Learning the Action for Long-Time-Step Simulations of Molecular Dynamics.

Area of Science:

• Computational Physics
• Machine Learning
• Classical Mechanics

Background:

• Classical mechanics models physical systems but requires small time steps for accuracy, limiting computational efficiency.
• Machine learning (ML) can extend time steps but often introduces artifacts like energy non-conservation.

Purpose of the Study:

• To develop data-driven, structure-preserving ML models for accurate, long-time-step classical dynamics.
• To demonstrate that these models learn the system's mechanical action.

Main Methods:

• Learning structure-preserving (symplectic, time-reversible) maps using ML.
• Deriving ML integrators from the learned mechanical action.
• Validating models on short reference trajectories and transferring them across conditions.

Main Results:

• Action-derived ML integrators eliminate artifacts seen in non-structure-preserving ML predictors.
• The proposed method enables long-time-step integration with improved accuracy and efficiency.
• Models can be iteratively applied to correct cheaper direct predictors.

Conclusions:

• Learning the mechanical action via structure-preserving ML provides a robust method for simulating classical dynamics.
• This approach significantly enhances computational efficiency without sacrificing physical accuracy.
• The technique offers a powerful tool for molecular dynamics and other complex systems.