Learning molecular dynamics: predicting the dynamics of glasses by a machine learning simulator.

Area of Science:

• Computational physics
• Materials science
• Artificial intelligence

Background:

• Simulating atom dynamics, like glass dynamics, is challenging due to complex physics laws and computational cost.
• Existing methods struggle to balance accuracy with efficiency in capturing these dynamics.

Purpose of the Study:

• To develop a novel framework for simulating complex atom dynamics, specifically glass dynamics, bypassing explicit physics laws.
• To leverage static structural information for predicting dynamic atomic behavior efficiently.

Main Methods:

• Introduction of an observation-based graph network (OGN) framework utilizing graph neural networks (GNN).
• Training and application of the OGN to predict atom trajectories from static structures in various atomistic systems.
• Comparison with traditional molecular dynamics (MD) simulations.

Main Results:

• The OGN successfully predicted atom trajectories for complex glass dynamics over hundreds of timesteps.
• Atom dynamics were found to be significantly encoded within the static structure of disordered phases.
• OGN simulations demonstrated enhanced computational speed (≥5x) compared to MD, conserving energy and momentum.

Conclusions:

• The OGN framework offers a powerful, efficient alternative for simulating many-body dynamics, particularly in disordered systems.
• The findings suggest that static structure contains substantial information about dynamic processes, enabling physics-agnostic simulations.
• This approach has the potential for broad applicability to various many-body dynamics problems.