acceptodds
Under review as a conference paper at ICLR 2027

Invertible continuous latent dynamic for long-term data assimilation in complex physical systems

Abstract

Forward forecasting and data assimilation are the two important aspects in physical simulation: one propagates the state forward, the other recovers unknown states from sparse observations. Learned surrogates are normally built and benchmarked for forward forecasting, however, whether a surrogate could attain good performance in data assimilation tasks is valuable as well, as inverse problems are of paramount importance in the scientific domain. In this paper, we propose a continuous-time Koopman autoencoder whose latent dynamics obey , yielding closed-form inference via at any horizon in a single step. This decouples forecast cost from forecast length at inference time, showing long-term stability and high efficiency in forward simulation, and also supports data assimilation as gradient-based optimization with cost independent of the assimilation window. Experiments are performed on the Kuramoto–Sivashinsky equation and a transient flow, and we compare our method against a range of baselines on the forward problem, including diffusion models and operator-learning models, and obtain a 110 inference speedup over strong diffusion baselines. We further test these baselines on an initial-state inference data assimilation task, and find that a strong forecaster does not guarantee a strong assimilator, while the continuous-time Koopman autoencoder achieves both higher accuracy and efficiency than surrogates of comparable forward performance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Autoencoder, Latent Dynamics, Physical Systems

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

This paper builds on Koopman theory to propose an efficient method for forecasting and estimating initial conditions from future observations. The key idea is to map the data to a latent space in which the dynamic is continuous and linear. This allows the latent state zτz_{\tau} to be computed explicitly at any time by exponentiating KK, the matrix that governs the latent dynamics. The dynamics can therefore be integrated efficiently without sequential rollout. The Jacobian can also be computed efficiently (without a lot of matrix-matrix products), enabling the implementation of a 4D-Var-like data assimilation methods.

Critical strengths
  • The paper is easy to follow, and the ideas are clearly presented, even for readers who are not familiar with the literature.
  • The proposed method is highly efficient while maintaining satisfactory forecasting performance compared with other baselines. In particular, the linear dynamics in latent space allow zτz_\tau to be computed at any time τ\tau without autoregressive rollout, but simply by evaluating the exponential of KK. The same principle applies to the Jacobian, helping to mitigate gradient issues and numerical errors.
  • The proposed method appears to remain stable over long integration horizons.
  • Since the dynamic is continuous in latent space, the solution can be evaluated at any time τ\tau, so not only at the timesteps seen during training. Figure 5c suggests that the method also produces accurate forecasts at previously unseen timesteps.
  • The low-rank matrix provides an elegant way to incorporate physical parameters (e.g., the Reynolds number) into the dynamics. It also acts as a regularizer and allows the model to handle multiple physical regimes. This is an interesting practical capability that deserves greater emphasis in the main text.
Critical weaknesses
  • Both forecasting and data assimilation are entirely deterministic. This is not necessarily a limitation for systems governed by known deterministic physical equations, such as those considered in the paper. However, the method may be less effective for complex systems such as the atmosphere (mentioned in the paper), for which deterministic predictions tend to oversmooth the range of possible outcomes. Similarly, the proposed data assimilation method appears to be an efficient implementation of 4D-Var in the latent space. It therefore provide only a point estimate of the solution rather than an ensemble representing the distribution of states compatible with the observations. In practice, observations are often sparse, and capturing the associated uncertainty and range of plausible solutions is important.
  • My understanding is that the observations are sparse only in time, but not in space. This seems somewhat unrealistic, as real-world observations are often both spatially sparse and noisy. I would therefore appreciate clarification on whether I have correctly understood the experimental setup. If so, could the authors discuss whether the proposed method can handle spatially sparse and noisy observations, rather than assuming access to the full system state at selected time points.
  • Is the matrix K0K_0 in Equation (11) directly learned during training? If so, I am concerned about the scalability of the method to high-resolution systems. As I understand it, Koopman methods require a latent space with a higher dimensionality than the original state space in order to obtain linear dynamics. For high-dimensional systems (like the atmosphere for example), this could lead to a very large latent space and, consequently, a prohibitively large K0K_0 matrix (whose memory cost scales quadratically with the latent dimension). I therefore wonder whether the authors also use a low-rank parameterization of this matrix in practice, and whether they have tested the method on larger systems. More generally, could the authors report in the main text the spatial resolutions of the systems considered in the experiments and the corresponding latent dimensions ?
Questions for authors

Could the authors clarify the points raised in the Weaknesses section (or explain why these concerns are incorrect) ? I'd be happy to increase my score if the authors provide convincing answers to my concerns.

Reviewer_i9dCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a continuous-time Koopman autoencoder (KAE) for long-horizon forecasting and variational data assimilation. The method uses a linear latent generator conditioned on physical parameters, allowing direct temporal propagation through matrix exponentiation without autoregressive rollouts. Experiments on the Kuramoto–Sivashinsky (KS) equation and fluid-flow benchmarks report competitive forecasting accuracy, computational efficiency, and initial-state reconstruction.

Critical strengths
  1. The paper provides a useful empirical comparison of forecasting and data assimilation, showing that models with better forward prediction accuracy do not necessarily achieve better state reconstruction. This highlights the importance of evaluating both tasks separately.

  2. The continuous-time Koopman formulation offers practical computational advantages by enabling direct temporal propagation and sensitivity computation without repeated autoregressive steps, particularly over long assimilation windows.

  3. The experimental evaluation covers multiple dynamical systems and physical regimes, including irregular sampling and unseen parameters. The analysis also considers observation noise, sensor sparsity, and assimilation horizons, providing useful insights into both the strengths and limitations of the proposed approach.

Critical weaknesses

1. Assimilation comparisons and inverse stability.

The assimilation experiments do not fully isolate the contribution of continuous-time Koopman dynamics. KAE optimizes a low-dimensional latent state, while U-Net and FNO optimize physical-state variables without explicit background regularization. Differences in optimization dimensionality and implicit regularization may therefore influence the results. Comparisons with regularized 4D-Var, nonlinear latent dynamics, and a discrete-time KAE would strengthen the conclusions.

Additionally, invertibility of the latent operator does not guarantee stable inverse reconstruction. The authors should analyze error amplification during backward propagation and the reconstruction limitations of the learned latent representation.

2. Insufficient discussion of related work and methodological novelty.

The proposed formulation is not sufficiently distinguished from existing work on Koopman operators and data assimilation. Several relevant studies are missing, including Data Assimilation in Operator Algebras [1], Learning Semilinear Neural Operators [2], Tensor-Var [3], LD-EnSF [4], KODA [5], and LEVDA [6]. Related forecasting methods such as KNF [7], Koopa [8], and the Markov Neural Operator [9] are also not discussed.

More importantly, continuous-time Koopman representations have already been investigated by Pan and Duraisamy [10] and Frion et al. [12]. Frion et al. [13] also investigate Koopman-based variational data assimilation with continuous temporal reconstruction. Recent work on invertible Koopman autoencoders [11] provides further context. Although the submission cites [12–15], it does not adequately explain how its formulation advances beyond these approaches.

The authors should clarify what constitutes the primary methodological novelty beyond established continuous-time Koopman modeling and latent-space assimilation.

3. Robustness under noisy and sparse observations.

Appendix J evaluates Gaussian observation noise, but Table 6 shows substantial degradation in KAE's assimilation performance on KS. SDA outperforms KAE at every nonzero noise level evaluated. Similarly, Table 7 shows worse performance for KAE at sensor coverage of 25% or less. The authors should explain these results and investigate robustness across different SNR settings or different noise distributions.

4. Inconsistencies in experimental results and ablations.

Section 5 claims that KAE outperforms ACDM in both incompressible extrapolation regimes, whereas Table 1 reports worse performance for KAE in the high-Reynolds-number regime 2.9 vs. 0.8. The KAE and Exponential KAE rows in Table 2 also report identical results and require clarification.

5. Computational efficiency and long-horizon fidelity.

The reported 110× speedup only compares per-step inference against a diffusion baseline and does not account for the iterative optimization required for assimilation. The authors should report end-to-end assimilation time at comparable reconstruction accuracy, including comparisons with deterministic baselines.

The long-horizon predictions also lose fine-scale details and converge toward smoother dynamics. The authors should provide stronger statistical and spectral evaluations to demonstrate that the model preserves the underlying physical dynamics rather than simply producing stable but oversmoothed predictions.

Data Assimilation and Neural Operators

[1] Freeman, D., Giannakis, D., Mintz, B., Ourmazd, A., & Slawinska, J. Data Assimilation in Operator Algebras. PNAS, 120(8), 2023. https://doi.org/10.1073/pnas.2211115120

[2] Singh, A., Borsoi, R., Erdogmus, D., & Imbiriba, T. Learning Semilinear Neural Operators: A Unified Recursive Framework for Prediction and Data Assimilation. ICLR, 2024. https://openreview.net/forum?id=ZMv6zKYYUs

[3] Yang, Y., et al. Tensor-Var: Efficient Four-Dimensional Variational Data Assimilation. ICML, 2025. https://proceedings.mlr.press/v267/yang25h.html

[4] Xiao, P., Si, P., & Chen, P. LD-EnSF: Synergizing Latent Dynamics with Ensemble Score Filters for Fast Data Assimilation with Sparse Observations. ICLR, 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/f7883c88d25e8cd0f05f06c92375febc-Abstract-Conference.html

[5] Singh, A., Singh, A., Imbiriba, T., Erdogmus, D., & Borsoi, R. KODA: A Data-Driven Recursive Model for Time Series Forecasting and Data Assimilation using Koopman Operators. arXiv:2409.19518, 2024. https://arxiv.org/abs/2409.19518

[6] Si, P., & Chen, P. LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics. arXiv:2602.19406, 2026. https://arxiv.org/abs/2602.19406

Koopman Forecasting and Neural Operators

[7] Wang, R., Dong, Y., Arik, S. Ö., & Yu, R. Koopman Neural Forecaster for Time Series with Temporal Distribution Shifts. ICLR, 2023. https://arxiv.org/abs/2210.03675

[8] Liu, Y., Li, C., Wang, J., & Long, M. Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors. NeurIPS, 2023. https://arxiv.org/abs/2305.18803

[9] Li, Z., et al. Learning Chaotic Dynamics in Dissipative Systems. NeurIPS, 2022. https://papers.neurips.cc/paper_files/paper/2022/hash/6ad68277e27b42c60ac228c9859fc1a2-Abstract-Conference.html

Continuous-Time and Invertible Koopman Models

[10] Pan, S., & Duraisamy, K. Physics-Informed Probabilistic Learning of Linear Embeddings of Nonlinear Dynamics with Guaranteed Stability. SIAM Journal on Applied Dynamical Systems, 19(1):480–509, 2020. https://doi.org/10.1137/19M1267246

[11] Frion, A., Drumetz, L., Dalla Mura, M., Tochon, G., & Aïssa El Bey, A. Augmented Invertible Koopman Autoencoder for Long-Term Time Series Forecasting. TMLR, 2025. https://openreview.net/forum?id=o6ukhJLzMQ

[12] Frion, A., et al. Leveraging Neural Koopman Operators to Learn Continuous Representations of Dynamical Systems from Scarce Data. ICASSP, 2023. https://arxiv.org/abs/2303.06972

[13] Frion, A., et al. Neural Koopman Prior for Data Assimilation. IEEE Transactions on Signal Processing, 72:4191–4206, 2024. https://doi.org/10.1109/TSP.2024.3416828

[14] Lusch, B., Kutz, J. N., & Brunton, S. L. Deep Learning for Universal Linear Embeddings of Nonlinear Dynamics. Nature Communications, 9, 2018. https://doi.org/10.1038/s41467-018-07210-0

[15] Azencot, O., Erichson, N. B., Lin, V., & Mahoney, M. W. Forecasting Sequential Data Using Consistent Koopman Autoencoders. ICML, 2020. https://proceedings.mlr.press/v119/azencot20a.html

Questions for authors
  1. How much of the reported assimilation improvement comes from continuous-time Koopman dynamics rather than optimizing in a lower-dimensional latent space? Since U-Net and FNO optimize directly over physical states without explicit regularization, can the authors provide comparisons with regularized 4D-Var, nonlinear latent dynamics, and a discrete-time KAE using the same encoder–decoder architecture? These comparisons would help isolate the contribution of the proposed continuous-time formulation.

  2. How does the stability of backward propagation change with increasing assimilation horizons, particularly when the learned Koopman generator contains dissipative modes? Since backward evolution can amplify errors in rapidly decaying modes, can the authors analyze the singular values of the inverse propagator and quantify how reconstruction errors grow with the observation gap? It would also be useful to report the reconstruction accuracy of the encoder–decoder independently of the assimilation procedure.

  3. Methodological novelty: Continuous-time Koopman representations, matrix-exponential propagation, and Koopman-based data assimilation have been investigated in earlier work, including Pan and Duraisamy (2020) and Frion et al. (2023, 2024). What specifically distinguishes the proposed formulation from these approaches? Is the main contribution the physics-conditioned operator, the efficiency of sensitivity computation, or the forecasting–assimilation evaluation? The authors should clarify which components are methodologically new and provide comparisons that support those contributions.

  4. Robustness to noise and sparse observations: Why does KAE exhibit substantially higher reconstruction errors than SDA under Gaussian observation noise on KS (Table 6), despite performing well in noiseless settings? Similarly, what explains its degradation under reduced sensor coverage (Table 7)? Can the authors investigate whether these failures arise from backward error amplification, limitations of the learned latent representation, or the absence of an explicit prior during assimilation? Additional evaluations across normalized SNR settings and different noise distributions would help establish the method's robustness.

Minor : Please cross check all the inconsistency between the table and the text.

4 comments

Thank you for a thorough and demanding review. We ran every comparison you requested and revised the claims where the evidence required it.

W1 + Q1 — Isolating continuous-time dynamics from latent optimisation. All three requested comparisons the full table, varying dynamics, control and first guess one at a time on identical trajectories (δf=9\delta_f=9), are in our reply to WFaJ (W3).

  • Regularised 4D-Var: a U-Net with a physical-space Gaussian background reaches 0.060 on KS, against 0.014 for the KAE. On transonic flow it leaves the U-Net unchanged (0.298→0.296).
  • Optimisation dimensionality / control variable: on KS the control matters, but not through dimensionality (the KS latent, 512, is larger than the state, 64). From the same first guess, KAE's closed-form pull-back D(e−Kτ1E(y1))\mathcal D(e^{-K\tau_1}\mathcal E(y_1)) (the first observation carried back to t0), the U-Net improves from 0.027 to 0.009 with the latent control, while KAE dynamics with a physical control stall at 0.067. The pull-back alone improves the U-Net 3× (0.085→0.027). On transonic flow the control does not matter (physical through the KAE: 0.061–0.070; latent: 0.068). There the pull-back alone is already the best estimate when the nearest observation is clean and complete (0.050, vs 0.25–0.30 for every baseline), and 4D-Var is what makes KAE robust when it is not (10% coverage: 0.252→0.088; =0.3: 0.142→0.054).
  • Nonlinear latent dynamics and a discrete-time KAE, trained in the identical settings (same encoder–decoder, NzN_z, data, seed): 0.256 and 0.069, against 0.014 for the continuous-time linear KAE (δf=9\delta_f=9), even though the nonlinear model forecasts 2.5× better. Their pull-backs are also worse (0.268 and 0.231 vs 0.072).

So both the starting estimate and the variables we optimise affect accuracy, and not because fewer variables are optimised. The U-Net driven through the KAE decoder from the pull-back beats the KAE (0.009 vs 0.014) but takes 18× longer (1165 s vs 64 s). The learned representation and starting estimate help determine accuracy, while linear latent dynamics make the optimisation cheap. Under the tested settings the continuous linear model assimilates better than the discrete linear and nonlinear alternatives. Continuous time allows arbitrary observation times, although discrete linear KAEs also propagate efficiently. We now present the KAE as a favourable accuracy–cost balance, and the contribution statement says so.

W1 + Q2 — Stability of backward propagation, singular values, error vs gap and reconstruction alone.

  • Where the inverse is used: latent 4D-Var never inverts the dynamics inside the optimisation. It fits z0z_0 through the forward maps eKτie^{K\tau_i}. The inverse e−Kτ1e^{-K\tau_1} is applied once, for the warm start (the pull-back above). For the KS model, ∥e−Kτ∥2\Vert e^{-K\tau}\Vert_2 is 2.06 at τ=0.9\tau=0.9, 5.2 at τ=2.5\tau=2.5 and 31 at τ=10\tau=10 (Lyapunov time 22.6).

  • Your concern about rapidly decaying modes is real for the warm start. The only place the inverse is applied is the warm start z0=e−Kτ1E(y1)z_0=e^{-K\tau_1}\mathcal E(y_1): the first observation is encoded and propagated backwards to t0t_0. Backward propagation turns damping into growth, so any error in E(y1)\mathcal E(y_1) along a strongly damped mode is amplified by up to ∥e−Kτ1∥2\Vert e^{-K\tau_1}\Vert_2 (Frion et al. 2024 report the same for backward prediction with discrete Koopman operators). To quantify this we compared four transonic generators: the paper model and three models retrained with the same architecture, data and training schedule. All four have equally benign forward maps (∥eKτ∥2≤6.3\Vert e^{K\tau}\Vert_2\le6.3 for τ≤2.5\tau\le2.5) and comparable forecasts. However, they differ in how strongly their fastest mode is damped (most negative real eigenvalue of KK), and therefore in ∥e−Kτ1∥2\Vert e^{-K\tau_1}\Vert_2 at the 9-frame gap (τ1=0.9\tau_1=0.9). We ran the full transonic sweep of the new Table 15 with each model (δf=9\delta_f=9, 26 settings of noise, coverage, recovery horizon and number of observations, 8 trajectories each, 500 iterations, objective with RR and BB), changing only the warm start. Mean analysis error over the 26 settings:

    modelmin Re λ(K)\lambda(K)∥e−Kτ1∥2\Vert e^{-K\tau_1}\Vert_2error, pulled-back starterror, regularised start
    paper model−1.550.0680.068
    retrain A−5.31300.1270.077
    retrain B−5.21160.1810.080
    retrain C−15.410610^61600.082

    The assimilation error follows the backward amplification: with an amplification of 10610^6 the starting point is so far off that the optimisation diverges. The fix is a regularised pull-back, z0=arg⁡min⁡z∥eKτ1z−E(y1)∥2+λ∥z−μz∥2z_0=\arg\min_z\Vert e^{K\tau_1}z-\mathcal E(y_1)\Vert^2+\lambda\Vert z-\mu_z\Vert^2 (λ=10−2\lambda=10^{-2}, μz\mu_z the prior mean), solved in closed form at no extra cost. Modes that the forward map damps below λ\sqrt\lambda cannot be recovered from y1y_1 anyway, so they are set to the prior mean instead of being amplified.

W1 + Q2 — Error growth with the gap is governed by identifiability (KS, clean, submitted objective, warm starts; rows 1–2: 12 problems; row 3: submitted Fig. 9a, SDA):

δf=1\delta_f=1δf=9\delta_f=9δf=20\delta_f=20
latent directions constrained by the observations (of 512)716243
linearised noise gain (no prior)1.445503
analysis error, clean: KAE / SDA / U-Net0.0023 / 0.0088 / 0.00060.011 / 0.020 / 0.0710.091 / 0.41 / 0.17

Encoder–decoder reconstruction, independent of assimilation: 0.017 relative L2L_2 on KS test states. Latent 4D-Var can go below this because it optimises z0z_0 directly rather than using E(x0)\mathcal E(x_0); optimising x0x_0 through the encoder instead stalls at 0.067 (W1 + Q1). We have removed the claim that invertibility of the latent flow implies stable inversion.

W2 + Q3 — Related work and novelty. Thank you for the references. All will be discussed in the revised Sec. 2 and a new appendix. We agree that the method itself is close to prior work:

  • Continuous-time Koopman latents with closed-form propagation are established: Pan and Duraisamy [10] learn a stable generator propagated with etKe^{tK}, and Frion et al. [12, 13] learn the continuous generator directly (K=eLK=e^{L}) or obtain it from a trained discrete operator via the matrix logarithm. [13] also compares forward and backward prediction by inverting the operator, as our pull-back does.
  • Our latent 4D-Var is the strong-constraint variant of Frion et al. [13], also used with the invertible KAE [11]: the latent initial state of a frozen Koopman autoencoder is optimised against a masked misfit. We use a Gaussian latent background instead of their spatial prior and apply it to 2-D compressible flow and KS. They also use temporal context (delay embedding). We aggregate it with an encoder.
  • Latent and operator-based assimilation is well developed: variational (Tensor-Var [3], LEVDA [6]), sequential (LD-EnSF [4], KODA [5]) and operator-theoretic [1, 2]. KNF [7], Koopa [8] and MNO [9] are forecasting models without assimilation.

We will clarify the relationship to Frion et al. (2024), which already develops continuous-time Koopman representations and latent variational assimilation and discusses differences between forward and backward reconstruction. These are foundations of our approach, and we claim no methodological novelty for them. Our work studies a specific adaptation to coupled spatial physical systems, with a shared latent generator and low-rank corrections conditioned on physical parameters such as Reynolds or Mach number. We evaluate this model for initial-state recovery and examine how dynamics, control parameterisation and initialisation affect assimilation accuracy and computational cost. The contribution is this adaptation to large spatial fields and varying physical parameters, and its controlled assessment for data assimilation. Specifically:

  1. The same surrogates evaluated both as forecasters and as assimilators across model families, on identical problems. On transonic flow the U-Net is the best short-horizon forecaster (KAE has the lowest 240-step MSE) but, with every 4D-Var baseline warm-started, KAE assimilates about 4× better (0.068 vs U-Net 0.298, FNO 0.265, ACDM+SDA 0.253).
  2. Under the tested settings the continuous linear model assimilates more accurately than discrete linear and nonlinear latent alternatives. The starting estimate and the control variable both have large effects. Latent solves are 18× cheaper than solves through an autoregressive emulator and 4D-Var adds robustness to noisy and sparse observations.
  3. A failure analysis covering noise, coverage, identifiability and backward amplification.

W3 + Q4 — Robustness to noise and sparse sensors. We tested the three explanations you proposed:

  • Backward error amplification is not the cause for the paper model: the inverse is not used in the optimisation, and its norm over our gaps is small (5 or less).
  • Limits of the latent representation contribute: the observations constrain only 62–71 of 512 latent directions, and the noise gain grows with the gap. At 10% coverage the problem is too poorly constrained for a Gaussian prior (0.54 vs 0.25 for SDA).
  • The missing explicit prior. Adding the observation-error model and latent background substantially improves robustness in the tested settings: the noise-induced error falls 3.5–4×, and at 25% coverage the KAE is better than SDA (0.029 vs 0.038; at 50% it already did).

We report noise as SNR (104, 21, 10, 5.2, 2.6). Laplace noise at matched variance behaves like Gaussian noise (within 10%) for the KAE and SDA alike. SDA keeps an edge under noise. It samples from a learned trajectory prior (both methods are given the true noise level).

W4 — Inconsistencies. Both have been corrected:

  • The InchighInc_{high} sentence was wrong: ACDM is better there (0.8 vs 2.9×10−52.9\times10^{-5}).
  • The KAE and Exponential-KAE rows were identical because RK4 at the training step and the exact exponential agree to 1.6×10−41.6\times10^{-4}. We will merge them into one row with this footnote.

We re-ran the five ablations against a full-loss model of the same architecture on identical problems (32 settings of δf\delta_f, noise, coverage, δl\delta_l, NN; 72 trajectories over three Mach regimes), and trained the current architecture with a core objective (reconstruction, multi-step prediction, forward latent consistency) against the full objective. Component by component: the latent energy term helps assimilation (without it 7–10% worse at δf=9\delta_f=9, p≤0.002p\le0.002) and does not affect forecasts; the spectral and physics terms mainly affect forecasts (18% worse at 50 frames without them); the core objective assimilates 7–17% better at full or moderate coverage but 30% worse at ≤10% coverage.

W5 — End-to-end cost and long-horizon fidelity.

  • Per-step speed: the submitted 110× mixed timing. With its latent advanced by RK4 as in training, KAE is 177× faster than ACDM per predicted step (0.80 vs 142.5 ms) and comparable to an FNO (1.2 ms). Computing the exact exponential afresh for each output time takes 5.36 ms (27×).

  • End-to-end cost at the 9-frame gap (clean, full coverage, KAE from its pull-back, wall-clock per batch, all iterations or sampling steps and draws):

    methoderrorwall-clock
    KS (δf=9\delta_f=9, 48 problems)KAE 4D-Var (8000 it., submitted objective)0.01464 s
    SDA (128 steps × 8 draws)0.021193 s (≈24 s per draw)
    U-Net 4D-Var, warm (8000 it.)0.0851149 s
    transonic (δf=9\delta_f=9, 8 problems)KAE 4D-Var (500 it., submitted objective)0.0687 s
    ACDM+SDA (4 draws)0.25345 s (≈11 s per draw)
    FNO 4D-Var, warm (500 it.)0.26558 s
    U-Net 4D-Var, warm (500 it.)0.298455 s

    At the 9-frame gap the KAE has the lowest error of these methods on both systems (1.5× and 3.7× lower than SDA), and its full solve costs 3× and 6.6× less than the SDA runs above. It is not the most accurate option everywhere: on KS the U-Net driven through the KAE decoder reaches 0.009, but takes 1165 s vs 64 s. We therefore describe the KAE as a favourable accuracy–cost trade-off. KAE's per-iteration cost is flat in the window length (≈1.9 ms for 1–50 frames, vs 15→1250 ms for U-Net 4D-Var), but total cost grows with the number of observation times and iterations; we have corrected the wording "cost independent of the assimilation window" accordingly.

  • Long-horizon fidelity. Yes, KAE oversmooths. Under one evaluator for energy and anomaly correlation, it keeps the dominant shedding frequencies but attenuates high wavenumbers and relaxes towards the mean flow: after 238 steps it retains 26% of the true fluctuation energy (ACDM 112%, best U-Nets 117–118%), and its anomaly correlation drops below 0.6 after 39 frames (ACDM and U-Nets: 52–56).

Minor — Cross-check of tables and text. We checked every table against its source files. Besides W4, every caption will state δf\delta_f, the number of trajectories, iterations, start and objective, the duplicated Table 8 will be removed, and the caption/text mismatches will be corrected.

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

This work proposes a continuous-time koopman AE whose latent dynamics are linear, z˙=K(ϕ)z\dot z = \mathbf{K}(\phi)z, with K(ϕ)=K0+LoRA(ϕ)\mathbf{K}(\phi)=\mathbf{K}_0+\text{LoRA}(\phi) conditioned on Re/Ma. The closed form z(τ)z(\tau) gives forecasts at any horizon in one single step. It also turns initial-state recovery into a latent 4DVar over z0z_0, and the cost does not grow with the window of assimilation. On the ACFM flow benchmarks and KS, the paper reports stable long roll-outs with 110x speedup over ACDM, and lower initial-state error than AR and score-based baselines. It also argues that good forecast does not mean good assimilation.

Critical strengths
  1. Scoring the same surrogates in forecast and DA at the same time is useful. On transonic flow, UNet forecasts well but can not do good assimilation. This is good to know when choosing the DA surrogate.

  2. The assimilation itself is fast and simple. eKτie^{\mathbf{K}\tau_i} is only computed once, and the observation times are off the training. One solver takes 1-2s compared with 10-100s in baselines. The Fig.9 suggests KAE has lowest error in most of the delta range on both KS and transonic flow with clean obs. KAE appears to be an efficient DA surrogate in that regime.

Critical weaknesses

My main concern is not whether KAE can do DA, because the clean-obs results already showed it can. It is 1) whether the continuous time koopman formulation really gives an advantage beyond optimizing a low-dom latent variable, and 2) whether this advantage still exist when observations are more realistic, e.g., noisy and sparse.

  1. Forecasting part is overclaimed. The section 5 states KAE outperforms ACDM on both InclowInc_{low} and InchighInc_{high}. However, table 1 reports 2.92.9 vs. 0.80.8 on InchighInc_{high}. On TralongTra_{long}, KAE has the lowest MSE but not the lowest LSiM (4.14.1 vs. 3.83.8). The 110× speedup is also relative to ACDM only, because ResNet, FNO and TF$_{Enc}$ have similar per-step cost.

  2. Assimilation degrades sharply with noisy and sparse obs. The main text (line 372) claims KAE improves over SDA by more than twice, but the appendix tables show the opposite. e.g., in tables 5, 6, 7, SDA can be better. The section 3.2 set a bayesian objective with a prior and an obs covariance RR, however, the KAE 4DVar used in the experiments, according to the algorithm in appendix, minimizes 1N∑i∣Mi⊙(D(eKτiz)−yi)∣2\frac1N\sum_i\\|M_i\odot(\mathcal D(e^{\mathbf{K}\tau_i}z)-y_i)\\|^2. The background term or noise weighting in that objective is missing, though the decoder does act as an implicit prior. This might be able to explain why it is sensitive to noise. However, sparse obs are usual DA regime...

  3. The comparison does not isolate what gives the gain of DA. Moving from UNet or FNO to KAE changes a lot at once: the model arch, the control variable, the decoder manifold also adds implicit regularization, etc. Besides, the appendix I.3 claims ACDM as a score model with a modified 20-step sampler. So it is unclear whether the gain comes from the continuous time linear dynamics or simply from optimizing a low-dim latent. This matters for the contribution.

Questions for authors
  1. How does KAE perform when the objective includes an explicit obs-error model and a background term? Does this help with the degradation under noise and low coverage?

  2. How does a continuous-time KAE compare with a discrete-time KAE and with nonlinear latent dynamics, both in forecasting and DA, given the fixed encoder-decoder and latent dim?

  3. The UNet KS error differs across the paper: ≈6×10−4\approx6\times10^{-4} in Fig. 9a. But I cannot find this number in the appendix J, where you provided many tables with only limited captions. Do they use different problem settings? Can you clarify the context of each table with some text?

3 comments

Thank you for the precise review. Continuous-time Koopman autoencoders, exponential propagation and latent variational assimilation are established foundations (Frion et al. 2024, who also compare forward and backward prediction). Our model adapts them to coupled multi-field flows with a shared latent generator and low-rank parameter-conditioned corrections, and we assess it for initial-state recovery. We answer both central questions with controlled experiments: (1) under the tested training and initialisation settings, the continuous linear KAE assimilates more accurately than discrete linear and nonlinear latent alternatives, and the starting estimate and the control variable both have large effects. KAE offers a favourable accuracy–cost trade-off rather than the lowest error everywhere. (2) with an explicit noise model and background, the KAE beats single SDA draws at 25–100% coverage on KS, while SDA remains better under observation noise and at 10% coverage.

W1 — Forecasting overclaimed. We agree on all points and will correct Sec. 5 and Table 1.

  • InclowInc_{low}: KAE beats ACDM on both metrics (MSE 1.31.3 vs 1.7×10−41.7\times10^{-4}; LSiM 6.16.1 vs 6.9×10−26.9\times10^{-2}). InchighInc_{high}: ACDM wins both (MSE 0.80.8 vs 2.9×10−52.9\times10^{-5}; LSiM 1.01.0 vs 1.71.7). TralongTra_{long}: the KAE has the lower MSE (14.914.9 vs 22.6×10−322.6\times10^{-3}), ACDM the lower LSiM (3.83.8 vs 4.1×10−14.1\times10^{-1}).
  • Speed was re-measured for every model under one protocol (240-step rollout, batch 1, ms per predicted step, encoding and decoding included): FNO 1.21, U-Nets 6.7–7.3, ResNets 9.2–9.3, ACDM 142.5. KAE takes 0.80 ms when its latent is advanced with RK4 at the training step, as in training. This is the basis of "177× faster than ACDM". Computing the exact exponential eKτe^{K\tau} afresh for each of the 240 output times (no caching, included in the time) takes 5.36 ms in our implementation, (27× faster than ACDM). Both give the same predictions to three significant figures. So the speed-up is large only against ACDM. KAE is comparable to FNO and 8–12× faster than the U-Nets and ResNets. We have corrected the abstract.

W2 + Q1 — Noise, sparsity, missing RR and BB. Agreed: the submitted runs minimised an unweighted misfit. Sec. 3.2/Alg. 1 will use by default

J(z)=12(σ2+σm2)∑i∥Mi⊙(D(eKτiz)−yi)∥2+12(z−μz)⊤Σz−1(z−μz),J(z)=\tfrac{1}{2(\sigma^2+\sigma_m^2)}\textstyle\sum_i\|M_i\odot(\mathcal D(e^{K\tau_i}z)-y_i)\|^2+\tfrac12(z-\mu_z)^\top\Sigma_z^{-1}(z-\mu_z),

with (μz,Σz)(\mu_z,\Sigma_z) from encoded training states and representation error σm\sigma_m from validation data. Results at δf=9\delta_f=9:

  • KS: 48 trajectories per setting, 8000 iterations. The clean and noise rows use the problems of Table 6. The coverage rows use a separate draw, those of Table 7.
  • Transonic: 8 trajectories per setting, 500 iterations.
  • Start: every 4D-Var method starts from the nearest observation (the KAE from its pull-back). This is why the transonic KAE values are lower than in submitted Table 3, where every method started from climatology.
settingKS: KAE, submitted objectiveKS: KAE, revised objectiveKS: SDAKS: U-Nettransonic: KAE, submitted objectivetransonic: KAE, revised objectivetransonic: ACDM+SDAtransonic: U-Net
clean0.01420.01780.02140.08540.0680.0680.2530.298
σ=0.05\sigma=0.050.21980.06260.02570.71400.0650.0620.2610.308
25% coverage0.08360.02900.03780.47070.0730.0720.2570.274
10% coverage0.57530.54190.24710.98880.0920.0880.2540.268

On KS, noise error falls 3.5–4× and KAE beats SDA at 25–100% coverage. SDA stays 1.2–2.7× better under noise and at 10% coverage. It samples from a learned trajectory prior, and both methods use the true noise level. On transonic flow, KAE is 3–5× more accurate than U-Net 4D-Var in every setting, and the revised objective is never worse (up to 25% better at σ=0.3\sigma=0.3).

W3 + Q2 — What drives the gain? We performed two controlled ablations on identical trajectories (clean, δf=9\delta_f=9, submitted objective; KS: 48 trajectories, 8000 iterations; transonic: 8 trajectories, 500 iterations).

(a) Control variable, first guess and dynamics, varied one at a time.

  • Control: latent z0z_0 (analysis D(z0)\mathcal D(z_0)) or physical x0x_0. With KAE dynamics, x0x_0 is propagated as D(eKτE(x0))\mathcal D(e^{K\tau}\mathcal E(x_0)), so the field the model actually uses is D(E(x0))\mathcal D(\mathcal E(x_0)). "U-Net, latent control": optimise z0z_0, decode with the frozen KAE decoder, roll out with the U-Net, backpropagate into z0z_0.
  • First guess: the first observation y1y_1 (9 frames after t0t_0), or the KAE's closed-form pull-back D(e−Kτ1E(y1))\mathcal D(e^{-K\tau_1}\mathcal E(y_1)), i.e. y1y_1 carried back to t0t_0.
dynamicscontrolfirst guessKS: error / timetransonic: error / time
none (first guess only)—y1y_10.183 / —0.272 / —
none (first guess only)—pull-back0.072 / —0.050 / —
KAElatentpull-back0.014 / 64 s0.068 / 7 s
KAEphysicaly1y_10.068 / 87 s0.070 / 8 s
KAEphysicalpull-back0.067 / 87 s0.061 / 8 s
U-Netphysicaly1y_10.085 / 1149 s0.298 / 455 s
U-Netphysicalpull-back0.027 / 1152 s0.146 / 454 s
U-Netphysical + Gaussian backgroundy1y_10.060 / 1158 s0.296 / 453 s
U-Netlatent (KAE decoder)pull-back0.009 / 1165 s0.070 / 456 s
  • KS: the U-Net–KAE gap (0.085 vs 0.014) has two parts. The pull-back first guess (KAE's closed-form backward map) improves the U-Net 3× (0.085→0.027). The latent control gives another 3× (0.027→0.009), more accurate than the KAE itself (0.014). With KAE dynamics a physical control stalls at 0.067 from either start: optimising through the encoder is the bottleneck, not dimensionality (the KS latent, 512, is larger than the state, 64). A physical-space background helps the U-Net only partly (0.060). Each U-Net solve costs 18× the KAE latent solve, so the most accurate KS configuration is also the slowest.

  • Transonic: KAE gives two estimates: the pull-back of the nearest observation alone, and 4D-Var combining all observations (8 trajectories per setting, 500 iterations; KAE with the revised objective):

    transonic, δf=9\delta_f=9pull-back onlyKAE 4D-VarU-Net 4D-VarFNO 4D-VarACDM+SDA
    clean, full coverage0.0500.0680.2980.2650.253
    noise σ=0.3\sigma=0.30.1420.0540.5470.5670.260
    10% coverage0.2520.0880.2680.2660.254

    With a clean, complete nearest frame the pull-back is already at the model's accuracy, and fitting later observations through an imperfect model costs a little (0.050→0.068; the U-Net from the same start drifts to 0.146). With noisy or sparse data the pull-back fails and 4D-Var is 2.6–2.9× more accurate. In every setting KAE is about 3–10× more accurate than U-Net and FNO 4D-Var. Optimising through KAE also recovers from y1y_1 (0.070), whereas U-Net 4D-Var does not (0.272→0.298). This does not depend on the iteration budget (same problems, 0–8000 iterations). With the revised objective, the error under noise or sparse coverage never increases with more iterations (8000 iterations: σ=0.3\sigma=0.3 0.054, 10% coverage 0.063), and only the clean case drifts from the pull-back (0.079). Sparse data benefit from 1000–8000 iterations, clean data does not need it.

  • The pull-back across latent dynamics: under our training settings, the discrete and nonlinear latent models give worse pull-backs on KS (0.231 and 0.268, table (b)). An invertible discrete operator also has a backward map. Ours was ill-conditioned, so we used its learned backward operator.

(b) Latent dynamics: same encoder/decoder, NzN_z, data, losses, seed, on KS:

latent dynamicspull-back first guessDA δf=9\delta_f=9forecast, lead 50
continuous linear (ours)0.0720.01420.579
discrete linear0.2310.06890.752
continuous nonlinear0.2680.25610.227

The nonlinear model forecasts 2.5× better yet assimilates 18× worse. Under the tested training and initialisation settings, the continuous linear model achieves lower assimilation error than the discrete linear and nonlinear latent alternatives, and the control experiments show substantial effects of initialisation and latent parameterisation. They do not establish that continuous time alone explains the gains. Continuous time lets us evaluate arbitrary observation times, but discrete linear KAEs also propagate efficiently. We therefore present the KAE as a favourable balance between accuracy and cost.

W3 — ACDM's modified sampler. On transonic flow ACDM+SDA uses a 20-step sampler with a three-frame blanket. It is the most accurate method at the shortest gaps (0.016 vs KAE 0.028 at δf=2\delta_f=2), but by δf=5\delta_f=5 it reaches the climatology level, and at δf=9\delta_f=9 it does not use the observations . We therefore do not treat it as evidence at δf=9\delta_f=9 : there the transonic comparison rests on warm-started U-Net and FNO 4D-Var. On KS, SDA uses the standard 128-step sampler with 8 draws.

Q3 — Why the U-Net KS numbers differ. Fig. 9a and the App. J tables used two U-Net settings

  • Fig. 9a (≈6×10−4\approx6\times10^{-4} at δf=1\delta_f=1): warm start — the nearest observation is the first guess of x0x_0; 8000 iterations, 24 trajectories. At δf=1\delta_f=1 it is already near the answer.
  • Tables 4–7 (0.095–0.19 at δf=1\delta_f=1 ): cold start from the data mean (a featureless field); Table 4 also used only 1000 iterations. From there, optimisation through the autoregressive rollout stalls far from the truth: the numbers reflect the start, not the model.
  • Fix: warm-started re-runs on exactly the Tables 6–7 problems give 0.0007 at δf=1\delta_f=1 (consistent with Fig. 9a) and 0.085 at δf=9\delta_f=9 (cold: 0.367). The warm U-Net is most accurate at the shortest gaps and KAE at longer ones (δf=9\delta_f=9 : 0.014 vs 0.085).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.