Constructing custom thermodynamics using deep learning.

Area of Science:

• Computational Physics
• Artificial Intelligence in Science
• Polymer Physics

Background:

• Traditional physical intuition struggles with complex phenomena.
• Automated scientific discovery can accelerate research by analyzing vast datasets.
• Physical principles like symmetries and conservation laws are crucial constraints.

Purpose of the Study:

• To develop a platform for automated discovery of macroscopic dynamics in stochastic dissipative systems.
• To learn interpretable thermodynamic coordinates and their associated dynamics directly from microscopic trajectories.
• To apply and validate the methodology on the problem of polymer chain stretching.

Main Methods:

• Utilized a generalized Onsager principle for learning system dynamics.
• Developed a platform to process microscopic trajectory data.
• Simultaneously constructed reduced thermodynamic coordinates and interpreted their dynamics.

Main Results:

• Successfully learned three interpretable thermodynamic coordinates for polymer stretching.
• Constructed a dynamical landscape of polymer stretching, identifying stable and transition states.
• Demonstrated control over the polymer stretching rate using the learned dynamics.

Conclusions:

• The developed AI platform effectively automates the discovery of macroscopic dynamics from microscopic data.
• The generalized Onsager principle provides a robust framework for learning interpretable reduced coordinates.
• This methodology has broad applicability across various scientific and technological domains.