acceptodds
Under review as a conference paper at ICLR 2027

Physics-Induced Koopman Learning for Mesh-Based Dynamics

Abstract

Finite-dimensional Koopman learning seeks to represent nonlinear dynamics through linear evolution in learned observable coordinates, but a generic learned realization need not respect the physical structure of the target system. We characterize families of linear latent generators with finite interaction radius, consistency across isomorphic local environments, and orientation covariance, showing that each such family is uniquely represented by a shared matrix-valued kernel on local interaction types. When the set of local interaction types is finite, this yields a finite expansion in orientation-graded relation operators with uniquely determined coefficients. We instantiate this reduction as Koopman-MeshFT, combining local covariant encoders and decoders with a growth-controlled continuous-time generator. Across four public PDE benchmarks, Koopman-MeshFT attains the lowest mean rollout error among the evaluated Koopman baselines and improves several benchmark-specific physical diagnostics over them, while remaining competitive with FNO as a nonlinear neural-operator reference. The learned linear realization also enables non-iterative state estimation and forecasting from sparse observations. These results establish a physics-induced hypothesis class for finite-dimensional Koopman learning without assuming exact finite-dimensional closure.

Primary area:
learning on time series and dynamical systems
Keywords:
Koopman learning, Operator learning, Scientific machine learning, Mesh-based dynamics, Mesh Field Theory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SsJyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a framework for learning finite-dimensional Koopman operators for mesh-based physical systems. The core contribution is a theoretical reduction of the hypothesis class of linear latent generators by enforcing four physical principles: (L) locality, (P) local-law consistency under mesh isomorphisms, (O) orientation covariance, and (E) storage non-increase. The authors instantiate this theory as Koopman-MeshFT, adding a "growth-controlled latent envelope" to suppress the amplification of closure errors over time. The approach is empirically evaluated on four public PDEBench datasets (Advection, Burgers, CNS1D, CNS2D), where it outperforms evaluated Koopman-based baselines (KNO, cKAE) in rollout accuracy and physical fidelity metrics.

Critical strengths

The approach is shown to empirically outperform models like FNO and KNO.

The experimental section is rigorous.

Critical weaknesses

This is a very poorly written paper, with heavy use of jargons, many of which, I am unable to find in reference literature. Terms are introduced before defining them. Key terms have no definitions. The abstract is very hard to make sense of, due to too many jargons. The paper relies on specific terminology that often lacks formal definitions or external grounding.

The authors repeatedly discuss "exact finite-dimensional closure" and "finite-dimensional Koopman closure" without providing a formal mathematical definition or a direct citation establishing the threshold for this closure within their specific mesh-based framework.

The concept of a "growth-controlled subfamily" is used to regulate the amplification of closure errors, yet it is presented without a direct citation to prior literature utilizing this exact control mechanism in Koopman latent spaces.

the paper utilizes a "substepped Taylor action". The authors use this phrase as a given without citing a standard numerical methods text for this specific nomenclature.

The optimization procedure includes a "manifold-cycle residual" alongside data-closure and rollout losses. While structurally clear in the text, the specific term lacks references linking it to similar cycle-consistency mechanisms in broader autoencoder literature.

Questions for authors

"orientation-graded relation operators with uniquely determined coefficients" -- please clarify these phrases. This is a jargon-heavy sentence and I could not find any reference to "orientation-graded relation"

"not every finite-dimensional linear dynamics is physically equally plausible" - Either provide citations or provide examples "Instantaneous interactions are typically local" -- what is an instantaneous interaction?

"the physical law should not depend on arbitrary relabeling of mesh entities, oriented quantities should transform consistently with orientation conventions, and distinguished physical storage quantities may be conserved or dissipated." -- what physical law? What's relabeling? What's an oriented quantity? What's a distinguished physical storage?

In Section 3.1, What is F(x)? Apart from mentioning it's Locally lipschitz, there's no description of what it actually maps.

Reviewer_UpiXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Koopman-MeshFT, a physics-informed neural Koopman framework for predicting nonlinear physical dynamics. The authors derive a structured linear dynamics operator from assumptions about locality, local consistency, and orientation covariance, and introduce growth constraints to improve stability. The model learns an encoder and decoder around this linear operator and is evaluated on four PDE benchmarks. Results show improved prediction accuracy and physical consistency over existing Koopman-based baselines, with competitive performance against FNO.

Critical strengths
  1. The paper provides a mathematical characterization of the structured linear dynamics operator, deriving its form from physical principles rather than relying solely on architectural design choices.

  2. The proposed method consistently outperforms the evaluated Koopman-based baselines across four PDE benchmarks, demonstrating the effectiveness of the proposed formulation.

Critical weaknesses
  1. It seems like the proposed architecture and constraints on the linear dynamics operator are closely related to existing approaches such as piDMD. While the mathematical characterization is valuable, the practical novelty beyond previous structured Koopman methods is unclear.

  2. Despite outperforming existing Koopman baselines, the method generally falls short of FNO in prediction accuracy. It remains unclear what practical advantages the proposed linear latent dynamics provide over nonlinear alternatives, especially without clear evidence of improved inference speed or broader generalization.

Questions for authors
  1. Could the authors clarify the key theoretical and practical contributions beyond piDMD and other existing physics-constrained Koopman methods?

  2. Could the authors provide further experiments demonstrating the practical benefits of the proposed linear latent dynamics over nonlinear alternatives, particularly in terms of computational efficiency, generalization, or long-term prediction?

Reviewer_Mt1POpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Koopman-MeshFT, a structured Koopman learning framework for mesh-based dynamics. Under locality, local-law consistency, and orientation covariance, the method characterizes local linear generators through shared relation kernels and introduces growth-controlled latent dynamics. Experiments on four PDE benchmarks show improvements over existing Koopman baselines and demonstrate applications to sparse-observation forecasting.

Critical strengths

The main strength of this paper is its efficient structured modeling. Koopman-MeshFT achieves competitive predictive performance with relatively few parameters and supports efficient state estimation from sparse observations. These results suggest practical benefits from the proposed structured representation.

Critical weaknesses
  1. Theoretical novelty and informativeness of the guarantees.

The theoretical contribution appears limited relative to the claims. Theorem 1 largely follows from the assumed local-law consistency and resembles existing characterizations of local parameter sharing and equivariant operators. The paper does not sufficiently establish what new theoretical insight this characterization provides.

Moreover, the finite-time influence bound in Eq. (3) and the prediction-error bounds in Eqs. (49)–(53) rely on elementary matrix exponential estimates and standard variation-of-constants arguments. The resulting bounds appear loose and provide limited insight into practical prediction accuracy. Taken together, the theoretical results are not sufficiently connected to establish a substantive new guarantee.

  1. Justification and validation of the physics-induced claims.

The connection between the proposed generator structure and the underlying physics remains insufficiently justified. The constraints are imposed on the latent generator rather than rigorously derived from the physical dynamics. In particular, the growth-controlled parameterization does not guarantee physical conservation and excludes canonical spatial diffusion operators. These limitations make the scope of the physics-induced interpretation unclear.

The experimental evidence also leaves important aspects of the framework untested. The main benchmarks use periodic grids with trivial orientation grading, so they do not sufficiently validate the generality or necessity of the proposed structure. Further justification is needed to distinguish its benefits from those of standard local parameter sharing and growth control.

Questions for authors
  1. Could the authors clarify the theoretical novelty of Theorem 1 relative to existing characterizations of locally shared equivariant operators? Which insight goes beyond the assumed local-law consistency?

  2. Could the authors provide tighter or more informative error bounds that characterize long-horizon prediction accuracy? In particular, what can be established beyond the standard matrix exponential and Lipschitz estimates used in the current analysis?

  3. Could the authors provide theoretical or empirical evidence that the generator structure is induced by physical principles rather than imposed as an architectural constraint? It would also be helpful to isolate its benefits from those of standard local parameter sharing and growth control.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.