Nonlinear GENERIC Informed Neural Networks (N-GINNs): learning GENERIC dynamics with non-quadratic dissipation potentials

This paper introduces Nonlinear GENERIC Informed Neural Networks (N-GINNs), a deep learning framework that enforces thermodynamic consistency through convex dissipation potentials to accurately discover evolution equations for systems exhibiting both conservative dynamics and non-quadratic dissipation.

The Gist

Imagine you are trying to teach a robot how to predict how a complex machine will move. You could just show the robot thousands of videos of the machine moving and let it guess the rules. But there's a problem: if the robot isn't careful, it might learn a rule that looks right for a few seconds but eventually breaks the laws of physics. It might invent a machine that creates energy out of thin air or one that gets colder while doing work, which is impossible in our universe.

This paper introduces a new tool called N-GINNs (Nonlinear GENERIC Informed Neural Networks). Think of this tool as a "physics safety harness" for artificial intelligence. Instead of letting the AI guess the rules freely, the researchers built the AI's brain so that it cannot break the fundamental laws of thermodynamics (energy conservation and entropy).

Here is a breakdown of how it works, using simple analogies:

1. The Two-Engine System

The paper focuses on systems that have two types of movement happening at once:

• The Reversible Engine (The Swing): Imagine a child on a swing. If there were no friction, they would swing back and forth forever. This is "conservative" motion. It's predictable and can go backward in time.
• The Irreversible Engine (The Friction): Now, imagine the swing has rusty hinges and air resistance. The swing slows down, and the energy turns into heat. You can't un-slow the swing. This is "dissipative" motion.

Most real-world machines (like car brakes, chemical reactions, or even your muscles) are a mix of both. The challenge for AI is to learn how to balance these two engines perfectly.

2. The "Safety Harness" (The Architecture)

The researchers created a special neural network architecture. Imagine building a car where the engine is designed so that it physically cannot produce more energy than you put into the gas tank.

• The "Energy" and "Entropy" Maps: The AI learns two maps: one for the system's total energy and one for its disorder (entropy).
• The "Friction" Map: The AI also learns a "dissipation potential." In simple terms, this is a map that tells the system how much energy turns into heat.
  • The Innovation: Previous AI models could only learn simple, straight-line friction (like a constant brake). This new model can learn complex, non-linear friction. Think of it like learning that a car's brakes work differently when they are cold versus when they are red-hot. The paper calls this "non-quadratic dissipation," which just means the friction rules can be curvy and complicated, not just straight lines.

3. The "Lock and Key" (The Constraints)

To make sure the AI doesn't cheat, the researchers built "locks" into the code.

• The Energy Lock: The code is written so that the "Reversible Engine" and the "Friction Engine" cancel each other out perfectly regarding total energy. The AI is forced to keep the total energy constant (unless heat is added from outside).
• The Entropy Lock: The code forces the "Friction Engine" to always generate heat (entropy). It is mathematically impossible for the AI to make the system get more ordered without an external push.

4. The Three Tests

The team tested this "safety-harnessed" AI on three very different scenarios to prove it works:

• Test 1: The Bouncing Ball in a Hot Room.
  A simple spring bouncing up and down while losing energy to a heat bath. This was the "easy" test to show the AI could learn standard physics.
• Test 2: The Chemical Motor.
  Imagine a piston (like in a car engine) filled with gas that is also undergoing a chemical reaction (like mixing baking soda and vinegar). The gas pushes the piston, but the chemical reaction creates complex, non-linear friction. This was a hard test because the rules were curvy and complicated. The AI successfully learned the rules.
• Test 3: The Stretching Metal.
  Imagine a metal bar being stretched. It behaves like a spring at first, but if you pull hard enough, it permanently deforms (plasticity) and heats up. This involves a whole sheet of metal moving, not just a single point. The AI learned how to predict the stretching, the permanent bending, and the heating all at once.

The Bottom Line

The paper claims that N-GINNs can look at data from these complex systems and figure out the exact mathematical rules governing them, while guaranteeing that the rules obey the laws of thermodynamics.

It's like giving a student a math test where they have to solve a problem, but the test paper itself has a built-in calculator that refuses to let them write down an answer that violates the laws of arithmetic. The result is a model that is not only accurate but also trustworthy because it is physically impossible for it to be "wrong" about the fundamental laws of energy and heat.

Technical Summary

Technical Summary: Nonlinear GENERIC Informed Neural Networks (N-GINNs)

Problem Statement
Data-driven discovery of governing equations often struggles to maintain compatibility with fundamental physical principles, particularly in computational mechanics where long-time predictions require adherence to conservation laws and entropy production constraints. While existing structure-preserving machine learning methods have successfully modeled reversible Hamiltonian dynamics (e.g., via Hamiltonian Neural Networks) and some dissipative systems, a significant gap remains in discovering the full set of building blocks for the General Equation for Non-Equilibrium Reversible–Irreversible Coupling (GENERIC) formalism. Specifically, current frameworks often rely on linear friction matrices or soft-constrained penalties, failing to robustly handle systems governed by general, potentially non-quadratic dissipation potentials. Such non-quadratic potentials are essential for modeling complex irreversible mechanisms like chemical reactions, jump processes, and plasticity, where Gaussian fluctuation assumptions fail.

Methodology
The authors propose Nonlinear GENERIC Informed Neural Networks (N-GINNs), a deep learning framework designed to discover evolution equations for systems governed by the nonlinear GENERIC formalism. The core evolution equation is given by:
$$\dot{x} = L(x)D_x E(x) + D_{x^*} \Xi(x, x^*)|_{x^*=D_x S(x)}$$
where $x$ is the state vector, $L(x)$ is the Poisson bivector, $E(x)$ is the energy functional, $S(x)$ is the entropy functional, and $\Xi(x, x^*)$ is the dissipation potential.

The methodology distinguishes itself by enforcing thermodynamic constraints by construction (hard constraints) rather than through penalty terms in the loss function. The architecture integrates the following components:

1. Reversible Dynamics (Hamiltonian Part): The framework learns the energy $E(x)$, entropy $S(x)$, and the bivector $L(x)$. To satisfy the reversible degeneracy condition $L(x)D_x S(x) = 0$ and the skew-symmetry of $L$, the authors employ a reparameterization strategy. A raw neural network output $\tilde{L}(x)$ is projected using a matrix $Q_S$ constructed from the entropy gradient, ensuring $L(x) = Q_S^T \tilde{L} Q_S$ is skew-symmetric and annihilates the entropy gradient.
2. Irreversible Dynamics (Dissipative Part): The dissipation potential $\Xi(x, x^*)$ is modeled using a Partially Input Convex Neural Network (PICNN). To satisfy the irreversible degeneracy condition $\Xi(x, x^* + \lambda D_x E(x)) = \Xi(x, x^*)$ and ensure $\Xi(x, 0)=0$ with convexity, the raw PICNN output is reparameterized via an orthogonal projection $P(x)$ onto the subspace orthogonal to $D_x E(x)$. The final potential is defined as $\Xi(x, x^*) = \tilde{\Xi}(x, P(x)x^*) - \tilde{\Xi}(x, 0) - D_{x^*}\tilde{\Xi}(x, 0)P(x)x^*$.
3. Thermodynamic Guarantees: This construction guarantees the first law (energy conservation) and the second law (non-negative entropy production) exactly, without requiring the network to learn these properties from data. The authors note that while the Jacobi identity (required for the Poisson structure to be a valid bracket) is not explicitly enforced by construction in high-dimensional settings due to complexity, it is not strictly necessary for thermodynamic admissibility.

Key Contributions

• First Framework for Non-Quadratic GENERIC: This is the first framework capable of discovering the full set of GENERIC building blocks (bivector, energy, entropy, and dissipation potential) specifically within the dissipation-potential formulation, accommodating both quadratic and non-quadratic cases.
• Hard-Constraint Architecture: The paper introduces specific reparameterizations for the bivector and dissipation potential that mathematically guarantee compliance with the first and second laws of thermodynamics, moving beyond soft-constrained approaches.
• Statistical-Mechanical Consistency: By parameterizing the dissipation potential $\Xi$ rather than a friction matrix, the approach aligns with large deviation theory, where $\Xi$ relates to the rate function of fluctuations, offering a more parsimonious and statistically interpretable representation.

Results
The authors validate N-GINNs on three distinct examples of increasing complexity:

1. Harmonic Oscillator in a Heat Bath: A system with quadratic dissipation. The model accurately recovered the classical dynamics, maintaining energy conservation and positive entropy production even when integrated with a standard non-structure-preserving RK4 scheme.
2. Idealized Chemical Motor: A system involving a non-canonical Poisson bracket (piston motion coupled to gas) and a non-quadratic dissipation potential (chemical reaction kinetics). The model successfully captured the nonlinear coupling between mechanical motion and chemical reactions, maintaining thermodynamic consistency across the test set.
3. One-Dimensional Viscoplastic Model (Perzyna Type): A continuum system governed by partial differential equations (PDEs) with a non-quadratic dissipation potential associated with viscoplastic flow. The framework successfully learned the spatially distributed fields (displacement, momentum, strain, temperature) and the underlying thermodynamic structure, demonstrating scalability to PDE-based settings.

In all cases, the learned models achieved low relative $L_2$ errors on test trajectories and strictly satisfied the thermodynamic constraints (constant total energy for closed systems, strictly positive entropy production).

Significance and Claims
The paper claims that N-GINNs provide a robust, structure-preserving tool for discovering thermodynamically consistent models from data, particularly for systems where non-quadratic dissipation is critical. The authors emphasize that their approach ensures exact compliance with the first and second laws of thermodynamics by design, addressing a limitation in existing data-driven methods that may violate physical principles.

The authors remain modest regarding the current scope, acknowledging that:

• The method is currently a deterministic framework trained on prescribed trajectories, returning single rollout predictions.
• It assumes state variables are fully observed, whereas many real-world applications involve partial observations or internal variables.
• The exact enforcement of the Jacobi identity in high-dimensional settings remains an open challenge and was not explicitly assessed in the examples.
• Future work is needed to address scalability, uncertainty quantification, and the integration of automatic state variable identification.

The paper positions N-GINNs as a significant step toward bridging the gap between complex non-equilibrium thermodynamics and modern deep learning, enabling the discovery of interpretable, physically consistent models for a broader class of irreversible processes.