Accelerating Molecular Dynamics with a Graph Neural Network: A Scalable Approach through E(q)C-GNN.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Ab initio molecular dynamics (AIMD) simulations are essential for studying thermal stability and non-adiabatic dynamics.
• AIMD simulations are computationally expensive, limiting their application in complex systems.

Purpose of the Study:

• To enhance the computational efficiency of AIMD simulations using graph neural networks (GNNs).
• To develop an equivariant GNN model for predicting properties of two-dimensional (2D) materials.

Main Methods:

• Implemented an equivariant GNN model trained on AIMD-simulated atomic coordinates.
• Applied the GNN model to predict potential energy, kinetic energy, entropy, and interatomic forces for 2D g-CN, WTe2, and g-CN/WTe2 systems.

Main Results:

• The GNN model accurately predicted key parameters with a fluctuation level of ±3%.
• Achieved a significant improvement in computational speed, by several orders of magnitude.
• Demonstrated the model's ability to handle varying atom connectivity in 2D systems.

Conclusions:

• Equivariant GNNs offer a computationally efficient alternative for extensive AIMD simulations of low-dimensional materials.
• This approach is suitable for homogeneous or symmetrically periodic 2D materials.
• The developed GNN model accelerates the study of material properties and dynamics.