Accelerating Molecular Dynamics with a Graph Neural Network: A Scalable Approach through E(q)C-GNN
Debasis Maji, Atish Ghosh, Debaditya Barman, Pranab Sarkar
Abstract
Ab initio molecular dynamics simulations are an integral part of any electronic structure calculation to access thermal stability and perform non-adiabatic dynamics but are computationally very demanding. To enhance the computational efficiency of crucial ab initio molecular dynamics simulations, in this work, we implemented the graph neural network (GNN)-accelerated predictions for the molecular dynamics simulation of two-dimensional systems with varying atom connectivity. In this work, we developed an equivariant GNN model that employs only the time-evolved AIMD-simulated atomic coordinates for training and successfully predicts the key parameters of stable two-dimensional g-CN, WTe 2, and g-CN/WTe 2, like potential energy and kinetic energy, while also delving into the structural and thermodynamical parameters like entropy and interatomic force variation, resulting in a fluctuation level of ±3%, and the computational speed has improved by several orders of magnitude. Hence, incorporating an equivariant GNN model will serve as a viable substitute for predictions in extensive AIMD simulations of homogeneous or symmetrically periodic low-dimensional materials.