Multiscale ForceField Model Based on a Graph NeuralNetwork for Complex Chemical Systems

Inspired by the QM/MM methodology, the ML/MM approach introduces a new opportunity for multiscale simulation, improving the balance between accuracy and computational efficiency. Benefited from the rapid advancements in molecular embedding methods, density functional theory level quantum mechanical (QM) calculations within the QM/MM framework can be accelerated by several orders of magnitude through the application of machine learning (ML) potential energy surfaces. As a problem inherited from the QM/MM methodology, challenges exist in designing the interactions between machine learning and molecular mechanics (MM) regions. In this study, electrostatic interactions between machine learning and MM atoms are treated by using a graphical neural network based on stationary perturbation theory. In this protocol, we process coordinates and MM charges to yield electrostatic energy and forces, resulting in a high-performance electrostatic embedding ML/MM architecture. The accuracy of the ML/MM energy was validated in aqueous solutions of alanine dipeptide and allyl vinyl ether (AVE). We investigated the transferability of parameters trained from AVE in a single solvent to various other solvents, including water, methanol, dimethyl sulfoxide, toluene, ionic liquids, and water–toluene interface environments. We then established a solvent-free protocol for data set preparation. Comparison of the free energy landscapes of the Claisen rearrangement of AVE in different solvation environments showed the catalytic effect of aqueous solutions, consistent with experiments.