Learning physics-consistent particle interactions.

Area of Science:

• Physics
• Materials Science
• Computational Science

Background:

• Interacting particle systems are crucial in science and engineering.
• Understanding particle interaction laws is key but often hindered by system complexity.
• Existing machine learning methods primarily focus on particle-level kinetics, leaving pairwise interaction learning as a challenge.

Purpose of the Study:

• To develop a machine learning algorithm capable of learning hidden pairwise particle interactions.
• To infer particle interaction laws that are consistent with underlying physical principles.
• To improve the generalization and robustness of models for interacting particle systems.

Main Methods:

• Adaptation of the Graph Networks framework with distinct edge and node components.
• Implementation of a deterministic operator in the node part for precise pairwise interaction inference.
• Training the model to predict particle acceleration to ensure physical consistency with Newton's laws.

Main Results:

• The proposed physics-induced graph network successfully infers pairwise interactions with superior performance across multiple datasets.
• The methodology demonstrates consistency with underlying physical laws, unlike previous approaches.
• The model exhibits enhanced generalization to larger systems and robustness to noise compared to baseline models.

Conclusions:

• The developed algorithm effectively learns physically consistent pairwise particle interactions.
• This approach advances the understanding and discovery of fundamental particle interaction laws.
• The methodology has the potential to guide the design of materials with specific properties.