acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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. The consequence is an operator that depends on state and a Jacobi identity that is forfeited, so the resulting structure is metriplectic-degenerate rather than metriplectic. Our approach closes this gap by taking degeneracy from the same place GENERIC itself takes it. For nonlinear transport, the reversible operator is the compatible Lie–Poisson pencil ; otherwise it is a constant Fourier multiplier, which is trivially Poisson. On an augmented state carrying a latent entropy density, the entropy is a Casimir of , so holds identically, with no projection required. A fixed mechanical quadratic, a learned gauge-free potential, and a convex internal energy together make up the energy. The friction operator satisfies pointwise, and its Onsager parity structure permits diffusion and damping while provably ruling out transport. For any parameters, skewness, positivity, both degeneracies, and the Jacobi identity (on the resolved band for the Lie–Poisson term) all hold to machine precision. Heat conduction and damped waves have closed-form friction operators that are representable exactly, the second law bounds physical energy under a checkable curvature condition, and exact discrete first and second laws follow from a discrete-gradient integrator. The model was tested on four PDEs in one and two dimensions with three backbones (FNO, Transolver, CNO): with the same backbone it wins 61 of 72 seed-level comparisons against unconstrained baselines, learns the exact transport and wave symbols, reproduces the true dissipation rate to within 13% on heat and Burgers, and dissipates nothing on advection. Whereas a constant- ablation isolates the price of exact Jacobi as the loss of Burgers, a learned-entropy ablation injects energy on every reversible–irreversible problem.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, structure-preserving learning, scientific machine learning, GENERIC, metriplectic dynamics, Poisson brackets, thermodynamically consistent machine learning, PDE surrogates, geometric integration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VbeAOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Poisson-GENERIC Neural Operators (PG-phys) to address the limitations of existing thermodynamically consistent neural operators (e.g., GENERIC-FNO). By augmenting the system state with a latent entropy density, the model makes entropy a Casimir invariant of the reversible Poisson operator, satisfying degeneracy conditions without global projections. The framework features an exact discrete-gradient integrator and an Onsager-structured friction operator that strictly excludes non-physical transport. The method demonstrates strong thermodynamic consistency across four PDEs in 1D and 2D settings.

Critical strengths

1.Theoretical Rigor: Successfully elevates the function-space thermodynamic structure from metriplectic-degenerate to fully metriplectic by elegantly treating entropy as a Casimir invariant.

2.Physical Fidelity: The Onsager parity structure strictly restricts the friction operator to diffusion and damping, mathematically preventing unphysical transport.

3.Exact Conservation: The proposed discrete-gradient integrator ensures exact compliance with the first and second laws of thermodynamics in discrete time, regardless of step size.

Critical weaknesses

1.High Computational Cost: Evaluating energy gradients and friction coefficients via automatic differentiation introduces a 25−40× training slowdown compared to unconstrained baselines.

2.Aliasing in the Lie-Poisson Term: The Jacobi identity exhibits significant residuals (0.17–0.78) on generic full-spectrum states due to aliasing, meaning nonlinear transport is strictly Poisson only on the resolved band.

3.Systematic Loss on Heat Equation: The PG-phys model systematically underperforms compared to unconstrained baselines on the heat equation across multiple seeds and dimensions.

Questions for authors

1.The training cost is reported to be 25−40× slower than baselines due to multiple automatic differentiation evaluations. Does this computational bottleneck fundamentally limit the scalability of the proposed framework to 3D fluid dynamics or complex multi-physics systems, and are there specific strategies to mitigate this overhead?

2.Regarding the systematic performance loss on the heat equation: Does the strict structural separation between reversible and irreversible operators restrict the neural network's optimization landscape compared to an unconstrained baseline?

3.The implicit discrete-gradient step fails to converge on the trained 2D heat Transolver models due to the high Lipschitz constant of the global attention mechanism. Since this step is essential for exact discrete conservation laws, how do you propose to resolve this in practice (e.g., structural regularization, Newton-Krylov solvers) without degrading predictive accuracy?

4.The Jacobi identity shows significant residuals on generic full-spectrum states due to aliasing. Since real-world PDE solutions rarely possess perfectly band-limited spectra, how does this violation of the Poisson structure on unresolved modes impact the long-term stability and thermodynamic consistency of the model in practical scenarios?

Reviewer_aasdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a development upon GENERIC-FNO which adds the following

  • Enforces a certain Jacobi identity for reversible dynamics
  • Ensures latent entropy is conserved under reversible dynamics
  • Combines fixed mechanical and learned potential energies, rather than freely learning energy as previous papers have done
  • Adds a state-dependent friction
  • Replaces explicit time-stepping with a linear Fourier evolution preserving first and second law of thermodynamics

The authors show significant improvements over previous models on 1D and 2D PDEs across FNO, Transolver and CNO backbones.

Critical strengths

The authors' method is a natural and well-motivated improvement over GENERIC-FNO. In particular, the derivation of the architecture from the underlying Poisson and GENERIC structure is principled and sensible, introducing physically motivated inductive biases rather than arbitrary architectural constraints. The use of Casimir entropy, physically grounded energy, and structured dissipation is particularly well justified. These choices appear to substantially improve learning performance, in some cases by over an order of magnitude. In my view, this is a solid demonstration of how careful application of physics and thermodynamic theory can lead to better neural operators.

Critical weaknesses

Despite positioning the method as an improvement over GENERIC-FNO, the authors do not directly benchmark against it. Furthermore, the proposed model incorporates substantially stronger physics-informed inductive biases than the unconstrained baselines, making it difficult to isolate which architectural choices are responsible for the reported improvements.

The writing quality deteriorates substantially in Section 4. While I found the preceding sections generally followable, despite considerable field-specific jargon, Section 4 is excessively compressed and frequently reports numbers already visible in the tables without providing meaningful interpretation or discussion of their significance.

Combined, these two weaknesses make me quite unsure about the practical relevance of the author's work, despite their solid motivation.

Questions for authors

Could Table 1 be formatted and explained more clearly? I found it difficult to interpret the majority of rows, and consequently struggled to understand how the results demonstrate the broader advantages of the method.

Why is there no direct comparison against the original GENERIC-FNO? The learned-energy/entropy ablation differs from GENERIC-FNO in several respects, making it difficult to isolate the improvements due to this newly introduced method.

The impact of the Jacobi identity limitation is unclear. Could the authors explain how much this matters in practice, particularly for long-term stability?

Given the reported 25–40× increase in training cost, could the authors provide a comparison against baselines trained with comparable computational budgets?

Reviewer_w2ZaOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors introduce neural operators for PDEs that are thermodynamically consistent, using the exisiting solid thermodynamic model so-called "GENERIC" formalism . The authors point out that existing work (e.g., GENERIC-FNO) gets the degeneracy condition by projecting the reversible operator onto the complement of the entropy gradient. In this paper, the degeneracy comes from a Casimir instead and the reversible operator is a compatible Lie-Poisson models when nonlinear transport is needed. In the main part of the paper, authors argued that a latent entropy density is a Casimir of the reversible operator. Due to this property, the degeneracy condition for the reversible operator holds identically without any projection. The energy consists of a learned gauge-free potential, and a convex internal energy.The authors claim that important structural constratins including skewness, positivity, degeneracy conditions, and the Jacobi identity hold for any parameters with marginal errors. In the experimental section, they demonstrate that the proposed model outperforms all existing methods both quantitatively, and qualitatively (e.g., physical metrics).

Critical strengths

None

Critical weaknesses

Personally, I think that writing a paper with the help of AI is not a big problem. However, in this paper, it goes too far, and the paper lacks even the most basic content that an ML paper should have, such as what the core contribution is, what the proposed model looks like, and what objective function is defined to train the model. For this reason, the manuscript cannot be evaluated in its current state, and given its poorly written form, I recommend rejection based on the quality of the paper.

Equation (2), which seems to be related to the core model of the paper, is introduced directly without any appropriate formal mathematical definition, and it is impossible to understand what each object means. For example, when the method is introduced, definitions at the level of those in Lines 951-956 are needed together with (2). The content of Appendix A.1 to A.5 is given in parallel with the main text, and it redefines or explains in more detail the definitions that are vague and unclear in the main text. Would it not be much easier to read if this discussion were explained in the main text with sufficient mathematical formality in the first place?

The narrative of this paper is very dependent on GENERIC-FNO. This paper also has the same problem, but in the main text, the authors should give a formal introduction covering the reason for introducing the GENERIC model, its narrative, its historical context, its role in ML, and related literature, together with a clear punchline on which part they want to fix. In this way, they should make the core contribution of this paper, the preservation of the Poisson property compared with GENERIC-FNO, stand out among the many undefined mathematical objects.

The limitation and scope section is important, but it could be moved further back. I would like the authors to revise the paper to present the related discussion of the Poisson structure in the main text as propositions or theorems, to help readers understand.

Questions for authors
  1. Compared with existing GENERIC-based NN parameterization methods, how is Eq. (2) for the energy in the main text different, and what are its core motivation and intuition?

  2. Please show, as a proposition connected to (2), how the Lie-Poisson formalism is involved in the proposed method.

  3. What is meant by “softplus network minus its degree-≤ 2 polynomial part”? In the degree ≤ 2 polynomial part, polynomial in what? Why is softplus used, what is the reason for the minus, and why is this whole statement valid? For what reason was it modeled this way?

  4. In “A quadratic hidden in φθ\varphi_\theta would otherwise absorb part of Q”, quadratic in what, and what is meant by “absorbing” part of Q?

Questions 3 and 4 point out only a few sentences at the beginning of the paper, but in fact, this kind of incomprehensible sentence repeats endlessly throughout the whole paper. Please fix all of them.

1 comment

I’m very sorry to say, but I think the rebuttal is also heavily AI-generated, and I don’t think the provided results answer my questions sincerely. I will keep my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.