Tensor improve equivariant graph neural network for molecular dynamics prediction.

Area of Science:

• Computational chemistry and materials science
• Artificial intelligence and machine learning for scientific discovery

Background:

• Molecular dynamics (MD) simulations are crucial for drug discovery, requiring accurate prediction of molecular motion.
• Existing methods often fail to fully account for chemical bonding constraints and primarily use scalar data.
• Incorporating group equivariance and tensor information can enhance the robustness and accuracy of MD predictions.

Purpose of the Study:

• To propose a novel Tensor-improved Equivariant Graph Neural Network (TEGNN) for molecular dynamics prediction.
• To address limitations in existing models by integrating chemical bonding constraints and richer tensor information.
• To improve the accuracy, efficiency, and constraint satisfaction of molecular dynamics simulations.

Main Methods:

• TEGNN materializes chemical bond constraints as geometric constraints using generalized coordinates for molecular kinematics.
• The model facilitates equivariant information transfer, enhancing prediction accuracy and computational efficiency.
• TEGNN incorporates equivariant locally complete frames to project tensor information onto scalar-only equivariant graph neural networks.

Main Results:

• TEGNN demonstrated superior prediction accuracy and constraint satisfaction compared to state-of-the-art Graph Neural Networks (GNNs).
• Experiments on simulated N-body datasets and the real MD17 dataset validated TEGNN's performance.
• The model showed improved data efficiency in molecular dynamics prediction tasks.

Conclusions:

• TEGNN advances equivariant neural networks by accommodating more tensor information and explicitly modeling chemical bonding constraints.
• The proposed model significantly improves the prediction of molecular kinematic states.
• TEGNN offers a more robust and accurate approach for molecular dynamics simulations in various scientific fields.