MGNN: Moment Graph Neural Network for Universal Molecular Potentials

Abstract

The quest for efficient and robust deep learning models for molecular systems representation is increasingly critical in scientific exploration. The advent of message passing neural networks has marked a transformative era in graph-based learning, particularly in the realm of predicting chemical properties and expediting molecular dynamics studies. We present the Moment Graph Neural Network (MGNN), a rotation-invariant message passing neural network architecture that capitalizes on the moment representation learning of 3D molecular graphs, is adept at capturing the nuanced spatial relationships inherent in three-dimensional molecular structures. From benchmark tests on public datasets, MGNN delivers multiple state-of-the-art results on QM9, revised MD17 and MD17-ethanol. Its generalizability and efficiency are also tested in additional systems including 3BPA and 25-element high-entropy alloys. The prowess of MGNN also extends to dynamic simulations, accurately predicting the structural and kinetic properties of complex systems such as amorphous electrolytes, with results that closely align with those from ab-initio simulations. The application of MGNN to the simulation of molecular spectra exemplifies its potential to offer a promising alternative to traditional electronic structure methods.