Abstract:Machine learning, with its remarkable ability for retrieving information and identifying patterns from data, has emerged as a powerful tool for discovering governing equations. It has been increasingly informed by physics, and more recently by thermodynamics, to further uncover the thermodynamic structure underlying the evolution equations, i.e., the thermodynamic potentials driving the system and the operators governing the kinetics. However, despite its great success, the inverse problem of thermodynamic model discovery from macroscopic data is in many cases non-unique, meaning that multiple pairs of potentials and operators can give rise to the same macroscopic dynamics, which significantly hinders the physical interpretability of the learned models. In this work, we propose a machine learning framework, named as Statistical-Physics-Informed Neural Networks (Stat-PINNs), which further encodes knowledge from statistical mechanics and resolves this non-uniqueness issue for the first time. The framework is here developed for purely dissipative isothermal systems. It only uses data from short-time particle simulations to learn the thermodynamic structure, which can be used to predict long-time macroscopic evolutions. We demonstrate the approach for particle systems with Arrhenius-type interactions, common to a wide range of phenomena, such as defect diffusion in solids, surface absorption and chemical reactions. Stat-PINNs can successfully recover the known analytic solution for the case with long-range interaction and discover the hitherto unknown potential and operator governing the short-range interaction cases. We compare our results with an analogous approach that solely excludes statistical mechanics, and observe that, in addition to recovering the unique thermodynamic structure, statistical mechanics relations can increase the robustness and predictability of the learning strategy.

Deep Energy-Based Modeling of Discrete-Time Physics

Deep Hidden Physics Models: Deep Learning of Nonlinear Partial Differential Equations

Time-Continuous Energy-Conservation Neural Network for Structural Dynamics Analysis

Learning Deep Dissipative Dynamics

Neural Operators Meet Energy-based Theory: Operator Learning for Hamiltonian and Dissipative PDEs

Deep physical neural networks enabled by a backpropagation algorithm for arbitrary physical systems

Combining physics and deep learning to learn continuous-time dynamics models

Physics-Informed Regularization for Domain-Agnostic Dynamical System Modeling

Physics-enhanced Neural Networks in the Small Data Regime

Latent Space Energy-based Neural ODEs

Learning Neural Hamiltonian Dynamics: A Methodological Overview

Physics-Informed Neural State Space Models via Learning and Evolution

Physics-informed learning of governing equations from scarce data

Advanced discretization techniques for hyperelastic physics-augmented neural networks

A nonintrusive hybrid neural-physics modeling of incomplete dynamical systems: Lorenz equations

Statistical-Physics-Informed Neural Networks (Stat-PINNs): A Machine Learning Strategy for Coarse-graining Dissipative Dynamics

Discover Physical Concepts and Equations with Machine Learning

Physics-informed machine learning

Learning to Represent Mechanics via Long-term Extrapolation and Interpolation

OnsagerNet: Learning Stable and Interpretable Dynamics Using a Generalized Onsager Principle

DeepPhysics: a physics aware deep learning framework for real-time simulation