Graph neural networks accelerated molecular dynamics.

Area of Science:

• Computational chemistry
• Materials science
• Artificial intelligence in scientific computing

Background:

• Molecular Dynamics (MD) simulations offer atomic-scale insights into matter dynamics but are computationally expensive for long timescales.
• Calculating forces in MD requires iterative computations of potential energy and spatial gradients, limiting simulation speed.

Purpose of the Study:

• To develop a deep learning model that accelerates Molecular Dynamics (MD) simulations.
• To bypass computationally intensive potential energy calculations by directly predicting forces.

Main Methods:

• Developed a Graph Neural Network Accelerated MD (GAMD) model.
• Trained the GNN on diverse data sources including classical MD and density functional theory.
• Validated GAMD on Lennard-Jones and water systems in the NVT ensemble.

Main Results:

• GAMD accurately predicts molecular dynamics by directly forecasting forces, bypassing potential energy calculations.
• The model demonstrates scalability to larger systems during inference, irrespective of training data scale.
• Benchmark tests show GAMD's performance is competitive with production-level MD software for large-scale simulations.

Conclusions:

• Graph Neural Network Accelerated MD (GAMD) offers a computationally efficient alternative for large-scale molecular dynamics simulations.
• The developed model shows promise for accelerating scientific discovery in fields relying on atomic-scale simulations.
• GAMD's ability to learn and predict forces agnostic to system scale represents a significant advancement in simulation methodology.