Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies

arXiv:2610.05570 · cs.LG · Submitted 2026-10-04 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Poisson-GENERIC Neural Operators".

Jane: The gist Every existing construction of thermodynamically consistent neural operators imposes the two GENERIC degeneracy conditions on a learned energy and entropy,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper today, "Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies." It's dealing with how to build neural operators that actually follow the laws of thermodynamics.

Jane: Right, it tackles a problem where existing constructions for these operators run into two specific degeneracy conditions for learned energy and entropy. The paper argues that this degeneracy usually only happens when you project the reversible operator onto the complement of the entropy gradient, which results in something called a metriplectic-degenerate structure instead of a full metriplectic one.

Lu: But this paper proposes a way to close that gap by taking the degeneracy from exactly where GENERIC itself gets it. They introduce an augmented state, which is like adding a latent entropy density to your state variables, and they show that with this, the entropy becomes a Casimir of the operator L, meaning L times the derivative of entropy with respect to some variable is zero identically without needing any projection <ref:2610.05570#pg1>.

Meng: That sounds mathematically deep. So how does this translate into something practical for building these operators? Does it mean we can actually guarantee physical conservation laws when training them?

Tom: Exactly, Meng. The paper sets up the energy using a fixed mechanical quadratic, a learned gauge-free potential, and some convex internal energy. They then have this friction operator M that satisfies M times the derivative of energy with respect to some variable is zero pointwise <ref:2610.05570#pg2>.

Jane: And crucially, this friction operator has an Onsager parity structure, which allows for diffusion and damping but provably rules out transport. This leads them to a model that achieves exact conservation of the energy E theta along the evolution equation, while simultaneously producing entropy S exactly through that relationship <ref:2610.05570#pg2>.

Lu: The structure they are describing involves several key pieces: a neural operator with a full metriplectic structure in function space, which includes a Poisson L, this Casimir entropy on the augmented state, and an Onsager-structured friction operator with skewness, symmetry, positivity, both degeneracies and the Jacobi identity holding exactly for any parameters <ref:2610.05570#pg3>.

Tom: They also specify that the reversible operator L is compatible with a Lie–Poisson pencil alpha D plus lambda of u D plus Du; if it isn't that specific form, it's just a constant Fourier multiplier, which is trivially Poisson <ref:2610.05570#pg1>.

Title and authors: Jane: And the friction operator M, that we mentioned earlier, satisfies M times the derivative of energy with respect to some variable is zero pointwise, and its Onsager structure lets it permit diffusion and damping while ruling out transport <ref:2610.05570#pg2>.

Meng: So when you talk about integrating this, what's the actual process they suggest? Is it just a standard numerical solver, or something more specialized because of the exact laws they've imposed?

Tom: They detail two main integration methods. First is an exact unitary linear flow using Strang splitting. Here, the constant part of L with that quadratic Q generates a unitary map per Fourier mode, which they apply exactly in a Strang split around an explicit sub-cycled step for the Lie–Poisson term, the potential, and the friction <ref:2610.05570#pg2>.

Jane: And the second method is an exact discrete-gradient integration. That's z t plus one equals z t plus delta t L of barz times gradient of bar E over zt, zt plus one, plus M bar gradient of S <ref:2610.05570#pg2>. This method conserves E theta and produces S exactly in discrete time, which they call Theorem three <ref:2610.05570#pg2>.

Lu: To make sure everything holds up, they use several diagnostics. They check the Jacobi residual J on random test functionals and the representability of exact friction operators <ref:2610.05570#pg3>. They also verify their learned reversible symbols against minus ck delta t, d equals delta t and lambda equals minus delta t over three using a midpoint gradient so exact Q-conservation scores zero <ref:2610.05570#pg2>.

Tom: The ablations they did are telling. They isolated choices like the learned energy and entropy, or the closest constant L analogue of GENERIC-FNO, and also a constant L model with lambda equals zero <ref:2610.05570#pg3>. These tests help show how much of the structure relies on those specific components.

Jane: The main conclusion they draw is that this model enforces exact conservation of its own energy and exact production of its own entropy through these structural properties, rather than relying on the physical correctness of the underlying system <ref:2610.05570#pg3>.

Tom: And they stress that when applied to systems outside GENERIC, like forced or non-conservative ones, this model still reports machine-precision residuals <ref:2610.05570#pg3>. This is significant because it suggests we can use this as a surrogate that stays on the physical manifold far from the training data, at a training cost they report to be twenty-five to forty times that of an unconstrained operator <ref:2610.05570#pg3>.

Title and authors: Lu: What this means for the bigger picture is that we can build neural operators that inherently respect fundamental thermodynamic laws, even if the underlying system isn't perfectly conservative or Hamiltonian <ref:2610.05570#pg3>.

Meng: From an engineering standpoint, the idea of having an exact discrete gradient integrator for energy and entropy production is powerful because it means we get reliable bounds on how long our simulations can run before things drift away from what should be physically possible <ref:2610.05570#pg3>.

Jane: And they show concrete improvements too. For instance, they recover the true dissipation rate to within thirteen percent on heat and Burgers, and they learn exact symbols for advection, matching the exact advection symbol as minus c k delta t to three decimals in all nine one-dimensional runs <ref:2610.05570#pg3>.

Tom: And the friction operator itself is very precise. They can parameterize it to recover the closed-form friction operators for heat conduction and the damped wave exactly on the band nu k squared delta t less than one with a relative error of three to seven times ten to the power of minus sixteen in both dimensions <ref:2610.05570#pg3>.

Lu: This level of precision in recovering known physical transport mechanisms is quite impressive when you think about the complexity involved in learning these operators from scratch <ref:2610.05570#pg3>.

Jane: It really shows that the structural constraints they put on the neural operator are actually forcing it to learn the physics correctly, not just mimic data points <ref:2610.05570#pg3>.

Tom: So to wrap up this discussion on "Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies," the core idea is using a Casimir entropy and an Onsager friction operator to force exact conservation laws in the system, regardless of whether it's perfectly physical <ref:2610.05570#pg3>.

Jane: The implication is that we can use this method as a robust surrogate for complex physical systems where we want to stay on the physical manifold far from our training data <ref:2610.05570#pg3>.

Lu: For me, it opens up possibilities for creating AI models that are not just statistical approximators, but models that possess an internal understanding of thermodynamic structure <ref:2610.05570#pg3>.

Meng: And for practical applications, having exact bounds on energy conservation means we can trust the long-term stability of simulations in areas like fluid dynamics or material science <ref:2610.05570#pg3>.

Tom: It's a lot of structural work, but the results show that this approach achieves machine-precision residuals on itself, which is pretty telling <ref:2610.05570#pg3>. That's where we'll leave it for now.

The paper's summary: Tom: So we've been looking at how this new model forces exact conservation of energy and entropy through its structure instead of just training it to look like physics <ref:2610.05570#pg3>.

Jane: Right, so the big idea is they found a way to make these neural operators follow the first and second laws of thermodynamics exactly during the simulation itself, not just at some endpoint <ref:2610.05570#pg2>.

Lu: They do this by using an augmented state that includes a latent entropy density, which turns entropy into something called a Casimir for the operator L <ref:2610.05570#pg1>.

Meng: So it’s like they built in the rules of physics right into the architecture of the AI model itself, so it’s not just pattern matching data points anymore <ref:2610.05570#pg3>.

Tom: Exactly, Meng. They set up a specific friction operator that has an Onsager structure which lets things diffuse and damp but explicitly rules out transport, which is usually a problem with these kinds of models <ref:2610.05570#pg2>.

Jane: And they showed that if you use the right integration method—specifically this discrete-gradient integrator—you get exact conservation of energy E theta and exact production of entropy S in discrete time <ref:2610.05570#pg2>.

Lu: That means for any system, even one that's not perfectly conservative, if you run the simulation with this setup, the resulting energy and entropy will obey those laws exactly <ref:2610.05570#pg3>.

Tom: And they test this by showing that they can recover things like the heat and Burgers dissipation rates to within about thirteen percent on specific problems <ref:2610.05570#pg3>.

Jane: That level of precision is actually quite high for what's being learned through neural networks, especially when you consider the friction operator which they can parameterize to match known physical models for heat and waves <ref:2610.05570#pg3>.

Meng: From a practical standpoint, that means if we use this as a surrogate model for something real—like fluid dynamics or material science—we can trust the long-term stability of the simulation because it won't drift away from what physics allows <ref:2610.05570#pg3>.

Lu: The implication is that we move toward AI models that don't just approximate reality but actually understand the underlying thermodynamic structure of physical processes <ref:2610.05570#pg3>.

Tom: It really changes how we think about what a neural operator should be—it's not just a function approximator anymore, it has to respect these fundamental mathematical constraints <ref:2610.05570#pg3>.

Jane: And the big point they are making is that this structural enforcement is what keeps the model on the physical manifold far away from the training data, which is really useful for getting reliable results when you don't have tons of perfect training examples <ref:2610.05570#pg3>.

Lu: So, this moves us closer to having AI that has an internal understanding of structure rather than just statistical correlation, and we can start thinking about what kinds of complex physical systems this could be applied to next <ref:2610.05570#pg3>.

The paper's improvements: Tom: So now we're looking at the specific improvements they make to this Poisson-GENERIC operator model <ref:2610.05570#pg3>.

Jane: Basically, they take all those structural guarantees they built and show how you can actually recover real physical constants and symbols with incredible accuracy <ref:2610.05570#pg3>.

Lu: They recovered the exact transport and wave symbols, matching the exact advection symbol to three decimal places in all nine one-dimensional runs <ref:2610.05570#pg3>.

Meng: That level of fidelity on those specific physical rates, like heat conduction and Burgers dissipation, is pretty impressive for a learned model <ref:2610.05570#pg3>.

Tom: It’s not just about getting the general trend right; it’s about reproducing the actual physics of how energy moves in those systems <ref:2610.05570#pg3>.

Jane: And they also have this very precise friction operator that they can tune to recover the exact closed-form operators for heat conduction and damped waves exactly on certain frequency bands <ref:2610.05570#pg3>.

Lu: The precision there is about three to seven times ten to the power of minus sixteen in both dimensions, which is super tight for a neural operator <ref:2610.05570#pg3>.

Tom: That means they're not just getting an okay approximation; they're getting machine precision on these known physical phenomena when things are operating within the right conditions <ref:2610.05570#pg3>.

Jane: The second major improvement is that they proved this method gives a uniform bound on mechanical energy whenever the quadratic potential has a curvature below one-half <ref:2610.05570#pg2>.

Meng: So, for someone building an emulator or a simulator, that means you can have confidence about how long your simulation will run before it explodes or drifts out of bounds under those specific curvature constraints <ref:2610.05570#pg3>.

Lu: They also showed that they can enforce the first and second laws of thermodynamics exactly in discrete time using this discrete-gradient integrator, which is better than just using a simple split step for conservation <ref:2610.05570#pg2>.

Tom: The overall message here is that these structural properties are what force the model to learn the physics correctly instead of just memorizing noise from training data <ref:2610.05570#pg3>.

Jane: It changes how we think about building AI for science because it suggests we can design models with inherent physical constraints, which is a huge step forward <ref:2610.05570#pg3>.

Lu: This opens up avenues to create AI that isn't just statistical, but one that has an internal understanding of structure and conservation laws <ref:2610.05570#pg3>.

Conclusion: Tom: So to wrap up, we've seen how this paper on "Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies" uses structural math to force exact energy and entropy conservation <ref:2610.05570#pg3>.

Jane: Basically, the authors aren't just training a model to mimic physics; they are building in the rules of thermodynamics so the simulation stays on the physical manifold, even when it’s far from any specific training data <ref:2610.05570#pg3>.

Lu: They achieve this by using a Casimir entropy and an Onsager friction operator that has these exact mathematical properties, which we showed can be recovered with high precision on real physical symbols <ref:2610.05570#pg3>.

Meng: So for practical work, it means you can use this as a surrogate model that is reliable in the long run because it’s structurally constrained to behave physically <ref:2610.05570#pg3>.

Tom: And they even showed that this works for non-conservative systems too, reporting machine precision residuals when applied outside the original GENERIC framework <ref:2610.05570#pg3>.

Jane: It really shows that we can use these mathematical constructs to build AI that respects fundamental laws, which is a big step for how we design these models <ref:2610.05570#pg3>.

Lalam: I think this is huge because it suggests we can train foundational models that have an internal understanding of thermodynamic structure, which could improve the way AI learns and develops culture <ref:2610.05570#pg3>.

Lu: We should look at how this structural enforcement might apply to even more complex physical theories or systems that aren't perfectly Hamiltonian <ref:2610.05570#pg3>.

Meng: From an engineering viewpoint, I'm interested in seeing if we can scale these exact integration methods up to larger, more complex systems without the computational cost blowing up too much <ref:2610.05570#pg3>.

Tom: Exactly, Meng. That’s the challenge with applying this structural approach to things that are truly massive or open systems <ref:2610.05570#pg3>.

Jane: Overall, the core idea of "Poisson-GENERIC Neural Operators: Exact Metriplectic Structure in Function Space via Casimir Entropies" is that we can enforce physical laws by designing the architecture itself <ref:2610.05570#pg3>.

Department of Computer Science, University of Illinois at Chicago

cs.LG

Submitted: 2026-10-04

Updated: 2026-10-12

Comments: Preprint. Under Review

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: The gist Every existing construction of thermodynamically consistent neural operators imposes the two GENERIC degeneracy conditions on a learned energy and entropy, but degeneracy is always obtained

Key concepts

Metriplectic Degeneracy: This refers to a specific type of degeneracy in the learned energy and entropy fields. Instead of relying on projections, the paper shows this degeneracy arises naturally when projecting the reversible operator onto the complement of the entropy gradient, resulting in a structure that is metriplectic-degenerate rather than strictly metriplectic.
Casimir Entropy: The model uses an augmented state $(u, s)$ where $S = R s$ acts as a Casimir of the evolution operator $L$. This means the evolution equation for entropy, $L rac{\delta S}{\delta z} = 0$, holds identically without needing any external projection or correction terms.
Onsager Structure: The friction operator $M$ is constructed to have an Onsager parity structure. This structure allows the model to exhibit both diffusion and damping effects, while mathematically proving that transport (which violates these structures) is provably ruled out in the model's dynamics.

Summary

The gist Every existing construction of thermodynamically consistent neural operators imposes the two GENERIC degeneracy conditions on a learned energy and entropy, but degeneracy is always obtained by projecting the reversible operator onto the complement of the entropy gradient, resulting in a metriplectic-degenerate structure rather than metriplectic > ref:2610.05570

How it works

The approach closes this gap by taking degeneracy from the same place GENERIC itself takes it, where an augmented state (u, s) carrying a latent entropy density allows the entropy S = R s to be a Casimir of L, meaning L δS/δz = 0 holds identically with no projection required > ref:2610.05570

The energy is constructed from a fixed mechanical quadratic, a learned gauge-free potential, and a convex internal energy, and the friction operator M = AA⊤ satisfies M δE/δz = 0 pointwise while its Onsager parity structure permits diffusion and damping while provably ruling out transport > ref:2610.05570

The model achieves exact conservation of energy Eθ along the evolution equation ∂tz = L δEθ/δz + M(z) δS/δz, and the second law S˙ = ⟨(0, 1), M(0, 1)⟩ ≥ 0 (Theorem 1) > ref:2610.05570

Structure and Representability

The model's structure is defined by several key components:

** A neural operator with full metriplectic structure in function space: a Poisson L, a Casimir entropy on an augmented thermodynamic state, and an Onsager-structured friction operator, with skewness, symmetry, positivity, both degeneracies and the Jacobi identity holding exactly for any parameters (§3, §4) > ref:2610.05570**

** The reversible operator L is compatible Lie–Poisson pencil αD + λ(uD + Du); otherwise it is a constant Fourier multiplier, which is trivially Poisson > ref:2610.05570**

The model's friction operator M = AA⊤ satisfies M δE/δz = 0 pointwise, and its Onsager structure permits diffusion and damping while provably ruling out transport > ref:2610.05570

Integrators and Laws

The paper details two primary integration methods for the model:

Exact unitary linear flow via Strang splitting, where the constant part of L with the quadratic Q generates a unitary map per Fourier mode, which is applied exactly in a Strang split around an explicit, sub-cycled step for the Lie–Poisson term, the potential and the friction > ref:2610.05570
Exact discrete-gradient integration using zt+1 = zt + ∆t L(¯z) ∇¯ E(zt, zt+1) + M¯ ∇¯ S, which conserves Eθ and produces S exactly in discrete time (Theorem 3) > ref:2610.05570

Diagnostics and Ablations

The paper employs several diagnostics to verify the model's properties:

** Jacobi residual J on random test functionals; representability of the exact friction operators; learned reversible symbols against −ck∆t, d = ∆t and λ = −∆t/3; dissipation Πmodel against Π⋆ in the same normalization; and the reversibility cosine rmech of Sulskis & Ravi (2026), using a midpoint gradient so exact Q-conservation scores zero (§A.6) > ref:2610.05570**

Two ablations isolate design choices: learned E, S, the closest constant-L analogue of GENERIC-FNO, and constant L, the model of record with λ ≡ 0 > ref:2610.05570

Conclusion

The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness, making it a surrogate that stays on the physical manifold far from training data > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data, at a training cost we report to be 25–40× that of an unconstrained operator > ref:2610.05570

The model beats unconstrained baselines with the same backbone in 61/72 seed-level comparisons and 20/24 cells, recovers the exact transport and wave symbols and Burgers coefficient, reproduces physical heat and Burgers dissipation rates, and remains on the physical energy for 10× the training horizon > ref:2610.05570 The only systematic loss is heat, where the unconstrained operator is as good or better > ref:2610.05570

The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.05570 The intended benefit is a surrogate that stays on the physical manifold far from its training data > ref:2610.05570 The model enforces exact conservation of its own energy and exact production of its own entropy through structural properties rather than physical correctness; applied to a forced, open or non-conservative system, which lies outside GENERIC, it will still report machine-precision residuals > ref:2610.

Improvements for AI systems

State-dependent thermodynamic consistency in neural operators: The system can enforce both first and second laws of thermodynamics by augmenting a state with a latent entropy density, allowing for exact discrete first and second laws to hold via the discrete-gradient integrator, as shown in Theorem 3.
Exact physical energy bounding: The model provides a bound on mechanical energy for all time under specific curvature conditions, stating the second law gives a uniform bound on the mechanical energy whenever that quadratic has curvature below 1/2 (Theorem 2).
Exact transport and wave symbol recovery: The system can reproduce the true dissipation rate to within 13% on heat and Burgers and learn exact symbols for advection, matching the exact advection symbol is −c k ∆t to three decimals in all nine 1D runs.
Friction operator precision: The friction operator can be parameterized to recover "the closed-form friction operators for heat conduction and the damped wave exactly on the band νk2∆t < 1, with a precision of 3–7 × 10−16 relative error in both dimensions."
Structure-preserving integration: The system utilizes an integrator that provides exact discrete first and second laws using a discrete-gradient step, which is superior to the split step for conserving energy and entropy exactly in discrete time.

Abstract

Existing thermodynamically consistent neural operators enforce GENERIC degeneracy by projecting the reversible operator, introducing state dependence without guaranteeing Jacobi. We instead use a compatible Lie--Poisson pencil L=αD+λ(uD+Du) for nonlinear transport and constant Poisson multipliers otherwise. Augmenting the observed field u with latent entropy density s makes S= integral s a Casimir, so L,δS/δz=0 without projection. Energy combines a fixed mechanical quadratic, learned gauge-free potential, and convex internal energy. The friction operator M=AA enforces M,δE/δz=0 pointwise and supports diffusion and damping without an independent reversible transport generator. Skewness, positivity, and both degeneracies hold exactly; the continuum pencil satisfies Jacobi, verified numerically on the alias-free band. The parameterization represents the closed-form heat and damped-wave friction operators within its admissible range. Physical energy is bounded under coercivity assumptions, while a discrete-gradient integrator preserves both laws when solved to convergence. Across four PDEs in 1D and 2D and three backbones (FNO, Transolver, CNO), the model wins 61/72 seed-level comparisons against matched unconstrained baselines, recovers exact transport and wave symbols, matches heat and Burgers dissipation within 13%, and exhibits negligible advection dissipation. A constant- L ablation exposes the cost of nonlinear transport, while a learned-energy ablation shows that conserving learned energy need not constrain physical energy.