DysNet: Learning Implicit Many-Body Interactions via Dynamically Attending to Body-Orders in Equivariant Graph...

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• First-principles simulations offer deep materials insights but are computationally expensive.
• Machine learning interatomic potentials (MLIPs) reduce costs, enabling large-scale simulations with DFT-level accuracy.
• Accelerating materials discovery requires efficient and accurate simulation methods.

Purpose of the Study:

• Introduce DysNet, a novel SE(3)-equivariant graph neural network.
• Address the trade-off between physical fidelity and computational efficiency in MLIPs.
• Improve the scalability and accuracy of molecular simulations for materials science.

Main Methods:

• Developed DysNet, a dynamic and spherical network architecture.
• Implemented a chemically gated interorder attention (CG-IOA) mechanism.
• Utilized an efficient many-body message passing framework and spherical harmonic tensor embeddings.

Main Results:

• DysNet achieves comparable or superior accuracy to state-of-the-art equivariant graph neural networks.
• Demonstrated remarkable adaptability across diverse chemical benchmark datasets (QM9, rMD17, 3BPA, SPICE).
• The CG-IOA mechanism effectively balances physical accuracy and computational efficiency.

Conclusions:

• DysNet offers a significant advancement in MLIPs for accelerating materials discovery.
• The model's design enables efficient and accurate large-scale dynamic simulations.
• DysNet shows promise for broad applications in computational chemistry and materials science.