ChronoCast: A time-series paradigm for molecular dynamics simulation using equivariant graph neural networks

Area of Science:

• Computational Physics
• Materials Science
• Machine Learning

Background:

• Traditional molecular dynamics simulations face significant computational costs for complex systems.
• Studying statistical physics of dynamical processes requires accurate and efficient simulation methods.

Purpose of the Study:

• To introduce ChronoCast, a time-series paradigm for molecular dynamics simulation.
• To overcome the computational challenges of traditional simulation algorithms.
• To develop an accurate and efficient method for studying complex dynamical processes.

Main Methods:

• Developed ChronoCast, an autoregressive Equivariant Graph Neural Network.
• Reformulated molecular dynamics integration as a time-series forecasting task.
• Incorporated velocity as a node feature and enforced momentum conservation.

Main Results:

• ChronoCast accurately reproduces diverse physical properties, including radial distribution function, mean-squared displacement, and vibrational density of states.
• Demonstrated accuracy on simple Si crystal and complex NbSe3 nanowires.
• Achieved orders of magnitude reduction in trajectory generation time compared to ab initio methods and over 50% reduction compared to state-of-the-art machine learning potentials.

Conclusions:

• ChronoCast offers a highly accurate and efficient time-series approach for molecular dynamics simulations.
• The method significantly reduces computational cost, enabling the study of complex physical problems.
• This work presents a promising advancement in simulating statistical physics of complex dynamical processes.