acceptodds
Under review as a conference paper at ICLR 2027

Stabilizing Neural Emulation of Spatiotemporal Physical Dynamics through Denoising Koopman Latent Dynamics

Abstract

Koopman operator theory provides a principled framework for reduced-order modeling of nonlinear spatiotemporal dynamics through linear evolution in an observable space. In practice, however, learned finite-dimensional Koopman representations are typically approximately invariant, causing unresolved dynamics and accumulated closure errors. Recent approaches, motivated by the Mori-Zwanzig formalism, address these unresolved effects through deterministic memory or closure modeling, yet robust extrapolation beyond the training horizon remains challenging. Here, we propose KoopFlow, an efficient prediction–correction framework built on two key insights: First, we identify a non-monotonic relationship between Koopman spectral coherence and predictive accuracy, revealing a trade-off between long-term information preservation and transient dynamics fidelity. Second, we reinterpret finite-dimensional Koopman error accumulation as a progressive distributional degradation process. Accordingly, KoopFlow treats Koopman predictions as noisy intermediate estimates and leverages a pretrained flow matching model as a probabilistic transport bridge toward the target latent distribution. By coupling structured linear Koopman prediction with generative correction, KoopFlow mitigates accumulated errors while retaining computational efficiency and temporal flexibility. Experiments across diverse nonlinear dynamical systems demonstrate improved long-horizon prediction accuracy and stability over representative baselines, particularly beyond the training horizon.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatiotemporal Physical Systems, Koopman Learning, Generative Model

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_WTd1OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper is concerned with the challenge of forecasting from historical data. The authors choose the Koopman operator formalism for describing spatiotemporal dynamics, and address the issues that arise from approximating this operator in finite dimensions. They note that forecasts from different start states may end at different locations despite deterministic underlying dynamics due to the propagation of errors, resulting in a distribution of degraded states that is a mixture of Gaussians.

They then introduce a procedure for correcting these errors, drawing on flow matching to learn the reverse stochastic process corresponding to the degradation. They demonstrate that this approach improves forecasting performance, both on train data and out of distribution data, compared to a variety of methods. They further analyze the effects of their hyperparameter choices (Koopman training horizon, propagation window, correction starting point) to assess and justify their experimental choices.

Critical strengths

This paper was well-written and very easy to read and understand the author's motivations and problem formulation. The primary strengths are:

  • Well-grounded motivation, and a clever application of flow-matching to an important problem: data-driven approximation and forecasting of nonlinear, deterministic dynamics. Section 3.3 in particular sets the framing for the paper well, providing a solid justification for using flow matching to reverse the progressive degradation.
  • A thorough set of experiments demonstrating the efficacy of this framework, especially on test set forecasting. The experiments exploring the effects of various hyperparameter choices and inference times are helpful as well.
Critical weaknesses

The main weaknesses of this paper are:

  • The related work section. While there is a section on generative modeling, this is a straightforward application of flow matching methods, and they should be covered in depth. Particularly relevant are Flow Matching for Generative Modeling (Lipman et al., 2023) and Flow straight and Fast: Learning to Generate and Transfer Data with Rectified Flow (Liu, Gong, & Liu 2022).
  • A lack of discussion of the implications of the Gaussianity assumption on errors. While the method clearly works, it would be helpful to give readers an understanding of how well this Gaussian characterization holds in practice (for instance, by fixing the Koopman dynamics and propagating errors forwards to create a rollout error distribution). This could also help give a better intuition for why different flow start points perform better at different propagation horizons.
Questions for authors
  • How well does the Gaussianity assumption hold in practice?
  • Do you an intuition for why Koopman training at moderate training horizons vs. shorter or longer horizons is better?
  • What was the dimensionality used for Koopman embeddings in your experiments? Did you try different dimensionalities, and how did that interact with the effect of training horizons & the flow stage?
  • What is the procedure for recursive correction? I couldn't find where it's explicitly defined in the paper.
4 comments

We sincerely appreciate your careful review and insightful comments, as well as your positive remarks on the clarity of our presentation, the well-grounded motivation and clever application, and the solid experimental evaluation. We address the concerns point by point below.

[W1] The related work section. While there is a section on generative d’samodeling, this is a straightforward application of flow matching methods, and they should be covered in depth. Particularly relevant are Flow Matching for Generative Modeling (Lipman et al., 2023) and Flow straight and Fast: Learning to Generate and Transfer Data with Rectified Flow (Liu, Gong, & Liu 2022).

Response: We thank the reviewer for this comment. We have expanded the discussion of Flow Matching in the Related Work section following your suggestion.

[W2/Q1] A lack of discussion of the implications of the Gaussianity assumption on errors. How well does the Gaussianity assumption hold in practice?

Response: We thank the reviewer for this constructive suggestion.

  1. We further validate the zero-mean Gaussian assumption through additional experiments. Specifically, we projected Koopman-induced residuals from held-out Navier–Stokes trajectories onto the first three principal directions at different prediction horizons.

    As shown in the table below, the residual means remain close to zero, the directional variances generally increase with the prediction horizon. Moreover, 18 of the 21 Shapiro–Wilk tests [1] do not reject the Gaussian hypothesis at the 5% significance level. These results support the reasonableness of modeling errors using Gaussian approximation with variance increasing with the prediction horizon.

  2. We have added Gaussian validation results in the revised manuscript. Visualizations of the residual distributions and their horizon-dependent statistics are provided in the Appendix H.1 (Figure 10). Furthermore, the Q-Q plots provided in Appendix H.1 (Figure 11) offer additional support for the Gaussian assumption.

[1] Shapiro, S. S., and Wilk, M. B. An analysis of variance test for normality (complete samples). Biometrika, 1965.

Direction indexDirection 1Direction 2Direction 3
Horizon151015202530151015202530151015202530
Mean9.6e-41.5e-3-1.6e-43.0e-39.9e-31.6e-21.4e-21.6e-6-2.4e-4-1.8e-3-2.0e-3-1.4e-33.9e-33.3e-3-4.9e-4-2.2e-3-1.0e-3-3.2e-3-7.8e-3-1.2e-2-8.4e-3
Std7.1e-32.8e-26.8e-21.1e-11.7e-11.9e-11.9e-18.0e-33.1e-27.2e-21.2e-11.8e-11.9e-11.9e-17.7e-33.6e-26.9e-21.3e-11.8e-12.0e-12.0e-1
Shapiro–Wilk p7.6e-15.6e-24.7e-13.2e-16.9e-16.9e-13.2e-16.9e-13.2e-17.5e-26.4e-45.6e-21.0e-16.9e-11.0e-11.8e-16.0e-56.3e-36.4e-25.8e-15.1e-1

[Q2] Do you an intuition for why Koopman training at moderate training horizons vs. shorter or longer horizons is better?

Response: We thank the reviewer for raising this question. From a spectral perspective, we interpret the effect of the training horizon as a trade-off between eigenbasis conditioning and predictive fidelity within a fixed-dimensional latent space.

  1. Short horizons provide limited constraints on multi-step evolution, potentially leading to highly correlated eigenvectors and sensitivity to approximation errors, as illustrated in Figure 6.
  2. Longer horizons may improve eigenbasis conditioning. However, since the finite-dimensional latent space is only approximately Koopman-invariant, enforcing linear consistency over excessively long horizons can compromise the accuracy of transient dynamics.
  3. Moderate horizons can balance these effects by providing sufficient temporal information to improve spectral conditioning while retaining the transient dynamics necessary for accurate prediction.

[Q3] What was the dimensionality used for Koopman embeddings in your experiments? Did you try different dimensionalities, and how did that interact with the effect of training horizons & the flow stage?

Response: We thank the reviewer for this helpful suggestion. We clarify the embedding dimension used in our original experiments and have added an ablation study to investigate its interaction with the training horizon and generative correction.

  1. All experiments reported in the original manuscript used a 128-dimensional Koopman embedding.

  2. We have evaluated embedding dimensions (m∈{32,64,128,256}m \in \{32,64,128,256\}) across Koopman training horizons (L∈{2,7,12}L \in \{2,7,12\}). The results show that an intermediate embedding dimension and training horizon (m=64m=64 and L=7L=7) yield the best overall performance after generative correction. This suggests that neither increasing the embedding dimension nor extending the training horizon alone necessarily improves the final KoopFlow performance. Moreover, generative correction consistently reduces prediction errors across all tested combinations of embedding dimensions and training horizons, demonstrating its effectiveness under different configurations. Due to time constraints, this ablation was conducted only on the Navier--Stokes dataset.

    We have added this experiment to the revised manuscript, included further analysis of the underlying reasons in Appendix H.1, and plan to extend the evaluation to other datasets in future work.

LLmmKoopman NMSEKoopFlow NMSEκ(V)\kappa(V)
2323.710e-014.534e-0228.2
2647.287e-011.355e-0159.6
21284.691e-015.523e-02236.6
22563.887e-015.939e-02296.4
7321.627e-013.586e-025.4
7646.817e-022.326e-0223.9
71281.406e-013.075e-02144.8
72569.225e-022.588e-02150.0
12328.859e-023.526e-0228.5
12645.851e-023.161e-0293.1
121288.296e-022.489e-0288.1
122562.680e-013.360e-02184.9

[Q4] What is the procedure for recursive correction? I couldn’t find where it’s explicitly defined in the paper.

Response: We thank the reviewer for this comment. This procedure divides a forecasting horizon into NN segments. At the end of each segment, generative correction is applied to the Koopman prediction, and the corrected latent state serves as the initial condition for the next Koopman rollout. We have added this explicit description and clarified the meaning of NN in the ablation study accompanying the revised manuscript.

We hope our responses have addressed your questions, and we welcome further feedback.

I thank the authors for their thoughtful responses, both to myself and the other reviewers. It's clear that they have considered the suggestions and questions provided carefully, and have performed a thorough set of experiments to bolster their results as well as proposed edits to improve the clarity of presentation. I am updating my score accordingly, to recommend a clear accept.

Dear Reviewer WTd1,

We are pleased that our responses have addressed your concerns. We sincerely thank you for raising your score and for your insightful comments and constructive suggestions!

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

This paper proposes KoopFlow, a two-stage, dual-pathway (prediction–correction) framework for long-horizon spatiotemporal PDE forecasting. In the first stage, a functional encoder-decoder pair is trained alongside a continuous-time linear Koopman generator (dzdt=Az\frac{dz}{dt} = Az) to advance latent states linearly via matrix exponentials. Recognizing that finite-dimensional linear Koopman evolution inevitably drifts off the true latent manifold over long rollouts, the second stage freezes the autoencoder and trains a latent Flow Matching model to capture the marginal distribution of ground-truth latent states. During inference, long-horizon forecasting alternates between inner-loop linear Koopman propagation and outer-loop periodic flow-based score/ODE rectification. Experiments on four 2D PDE benchmarks (Navier–Stokes, Reaction–Diffusion, 2DHeat, and 2DEM) evaluate both within-horizon and temporal-OOD performance.

Critical strengths

Intuitive Dual-Pathway Motivation: The high-level conceptual idea—viewing Koopman rollout error as off-manifold drift caused by linear approximation and pairing fast linear propagation with a second corrective pathway that projects states back toward the valid latent manifold—is intuitive and well-motivated for long-horizon rollout stabilization. Strong Empirical Stabilization in Temporal OOD: Across the four tested 2D benchmarks, periodically applying the second-stage correction noticeably suppresses the exponential error blow-up typically seen in pure linear Koopman rollouts beyond the training horizon.

Critical weaknesses

Severe Disconnect Between Theory, Mathematical Equations, and Actual Implementation: Theory vs. Training Objective Mismatch: In Section 3.3 and Theorem 1, the authors spend considerable space defining a horizon-dependent "Koopman-induced degradation distribution" qdeg(z~K∣zK)q_{\text{deg}}(\tilde{z}K | z_K) (modeled as a Gaussian mixture in Proposition 2) and proving the existence of a stochastic bridge between zKz_K and z~K∼qdeg\tilde{z}K \sim q{\text{deg}}. However, when actually training Stage II (Eq. 14), qdegq{\text{deg}} and the Koopman operator are completely absent! Stage II simply trains a standard unconditional Rectified Flow / Flow Matching model between isotropic Gaussian noise ξ∼N(0,Im)\xi \sim \mathcal{N}(0, I_m) and clean latent codes z∼pzz \sim p_z via zs=(1−s)ξ+szz_s = (1-s)\xi + sz. Thus, the theoretical bridge derived in Section 3.3–3.4 does not match what is actually optimized in Eq. (14).

Trivial Theoretical Proofs Unaligned with Claims: The formal propositions and theorems in the main text (proved in Appendix A.1–A.4) are mathematically trivial and do not substantiate the algorithmic contribution. Lemma 1 and Proposition 1 are just a first-order binomial expansion of (K+ΔK)n−Kn(K+\Delta K)^n - K^n and the triangle inequality; Proposition 2 simply assumes by fiat that deterministic Koopman rollout errors follow zero-mean Gaussians N(0,Σn)\mathcal{N}(0, \Sigma_n); and the proof of Theorem 1 (Appendix A.4) literally just writes down a linear interpolation Zs=(1−s)z0+sz1Z_s = (1-s)z_0 + sz_1 and its time-reversal Zˉs=Z1−s\bar{Z}s = Z{1-s} to claim the existence of a continuous reverse process.

Missing Inference Equations: In Section 3.5 (lines 355–366), after noting that deterministic ODE integration cannot correct z~K\tilde{z}K, the authors vaguely state that they use the score approximation gϕ(z,s)g\phi(z, s) (Eq. 16) to "apply a score-guided stochastic correction" before integrating back to s=1s=1. Yet the exact mathematical update rule, noise injection scale, step size, and number of ODE/SDE integration steps for this crucial correction step are completely omitted from both the main text and the appendix. Why Not Directly Add Nonlinearity to the Latent ODE Instead of a Complex Dual-Pathway Setup? Since the root cause of the drift is restricting the latent dynamics to a strictly linear ODE (dzdt=Az\frac{dz}{dt} = Az), a natural and much simpler question arises: why not directly incorporate a lightweight nonlinear residual term into the latent dynamics (e.g., a semi-linear Neural ODE dzdt=Az+fϕ(z)\frac{dz}{dt} = Az + f_\phi(z) or a deterministic MLP corrector in latent space) using the exact same parameter budget? The paper provides no ablation comparing its complex two-stage generative score-rectification pipeline against simply adding nonlinearity to the latent evolution operator.

Opaque Computational Overhead, Unspecified Backbone Architectures, and Questionable Fairness in Baseline Comparisons: Incomplete Complexity & Runtime Reporting: In Figure 7, the authors only compare inference FLOPs and runtime against two heavy generative baselines (LoLA with 46.66M parameters and CFO with 8.10M parameters), while omitting all deterministic and Koopman baselines (FNO, KNO, InfoKoop, MERLIN) from the efficiency comparison. Because KoopFlow requires a two-stage training pipeline (800–1000 epochs for Stage I + 500 epochs for Stage II, Table 5) and invokes iterative score-guided flow integration NN times during inference, its overhead relative to standard Koopman/operator methods is never transparently quantified. Missing Architecture Specifications & Training Budget Discrepancies: Remarkably, the paper never specifies the exact architectures and parameter counts of its own functional encoder EθE_\theta, decoder DθD_\theta, and velocity network vϕv_\phi, nor the number of NFE (function evaluations) used per flow correction. Meanwhile, Appendix D & E (Table 5 vs. lines 900–940) reveal that KoopFlow is trained for 1,300–1,500 total epochs (800–1,000 in Stage I + 500 in Stage II) with a 128-dimensional latent space, whereas baselines like FNO, DeepONet, KNO (width 32), MERLIN, and InfoKoop are trained for only 500 epochs. This makes it difficult to disentangle whether the gains come from the proposed dual-pathway mechanism or simply from a stronger autoencoder backbone, extra model complexity, and 2.5×–3×2.5\times\text{–}3\times more training epochs.

Questions for authors

Theory vs. Implementation: In Eq. (14), Stage II is trained using standard Gaussian noise ξ∼N(0,Im)\xi \sim \mathcal{N}(0, I_m) rather than samples from the Koopman degradation distribution qdeg(z~K∣zK)q_{\text{deg}}(\tilde{z}_K | z_K) derived in Definition 1 / Proposition 2. Why is there such a disconnect between Sections 3.3–3.4 and Eq. (14)? Furthermore, could you write out the exact mathematical equations and hyperparameters (step sizes, injected noise variance, number of integration steps) used for the "score-guided stochastic correction" following Eq. (16)?

Comparison with Direct Nonlinear Latent Dynamics: If the drift is caused by the linearity of dzdt=Az\frac{dz}{dt} = Az, what happens if the second-stage network vϕv_\phi is replaced by a direct nonlinear latent ODE term dzdt=Az+fϕ(z)\frac{dz}{dt} = Az + f_\phi(z) or a simple deterministic residual MLP corrector trained on Koopman rollout errors?

Fairness, Architectures, and Overhead vs. Koopman Baselines: Could you provide the exact architectural details of Eθ,Dθ,vϕE_\theta, D_\theta, v_\phi, report the training/inference runtime overhead compared to KNO/FNO/MERLIN, and clarify whether baselines match KoopFlow's performance when given the same latent dimension (128) and total training epochs (1,300–1,500)?

2 comments

We sincerely appreciate your careful review and insightful comments, as well as your positive remarks on the well-motivated framework and its strong empirical stabilization in temporal OOD. We address the concerns point by point below.

[W1-1/Q1-1] Theory vs. Implementation: In Eq. (14), Stage II is trained using standard Gaussian noise rather than samples from the Koopman degradation distribution derived in Definition 1 / Proposition 2. Why is there such a disconnect between Sections 3.3–3.4 and Eq. (14)?

Response: We thank the reviewer for raising this important question. Our clarifications are as follows:

  1. The standard Gaussian noise in Eq. (14) is used to construct the flow matching training path for learning an unconditional generative prior over clean latent states, rather than to model the Koopman-induced degradation distribution. Inspired by the use of generative priors in inverse problems [1], we leverage this learned prior to correct degraded Koopman predictions.

  2. We agree that directly learning a bridge from Koopman-degraded predictions to their corresponding reference latent states is a feasible alternative. As shown in the table below, both methods achieve comparable prediction performance across the evaluated benchmarks, while our KoopFlow exhibits higher Stage-II training efficiency. These results support our choice of a more efficient training strategy while maintaining comparable prediction performance.

[1] Diffusion posterior sampling for general noisy inverse problems. ICLR, 2023.

MethodNSRD2DHeat2DEMStage II Training time
Bridge2.395e-021.914e-034.973e-031.803e-0220.4 GPU-h
KoopFlow2.489e-022.121e-031.730e-031.756e-023.5 GPU-h

[W1-2/Q1-2] Furthermore, could you write out the exact mathematical equations and hyperparameters (step sizes, injected noise variance, number of integration steps) used for the ”score-guided stochastic correction” following Eq. (16)?

Response: We have added the exact equations, derivation, and hyperparameters in Section 3.5, Appendix B, and Appendix F (Table 7):

  1. Exact equations and derivation: The score-guided correction applied before each forward ODE step is explicitly given in Eq. (17) (Section 3.5), with its derivation and pseudocode provided in Appendix B.
  2. Hyperparameters: In all main experiments, we set sr=0.6s_r = 0.6 and γ=0.05\gamma = 0.05, yielding an injected Gaussian noise variance of γ(2−γ)(1−s)2\gamma(2-\gamma)(1-s)^2 at flow time ss. Both backward (1→sr1 \to s_r) and forward (sr→1s_r \to 1) integrations use 20 uniform Euler steps with step size Δs=0.02\Delta s = 0.02 (Appendix F, Table 7).

[W1-3/Q2] why not directly incorporate a lightweight nonlinear residual term into the latent dynamics (e.g., a semi-linear Neural ODE or a deterministic MLP corrector in latent space) using the exact same parameter budget?

Response: We thank the reviewer for this constructive suggestion. Following your suggestion, we evaluated two deterministic correction methods under the same parameter budget (~3.77M): (i) a lightweight MLP-based corrector and (ii) an iterative refinement neural operator (IRNO) following [1]. As shown in the table, KoopFlow consistently achieves the lowest NMSE across training and OOD evaluations, reducing the overall NMSE by 33.0% compared with the stronger deterministic baseline, IRNO.

[1] Iterative refinement neural operators are learned fixed-point solvers: A principled approach to spectral bias mitigation. ICML, 2026.

MethodTrainOODTotalParam
MLP1.428e-028.157e-024.292e-023.77M
IRNO5.914e-037.936e-023.717e-023.77M
KoopFlow2.026e-035.572e-022.489e-023.77M

[W1-4/Q3] Fairness, Architectures, and Overhead vs. Koopman Baselines: Could you provide the exact architectural details of Encoder,Decoder,Dit, report the training/inference runtime overhead compared to KNO/FNO/MERLIN, and clarify whether baselines match KoopFlow’s performance when given the same latent dimension (128) and total training epochs (1,300–1,500)?

Response: We thank the reviewer for this comment. We clarify the architectural details and provide an additional extended-training comparison below:

  1. Architectural details. We have provided architectural details of the Encoder, Decoder, and Dit in Appendix B (Table 6).

  2. Computational overhead. (i) We evaluated full-trajectory inference runtime on an RTX 4090 GPU. As shown in the following table, KoopFlow achieves the lowest prediction errors among all evaluated methods. Compared with the deterministic baselines FNO, KNO, and MERLIN, its inference is slower than FNO and MERLIN but faster than KNO, while its total training time is lower than that of MERLIN and KNO but higher than FNO. Compared with the generative baselines CFO and LoLA, KoopFlow achieves faster inference. (ii) Figure 8 compares KoopFlow with CFO and LoLA across different spatial resolutions, demonstrating its advantages in inference FLOPs, runtime, and parameter count.

CategoryDeterministic ModelGenerative Model
MethodFNOMERLINKNOKoopFlowCFOLoLA
Train NMSE8.096e-032.097e-023.254e-032.026e-035.313e-024.648e-02
OOD NMSE1.154e-011.626e-011.940e+005.572e-023.388e-011.300e-01
Inference time (ms)6.228.3100.670.6314.6845.7
Total training time (GPU-h)2.816.318.715.54.621.1
  1. Representation dimension. The latent dimension has different meanings across architectures. KoopFlow represents the system using a single 128-dimensional global latent vector, whereas a channel width of 128 in FNO or KNO corresponds to a spatial feature with 128 channels at each grid location.

  2. Extended training and model capacity To assess whether FNO and KNO benefit from increased model capacity and longer training, we increased their channel width from 32 to 128 and extended training from 500 to 1,300 epochs, comparable to KoopFlow's total training budget. As shown in the table below, these configurations incur higher parameter counts and computational costs while yielding limited and inconsistent accuracy gains. We therefore retain width 32 and 500 training epochs for FNO and KNO in the main experiments. MERLIN follows the same training epoch budgets as KoopFlow in both Phase I and Phase II. Across these evaluated configurations, KoopFlow maintains the lowest prediction errors, including against the wider and longer-trained FNO and KNO variants.

Model configurationComputational costPrediction error
MethodWidth/dEpochsParams (M)GFLOPsTrain NMSEOOD NMSE
Stage IStage II
FNO32500–2.1050.3268.096e-031.154e-01
FNO1281300–33.6593.7217.554e-031.120e-01
KNO32500–0.2635.4683.254e-031.940e+00
KNO1281300–8.40644.0803.528e-031.312e+00
Merlin1288005002.8631.0602.097e-021.626e-01
KoopFlow1288005003.7736.4732.026e-035.572e-02

We hope our responses have addressed your questions, and we welcome further feedback.

Reviewer_7UvTOpenReview ↗
Rating
2(−2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a generative correction of latent dynamic a Koopman operator of dynamical system. Apply that on commonly used benchmark problems and demonstrate competitive advantage in accuracy for both t<t_train and t>t_train (I intentionally avoided the word OOD ). They have also establish Lemmas regarding structure of error accumulation in the appendix which i haven't verified.

The proposed approach corrects the effect of unrolling of the latent space using a generative model, the make the connection to Mori-Zwanzig formalism.

Critical strengths

The experiments mostly support the claim. The Lemmas and propositions give interesting insight.

Critical weaknesses
  • The connection to Mori-Zwanzig formalism is either weak ∨ unnecessary ∨ poorly supported. I don't find any strong evidence that the error accumulation can be reduced by leveraging the history (the immediate consequence of Mori-Zwanzig).

  • Seeing the results on more difficult problems (something that this method fails), though I expect other methods to fail as well (simple example: 1D convection)

Questions for authors

Figure 1-b: Why \DetlaT\Detla T is plotted in reverse? Is that even surprising or worthy of mentioning? The point of it is not clear. Figure 1-c: I like to see the initial condition leading to GT at t=50, this helps the reader understand how much further in [simulation] time the prediction is extended. Figure 3: Does not give any sense in complexity of the problem or importance of reduction in error (maybe replace one of the snapshots from figures in appendix).

S4: I understand that the used benchmark is often used for PDE problem in ML community, but I strongly suggest evaluating the method on more challenging datasets, as I believe the claims on complexity of the mentioned problem is unsupported and exaggerated. I give the benefit of the doubts to the authors as they needed some baselines to establish the results.

Table 1: Building on the comment about, I am not sure if the solution outside the t<t_train can be regarded as OOD, for a true example of ODD, see Figure 6 and the discussion in https://ojs.aaai.org/index.php/AAAI/article/view/16116 (Low-Rank Registration Based Manifolds for Convection-Dominated PDEs). For example, the time horizon of the NS example is so short, that the no claims of t>t_train for accuracy is meaningful. I personally encourage showing some results, given the tight timeline, the discussion of this case in summary is suggested.

14 comments

We sincerely appreciate your careful review and insightful comments, as well as your positive remarks on the competitive advantage in accuracy and interesting insight. We address the concerns point by point below.

[W1] The connection to Mori-Zwanzig formalism is either weak “unnecessary” poorly supported. I don’t find any strong evidence that the error accumulation can be reduced by leveraging the history (the immediate consequence of Mori-Zwanzig).

Response: We clarify the relationship between our approach and the Mori–Zwanzig (MZ) formalism as follows:

  1. Theoretical connection. KoopFlow and the MZ formalism share a common motivation of addressing unresolved dynamics in reduced-order dynamical modeling. Specifically, the MZ formalism characterizes the effects of unresolved dynamics through memory and orthogonal-dynamics terms, whereas KoopFlow statistically characterizes accumulated Koopman approximation errors arising from both finite-dimensional reduction and operator learning.

  2. Methodological distinction. Unlike MZ-based approaches that explicitly model history-dependent memory effects, KoopFlow does not construct an MZ memory kernel or orthogonal-dynamics term. Instead, we employ generative correction to mitigate accumulated Koopman prediction errors, providing a complementary data-driven strategy rather than an explicit implementation of the MZ formalism.

  3. Supporting evidence. Previous studies [1-4] have demonstrated the benefits of incorporating historical information in reduced-order dynamical modeling. While these findings support the importance of accounting for unresolved dynamics, KoopFlow achieves its improvements through generative correction rather than explicit memory modeling. We have revised Appendix C to clarify this conceptual connection and methodological distinction.

[1] Learning reduced systems via deep neural networks with memory. JMLMC, 2020.

[2] Data-driven learning for the mori–zwanzig formalism: A generalization of the koopman learning framework. SIAM JADS, 2021.

[3] Mori–zwanzig latent space koopman closure for nonlinear autoencoder. PRS, 2025.

[4] On the benefits of memory for modeling time-dependent pdes. ICLR, 2025.

[W2/Q2] I understand that the used benchmark is often used for PDE problem in ML community, but I strongly suggest evaluating the method on more challenging datasets.

Response: We thank the reviewer for this constructive suggestion. Our responses are as follows:

  1. Additional challenging benchmark. To further evaluate KoopFlow on challenging real-world dynamics, we conducted additional experiments on the sea surface temperature (SST) dataset from the CMEMS Global Ocean Physics Reanalysis, with a native spatial resolution of 1/12° following [5,6]. We constructed 1,000 sequences from four non-overlapping oceanic regions, using the first 30 days for training and the subsequent 20 days for temporal extrapolation evaluation.
  2. Experimental results. As shown in the newly added results, KoopFlow achieves the lowest NMSE of 5.516e-03 among the compared methods. These results provide additional empirical support for its effectiveness in real-world spatiotemporal forecasting. Corresponding visual comparisons are provided in Appendix H.2 (Figure 17).
  3. Due to the limitations of the low-dimensional linear approximation in the Koopman latent representation, KoopFlow may perform less effectively on systems with stronger nonlinearities or abrupt structural changes. Given the time constraints, we only conducted a preliminary validation of this issue on a challenging real-world dataset. We have added a discussion of these limitations and the potential for further evaluation on more strongly nonlinear systems in the Future Work section.

[5] Continuous pde dynamics forecasting with implicit neural representations. ICLR, 2023.

[6] Deep learning for physical processes: Incorporating prior scientific knowledge. JSM, 2019.

MethodDeepONetFNOKNOLoLAInfoKoopCFOMERLINKoopFlow
NMSE6.124e-012.797e-013.662e-011.965e-022.883e+001.100e+009.614e-025.516e-03

[Q1-1] Figure 1-b: Why is plotted in reverse? Is that even surprising or worthy of mentioning? The point of it is not clear. Figure 1-c: I like to see the initial condition leading to GT at t=50, this helps the reader understand how much further in [simulation] time the prediction is extended.

Response: We thank the reviewer for the helpful comment and suggestion.

  1. In Figure 1(b), the descending order of ΔT\Delta T is intentional to illustrate how Koopman predictions starting from different historical states exhibit varying degrees of degradation when evaluated at the same target time, as characterized in Proposition 1.
  2. We have revise Figure 1(c) by explicitly illustrating the different starting points and their prediction trajectories toward the common target state at T=50T = 50.

[Q1-2] Figure 3: Does not give any sense in complexity of the problem or importance of reduction in error (maybe replace one of the snapshots from figures in appendix).

Response: We thank the reviewer for this valuable suggestion. We have revised Figure 3 by adding visual comparisons of prediction snapshots before and after generative correction, providing a clearer illustration of the complexity of dynamics and the effectiveness of error reduction. The revised figure is provided in the updated manuscript.

[Q3] Table 1: Building on the comment about, I am not sure if the solution outside the t<ttraint < t_{\mathrm{train}} can be regarded as OOD. I personally encourage showing some results, given the tight timeline, the discussion of this case in summary is suggested.

Response: We thank the reviewer for this comment. Our clarifications are as follows:

  1. The NS, RD, and 2DHeat datasets are obtained from established PDE benchmarks [1,2], and we follow the corresponding evaluation protocols widely adopted by the research community to assess temporal extrapolation performance.
  2. To further assess long-horizon extrapolation, we generated extended 2DEM trajectories using MEEP simulations [3]. As shown in the new results, KoopFlow achieves the lowest prediction errors across all four evaluated horizons (T=75,100,125,150T = 75,100,125,150), with increasing improvements over the strongest competing baseline (LoLA), further demonstrating its long-horizon temporal extrapolation ability.
  3. Moreover, the generalization experiments in the Appendix H.1 (Table 8) present results on the heat equation, where the trained model is directly evaluated on unseen parameter values. These results provide further evidence of KoopFlow's generalization capability under the evaluated parameter settings.
  4. We have also added a brief discussion in the Conclusion to clarify this issue and the scope of our extrapolation results.

[1] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. ICLR, 2021.

[2] Felix Koehler, Simon Niedermayr, Rudiger Westermann, and Nils Thuerey. APEBench: A benchmark for autoregressive neural emulators of pdes. NIPS, 2024.

[3] Ardavan F Oskooi, David Roundy, Mihai Ibanescu, Peter Bermel, John D Joannopoulos, and Steven G Johnson. MEEP: A flexible free-software package for electromagnetic simulations by the FDTD method. CPC, 2010.

TTDeepONetFNOKNOLoLAInfoKoopCFOMERLINKoopFlow
752.125e+008.757e-015.452e+013.717e-011.165e+001.010e+001.139e+009.99e-02
1002.609e+001.500e+006.852e+017.654e-011.243e+006.604e+021.659e+001.875e-01
1253.346e+003.399e+009.535e+011.097e+001.368e+00–2.411e+002.447e-01
1505.202e+001.523e+011.68e+021.447e+001.681e+00–4.129e+002.907e-01

We hope our responses have addressed your questions, and we welcome further feedback.

  1. The added MZ supports my earlier worry, as it provides reasons for "orthogonality of the proposed approach to MZ", so all the discussions surrounding the MZ are of no value. It is obvious that discretization and reduction of resolution causes deviation of emulator from the underlying dynamics, there is no connection of this work to MZ and it seems the justification is just used as stuffing more words in the paper.

  2. The additional "challenging" problems are of the same kind of the previous examples and provide no additional insight.

I have lowered my score to reflect the circular logic of the paper.

Reviewer_ipbAOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KoopFlow, a prediction-correction framework for long-horizon forecasting. The method treats errors accumulated during finite-dimensional Koopman rollouts as a progressive degradation process and uses a flow-matching generative model to correct the resulting latent predictions. Experiments on four PDE benchmarks show promising forecasting performance, particularly at prediction horizons beyond those used in training.

Critical strengths

The main strength of this paper is its empirical performance on long-horizon prediction. KoopFlow achieves competitive results across four PDE benchmarks, including forecasts beyond the training horizon. The ablation studies further show that generative correction substantially improves the accuracy of the underlying Koopman model, supporting the practical value of the proposed prediction-correction framework.

Critical weaknesses
  1. Theoretical justification and assumptions. The claimed guarantee for generative recovery appears stronger than the theoretical analysis supports. Theorem 1 constructs a continuous path between a degraded prediction and its corresponding reference latent state, thereby establishing the existence of a reverse process. However, it does not establish that the velocity field learned through Eq. (14) can recover that reference state from the degraded Koopman prediction. The analysis therefore motivates generative correction, but the connection between the constructed reverse process and the trained model remains insufficiently justified. Moreover, Proposition 2 assumes that accumulated Koopman rollout errors are Gaussian, have zero mean, and satisfy Σ_(n+1)⪰Σ_n. Under these assumptions, Eq. (10) follows directly from the law of total probability, while the assumptions themselves are not sufficiently examined. In particular, Eq. (9) defines a discrete empirical distribution from predictions originating at different historical states, whereas Eq. (10) introduces a continuous Gaussian error distribution at each prediction horizon. The source of this additional randomness is unclear. The zero-mean assumption also requires further justification. Equation (10) centers every Gaussian component at z_K, implying E[\tilde{Z}_K-z_K∣z_K]=0. However, the discussion in lines 261–265 allows for a systematic displacement of the degradation distribution from the reference state. From the Mori–Zwanzig perspective, unresolved dynamics can also produce memory effects and history-dependent prediction biases, even when the unresolved forcing is unbiased. These considerations make the zero-mean Gaussian assumption difficult to justify in general.
  2. Clarity and reproducibility. The paper lacks a dedicated reproducibility statement. Although Appendix D and Table 5 report training configurations, the generative correction procedure in Section 3.5 remains insufficiently specified. In particular, Eqs. (15)–(17) do not clearly explain how a Koopman prediction is initialized at the intermediate flow time s_r or how the correction is then carried out. Pseudocode covering both training and inference would help make the procedure reproducible. Table 4 studies the effects of s_r and the Koopman propagation window, but the values of s_r used in the main experiments are not clearly reported. It is also unclear whether s_r varies with the prediction horizon. In addition, L denotes both the training horizon in Eq. (13) and the propagation window in Table 4. Distinguishing these quantities, together with a brief introduction and foundational references for flow matching, would improve the clarity of the methodology.
Questions for authors
  1. Could the authors visualize the Koopman-induced degradation distribution on held-out trajectories for at least one benchmark, such as Navier–Stokes? Residual distributions and their means along representative latent directions, together with directional variances at different prediction horizons, would help assess the Gaussian, zero-mean, and covariance-growth assumptions in Proposition 2.

  2. Could the authors provide training and inference pseudocode, with explicit details of the generative correction step? In particular, please clarify the initialization at the intermediate flow time s_r, the correction procedure, the values of s_r used in the main experiments and whether they vary with the prediction horizon, and the selection of correction checkpoints during recursive prediction.

4 comments

We sincerely appreciate your careful review and insightful comments, as well as your positive remarks on the forecasting performance and practical value of the framework. We address the concerns point by point below.

[W1-1] The claimed guarantee for generative recovery appears stronger than the theoretical analysis supports. The analysis therefore motivates generative correction, but the connection between the constructed reverse process and the trained model remains insufficiently justified.

Response: We thank the reviewer for this comment. Our detailed explanations are as follows:

  1. We agree that Theorem 1 establishes the existence of a continuous recovery path but does not guarantee exact recovery by the learned velocity field. Since the true zKz_K is unavailable during inference, the constructed path cannot be directly implemented.
  2. To address this practical limitation, inspired by the use of learned generative priors to guide the recovery of unknown states in inverse problems [1], we train an unconditional flow matching model on clean reference latent states and then leverage this learned prior to correct degraded Koopman predictions. This provides a practical, data-driven correction mechanism rather than a direct theoretical consequence of Theorem 1.
  3. Experiments demonstrate improved prediction accuracy and stability, providing empirical support for the effectiveness of our generative correction mechanism. In response to your comment, we have revised the relevant statements to clarify that Theorem 1 primarily provides theoretical motivation for generative correction rather than a rigorous guarantee of exact recovery.

[1] Diffusion posterior sampling for general noisy inverse problems. ICLR, 2023.

[W1-2] In particular, Eq. (9) defines a discrete empirical distribution from predictions originating at different historical states, whereas Eq. (10) introduces a continuous Gaussian error distribution at each prediction horizon. The source of this additional randomness is unclear.

Response: Our clarifications are as follows:

  1. Eq. (9) defines a discrete empirical distribution of Koopman predictions originating from different historical states along the same trajectory. For a fixed trajectory and learned Koopman operator, the prediction errors remain deterministic, arising from operator learning errors and finite-dimensional approximation errors.
  2. To statistically characterize the variability of these accumulated errors across trajectories governed by the same dynamics, we adopt a zero-mean Gaussian approximation with horizon-dependent covariance. This provides a tractable statistical characterization supported by the residual analysis in Table 4 and Figure 7 in the manuscript, rather than implying additional randomness in the underlying dynamics.
  3. Accordingly, Eq. (10) provides a continuous statistical approximation of the degradation distribution described by Eq. (9), rather than an exact reformulation.

[W1-3/Q1] Could the authors visualize the Koopman-induced degradation distribution on held-out trajectories for at least one benchmark, such as Navier–Stokes? Residual distributions and their means along representative latent directions, together with directional variances at different prediction horizons, would help assess the Gaussian, zero-mean, and covariance-growth assumptions in Proposition 2.

Response: We thank the reviewer for this helpful comment. Our clarifications are as follows:

  1. We have added visualizations of the Koopman-induced degradation distribution on held-out Navier–Stokes trajectories in Appendix H.1 (Figure 12).
  2. To assess the assumptions in Proposition 2, we analyze Koopman rollout errors projected onto the first three principal directions across different prediction horizons. As reported in the following table, the error means remain close to zero, while the directional variances exhibit an overall increasing trend. Moreover, 18 out of 21 Shapiro–Wilk tests [1] do not reject normality at the 5% significance level, providing empirical support for the approximate Gaussian characterization.
  3. We have included these statistics, error distribution visualizations, and Q–Q plots in Appendix H.1 (Table 10 and Figures 10–12), providing further empirical support for the adopted statistical assumptions.

[1] Shapiro, S. S., and Wilk, M. B. (1965). An analysis of variance test for normality (complete samples). Biometrika.

Direction indexDirection 1Direction 2Direction 3
Horizon151015202530151015202530151015202530
Mean9.6e-41.5e-3-1.6e-43.0e-39.9e-31.6e-21.4e-21.6e-6-2.4e-4-1.8e-3-2.0e-3-1.4e-33.9e-33.3e-3-4.9e-4-2.2e-3-1.0e-3-3.2e-3-7.8e-3-1.2e-2-8.4e-3
Std7.1e-32.8e-26.8e-21.1e-11.7e-11.9e-11.9e-18.0e-33.1e-27.2e-21.2e-11.8e-11.9e-11.9e-17.7e-33.6e-26.9e-21.3e-11.8e-12.0e-12.0e-1
Shapiro–Wilk p7.6e-15.6e-24.7e-13.2e-16.9e-16.9e-13.2e-16.9e-13.2e-17.5e-26.4e-45.6e-21.0e-16.9e-11.0e-11.8e-16.0e-56.3e-36.4e-25.8e-15.1e-1

[W2-1/Q2] Could the authors provide training and inference pseudocode, with explicit details of the generative correction step? In particular, please clarify the initialization at the intermediate flow time srs_r, the correction procedure, the values of srs_r used in the main experiments and whether they vary with the prediction horizon, and the selection of correction checkpoints during recursive prediction.

Response: We have expanded Section 3.5 and Appendix B, with detailed configurations provided in Appendix F (Table 7). Our clarifications are as follows:

  1. Algorithms 1 and 2 in Appendix B provide the complete two-stage training and recursive inference procedures, respectively.
  2. At each checkpoint, we first initialize zsrz_{s_r} by mapping the Koopman forecast z~K\widetilde z_K from s=1s=1 to sr=0.6s_r = 0.6. Starting from zsrz_{s_r}, we then refine the state through forward integration back to s=1s=1, applying the score-guided correction (Eq. (17)) before each step (Section 3.5 and Appendix B). The corresponding derivation is provided in Appendix B (Eqs. (18)–(19)).
  3. We partition the forecasting horizon into (N=7N = 7) equally spaced segments, applying correction at each checkpoint and using the corrected state to initialize the next segment, without accessing future ground-truth states. In the main experiments, we keep sr=0.6s_r = 0.6 across all checkpoints and prediction horizons. The sensitivity of srs_r to different uncorrected propagation horizons is analyzed in Table 5.

[W2-2] In addition, LL denotes both the training horizon in Eq. (13) and the propagation window in Table 4.

Response: We thank the reviewer for pointing out this ambiguity. In Eq. (13), LL denotes the training horizon, whereas in Table 4, it denotes the temporal length of the Koopman propagation window. To avoid confusion, we have replaced LL with WW in Table 4.

[W2-3] Distinguishing these quantities, together with a brief introduction and foundational references for flow matching, would improve the clarity of the methodology.

Response: We thank the reviewer for this valuable suggestion. Following your suggestion, we have added a brief introduction to Flow Matching and its foundational references in the Related Work section.

We hope our responses have addressed your questions, and we welcome further feedback.

The authors have addressed my major concerns raised in the first round of review. I appreciate the efforts made by the authors, and therefore update the scores and the acceptance suggestions as well.

Dear Reviewer ipbA,

We are glad that our responses have addressed your concerns. We sincerely thank you for increasing your score, as well as your insightful feedback and constructive suggestions! Please feel free to reach out with any further questions!

Comments on the paper

1 comment

We sincerely thank the reviewers for their thoughtful comments and constructive suggestions, which have helped us substantially improve our work, and the Area Chair for their time and consideration. In response to the reviewers' feedback, we have made the following major revisions:

  • Gaussian error assumption (Reviewers ipbA and WTd1): We added statistical analyses of Koopman rollout errors on held-out Navier–Stokes trajectories, including error distributions, directional means and variances, and normality tests, providing further empirical support of our theoretical assumptions.
  • Method clarity and reproducibility (Reviewers ipbA, GGaA, and WTd1): We added training and inference pseudocode, detailed architectural specifications, explicit correction equations and inference configurations, and clarified the recursive correction procedure. We also clarified the relationship between the generative prior and the theoretical degradation analysis and expanded the discussion of Flow Matching with foundational references.
  • Ablation and fair comparisons (Reviewers GGaA and WTd1): We added latent-dimension ablations, parameter- and training-budget-aligned comparisons, including a parameter-matched MLP correction baseline, and more comprehensive computational cost evaluations against representative Koopman and neural operator methods.
  • Mori–Zwanzig connection (Reviewer 7UvT): We clarified the role of MZ as theoretical motivation and a contrasting perspective, and added extensive comparisons with MERLIN, a representative MZ-inspired method, across multiple datasets.
  • Challenging benchmarks and extrapolation (Reviewer 7UvT): We added evaluations on more challenging datasets, including real-world SST and 1D convection, together with additional Burgers' equation temporal extrapolation experiments. We also clarified the meaning of OOD in this work and provided further evidence of the temporal extrapolation capability of KoopFlow.

The revised PDF highlights manuscript changes in blue for ease of review. We sincerely thank the reviewers again for their valuable feedback, which has helped us strengthen the quality and completeness of our work, and thank the Area Chair for coordinating the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.