Toward Generalizable Surrogate Models for Molecular Dynamics via Graph Neural Networks.

Area of Science:

• Computational Physics
• Materials Science
• Artificial Intelligence

Background:

• Traditional molecular dynamics (MD) simulations are computationally intensive due to repeated force evaluations and numerical integration.
• Accurate prediction of atomic-level behavior is crucial for understanding material properties and dynamics.
• Developing efficient computational frameworks is essential for advancing atomistic simulations.

Purpose of the Study:

• To introduce a novel graph neural network (GNN) based surrogate framework for molecular dynamics (MD) simulations.
• To enable direct prediction of atomic displacements and learn the system's evolution operator.
• To provide a computationally efficient alternative to conventional MD for accelerated atomistic simulations.

Main Methods:

• Representing atomic environments as graphs and utilizing message-passing layers with attention mechanisms.
• Developing a surrogate model that propagates atomic configurations forward in time without explicit force computation.
• Applying the model in an autoregressive manner for multistep temporal evolution.
• Training the GNN surrogate on classical MD trajectories of bulk aluminum.

Main Results:

• The GNN surrogate achieves sub-angstrom accuracy within the training horizon.
• The model demonstrates stable temporal extrapolation capabilities for short to mid-term predictions.
• Validated structural and dynamical fidelity through agreement with radial distribution functions and mean squared displacement.
• Preservation of key physical signatures beyond simple coordinate accuracy.

Conclusions:

• GNN-based surrogate integrators offer a promising and computationally efficient complement to traditional MD.
• The developed framework accelerates atomistic simulations within validated settings.
• The approach effectively captures local coordination and many-body interactions in metallic systems.