Accelerating Molecular Dynamics Simulations with Quantum Accuracy by Hierarchical Classification

There has been substantial interest in the development of machine learning methods to compute the electronic structure of molecular systems with quantum accuracy with significantly reduced cost. However, due to the numerical complexity of high dimensional embeddings of the molecular information, these methods are still far more costly than traditional molecular-dynamics simulations. Here we demonstrate that a three-layer partitioning strategy of the molecular ensemble, accompanied by system-specific training of inter- and intra-molecular interactions, can accelerate reactive ML-based simulations. At the example of hydrogen-transfer reactions of hydronium in aqueous solution, we find an acceleration of three-orders of magnitude in comparison to the underlying ML model without significant loss of accuracy. This finding paves the way to enable the use of efficient ML models in large-scale molecular dynamics simulations that are applicable to problems of current interest.