Abstract
We propose a novel approach for learning the evolution that employs differentiable neural networks to approximate the full GENERIC structure. Instead of manually choosing the fitted parameters, we learn the whole model together with the evolution equations. We can reconstruct the energy and entropy functions for the system under various assumptions and accurately capture systems behaviour for a double thermoelastic pendulum and a rigid body.
Introduction
In classical physics and engineering, modeling complex systems typically involves assuming a specific set of governing equations and fitting parameters like viscosity or friction. However, when equations are unknown or systems are too complex for standard formulas, traditional methods struggle. While standard "black-box" neural networks can predict motion, they often fail to respect fundamental physical laws, such as energy conservation or the second law of thermodynamics. This research introduces a machine learning model that is flexible enough to learn unknown dynamics while remaining strictly principled through the GENERIC framework.
The GENERIC Framework
The authors utilize the General Equation for Non-Equilibrium Reversible-Irreversible Coupling (GENERIC). This framework decomposes the evolution of a system into two distinct, thermodynamically consistent parts:
- Reversible (Hamiltonian) Dynamics: Driven by Energy ($E$) and a Poisson bivector ($L$).
- Irreversible (Dissipative) Dynamics: Driven by Entropy ($S$) and a friction matrix ($M$).
The governing equation is expressed as: $$\dot{\mathbf{x}} = \mathbf{L}(\mathbf{x}) \cdot \nabla E(\mathbf{x}) + \mathbf{M}(\mathbf{x}) \cdot \nabla S(\mathbf{x})$$
By using deep neural networks to approximate all four components ($L, M, E, S$) simultaneously, the model can reconstruct the underlying physics directly from data. To ensure physical realism, "degeneracy conditions" are integrated into the loss function, guaranteeing energy conservation during reversible motion and preventing entropy decrease.
Experimental Results
The model was evaluated on several simulated physical systems to test its ability to learn complex, non-canonical structures.
Rigid Body Dynamics
Rigid body dynamics were used to test if the model could learn non-canonical Poisson brackets and quadratic energy. The results demonstrated that the network successfully captured the system's structure, with training and validation losses vanishing quickly.
As shown in the figure above, the learned model (represented by the "learned" lines) tracks the ground truth components of the angular momentum ($m_x, m_y, m_z$) with remarkable precision over a long duration, demonstrating the stability of the GENERIC-informed approach.
Double Thermoelastic Pendulum
In the case of the double thermoelastic pendulum, the model had to learn the energy and friction components. The learned trajectories closely matched the ground truth, even for initial conditions not seen during training.
Key Insights
- Full Reconstruction: This work represents the first successful reconstruction of all four main components of the GENERIC evolution from data alone.
- Physics-as-Regularization: By including terms for the Jacobi identity and degeneracy conditions in the loss function, the neural network is forced to discover a coordinate system that behaves according to real physical laws.
- Energy Conservation: The learned models maintained energy conservation within an error margin of approximately 1%.
- Handling Chaos: While chaotic systems like the double pendulum eventually diverge due to inherent sensitivity, the model maintains the correct physical "character" of the motion throughout the simulation.
Conclusion and Future Work
The proposed method demonstrates that differentiable neural networks can effectively learn the underlying thermodynamic structure of physical systems. Future work will focus on implementing the Jacobi identity more robustly, improving long-term stability in chaotic regimes, and applying the framework to noisy, real-world experimental data.
Summary updated