predict forces and energies, enabling the use of a smaller, faster model between evaluations of a large reference MLFF. The approach provides up to 8x speedup over the ground truth reference model. Testing on unseen dipeptides demonstrates that BoostMD accurately generalises and reproduces Boltzmann-distributed samples, making it a robust tool for efficient, long-timescale molecular simulations.
