Modeling Dual-Range Atomic Interactions with Physicochemical Principles for Molecular Force Fields

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine Learning Force Fields (MLFFs) accelerate molecular dynamics simulations.
• Existing MLFFs struggle with long-range interactions and conformational variations.
• There's a need for adaptive MLFFs balancing short- and long-range forces.

Purpose of the Study:

• To introduce GeoNet, a physicochemical-principle-guided framework for dual-range atomic interactions.
• To improve the modeling of geometric characteristics in long-range interactions.
• To enhance the robustness and efficiency of MLFFs.

Main Methods:

• GeoNet utilizes geometric attention over atom-fragment bipartite graphs for long-range dependencies.
• Dual-level augmentation enforces semantic consistency across molecular conformations.
• An adaptive fusion module dynamically balances short- and long-range interactions.

Main Results:

• GeoNet consistently outperforms ten state-of-the-art baselines.
• Achieves the smallest model size and shortest training time.
• Demonstrates superior predictive performance and computational efficiency.

Conclusions:

• GeoNet offers a significant advancement in MLFFs for molecular dynamics.
• The framework provides accurate and efficient modeling of atomic interactions.
• GeoNet addresses key limitations of existing MLFF approaches.