ViSNet: a scalable and accurate geometric deep learning potential for molecular dynamics simulation

Geometric deep learning has been revolutionizing the molecular dynamics simulation field over a decade. Although the state-of-the-art neural network models are approaching ab initio accuracy for energy and force prediction, insufficient utilization of geometric information and high computational costs hinder their applications in molecular dynamics simulations. Here we propose a deep learning potential, called ViSNet that sufficiently exploits directional information with low computational costs. ViSNet outperforms the state-of-the-art approaches on the molecules in the MD17 and revised MD17 datasets and achieves the best prediction scores for 11 of 12 quantum properties on QM9. Furthermore, ViSNet can scale to protein molecules containing hundreds of atoms and reach to ab initio accuracy without molecular segmentation. Through a series of evaluations and case studies, ViSNet can efficiently explore the conformational space and provide reasonable interpretability to map geometric representations to molecular structures.