Teachers that teach the irrelevant: Pre-training machine learned interaction potentials with classical force fields for robust molecular dynamics simulations

Abstract: Machine learned interaction potentials (MLIPs) have become a critical component of large-scale, high-quality simulations for a range of chemical and biochemical systems. Yet, despite their in-distribution accuracy, molecular dynamics simulations using MLIPs exhibit numerical instabilities due to underlying data insufficiencies when encountering new regions of the potential energy surface. Here we propose a pre-training learning scheme that uses low-quality, practically free, single-molecule non-reactive force field data while all intermolecular interactions and reactive properties are learned at a fine-tuning stage with a small amount of computationally more expensive labels. We show that the force field pre-training approach followed by data efficient ab initio fine tuning allows for stable and accurate molecular dynamics and metadynamics simulations of gas phase molecules, liquid water, and hydrogen combustion reactions compared to models trained from scratch.