Deep learning of thermodynamic laws from microscopic dynamics.

Area of Science:

• Physics
• Thermodynamics
• Machine Learning

Background:

• Macroscopic thermodynamic laws govern systems with many particles.
• Understanding emergent laws from microscopic interactions is a fundamental challenge.

Purpose of the Study:

• To demonstrate that deep neural networks can infer macroscopic thermodynamic laws from microscopic data.
• To explore machine learning's potential in discovering emergent physical principles.

Main Methods:

• Generating molecular dynamics simulation data of gas particles in adiabatic processes.
• Training a deep neural network (DNN) to predict the temporal order of particle snapshot pairs.

Main Results:

• The DNN learned an order relation consistent with adiabatic accessibility and thermodynamic axioms.
• The network's internal representation effectively functioned as entropy.
• The DNN successfully generalized thermodynamic laws from microscopic simulations.

Conclusions:

• Machine learning can discover macroscopic physical laws from underlying microscopic data.
• This approach offers a new pathway for data-driven discovery in physics.
• Deep learning models can uncover emergent phenomena in complex systems.