acceptodds
Under review as a conference paper at ICLR 2027

Coupled Horizons as Operator Coordinates for PDE Trajectory Prediction

Abstract

Neural PDE surrogates aim to predict finite-time trajectories from observed states. Existing approaches either advance trajectories sequentially or tie joint prediction to a prescribed temporal discretization, making cross-horizon interaction and flexible temporal queries difficult to realize simultaneously. We introduce CURVO, a horizon-coupled neural operator that treats prediction horizons as operator coordinates and represents finite-time evolution through interval-average responses. A multiscale spatial–horizon architecture with shared interval-conditioned refinement couples the queried horizons and reconstructs all requested states from a single observed state in one pass. Trained with a modest fixed query cardinality, the same operator directly evaluates query sets of different cardinalities and placements and supports dense readout of trajectories with up to one hundred states. Across five PDE configurations, CURVO reduces trajectory-averaged error by – over the strongest one-pass comparator. The formulation further extends to teacher-free one-step image generation with jointly predicted intermediate transport states. CURVO thus makes temporal resolution an inference-time choice while retaining explicit cross-horizon interaction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
operator learning, neural operators, scientific machine learning, PDE surrogate modeling, generative models, flow matching, one-pass generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CuqWOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CURVO treats prediction horizons as operator coordinates and learns interval-average responses, allowing one-pass joint prediction of multiple future PDE states with inference-time temporal resolution. The architecture combines a multiscale spatial–horizon U-Net, spectral horizon mixing, physical-coordinate attention, and shared interval-conditioned refinement. Experiments cover five PDE configurations and a CIFAR-10 extension.

Critical strengths

1.Original and well-motivated formulation: avoids sequential rollout and independent per-horizon prediction.

2.Architecture is coherent; cross-horizon coupling and variable-cardinality attention are technically sound.

3.Strong main results on several PDE families; good generalization to unseen query cardinalities, placements, and off-grid horizons.

4.Competitive memory and latency at larger query cardinalities.

Critical weaknesses

1.Limited experimental scope: mainly 64×64 2D PDEs. No 3D, irregular meshes, real-world scientific data, or long-horizon/chaotic extrapolation.

2.Theory is narrow: Proposition 1 covers linear dissipative modes; nonlinear response families lack theoretical guarantees.

3.Baseline comparison is incomplete: AR-FNO2D outperforms CURVO on both Navier–Stokes settings. Latency/FLOPs comparisons are selective and not reported across all datasets.

4.No statistical rigor: appears single-seed; no confidence intervals or variance.

Questions for authors

1.Multi-seed results: Are Table 1 results single-seed? Please report mean±std over ≥3 seeds for the main PDE benchmarks. Small ranking differences, especially vs. AR-FNO2D and U-NO, are otherwise not reliable.

2.Navier–Stokes claim: AR-FNO2D outperforms CURVO on both NS settings. What is the precise claim—accuracy, latency, memory, or compute-matched trade-off? This directly affects the “best sequential” comparison.

3.Ablation clarity: Table 3 is garbled. Please provide a clean table with exact definitions and full settings for horizon coupling, latent refinement, and joint vs. independent prediction.

4.Scope/generality: Are there any 3D, irregular-mesh, or real-world scientific datasets? If not, please state that the claim is limited to 2D regular-grid PDEs.

1 comment

We thank the reviewer for the careful assessment and appreciate the recognition of our formulation, which avoids both sequential rollout and independent per-horizon prediction, the soundness of the cross-horizon coupling and variable-cardinality attention, and the generalization to unseen query sets and horizons. Below we address the experimental scope, the theory, the comparison with AR-FNO2D, the multi-seed results and the ablation table, supported by the new Tables R1 and R2.

W1 & Q4. We thank the reviewer for this suggestion and agree that these settings would broaden the evaluation. This work focuses on the temporal side of the problem: we chose 2D benchmarks ranging from viscous decay to strongly nonlinear evolution (Table 1) and tested generalization to unseen query sets (Fig. 5, Table 10), off-grid horizons (Table 2), readouts of up to 100 states (Table 8) and a higher spatial resolution (Table 15). Sec. 5 already notes that CURVO predicts within the trained horizon window. Extending it to 3D domains, irregular meshes and real-world data mainly concerns the spatial operator and the data, and we will further discuss these directions in Sec. 5 as future work.

W2. We thank the reviewer for this comment. Linear dissipative dynamics are a special case of the nonlinear systems we study. They arise whenever the nonlinear terms vanish and, to leading order, for small perturbations of a steady state, where each spatial mode evolves at its own wavenumber-dependent rate [2]. For example, the Taylor–Green vortex is an exact solution of the 2D Navier–Stokes equations whose vorticity decays purely linearly [1]. Proposition 1 shows that even in this special case no fixed temporal basis shared by all spatial scales represents the interval responses exactly. As a necessity result, it therefore extends to every nonlinear family that contains this case, which motivates coupling horizons within each spatial scale in CURVO. We will make this argument explicit in Sec. 3.2.

W3 & Q2. We thank the reviewer for this observation. Figure 1 shows two datasets for space reasons, and the revised appendix will report accuracy, latency and parameter count for all methods and datasets. With three seeds (Table R1), AR-FNO2D is more accurate only on NS ν=10−3\nu=10^{-3}, where CURVO halves its latency (62 vs.\ 120 ms). On SWE, Gray–Scott and Active Matter, CURVO is both more accurate and 1.6–3.9× faster, with 4.9M parameters versus 16–31M for U-Net, U-NO and MeanFlow. Our claim is thus higher accuracy at lower latency with fewer parameters on most systems, which we will make explicit in the Discussion.

W4 & Q1. We thank the reviewer for this suggestion. Table 1 was single-seed. Within the rebuttal period, we have retrained CURVO and most baselines with seeds 42, 123 and 2024 and report mean ±\pm std in Table R1; the remaining baselines will follow in the revision, where we will update Table 1 accordingly.

Q3. We thank the reviewer for pointing this out. Table R2 gives a clean ablation, scored on the full trajectory as in Table 1, with its settings stated in the caption. Coupling denotes the horizon interaction Hℓ,h\mathcal H_{\ell,h} of Eq. (5), without which each horizon is predicted independently, and refinement denotes the shared interval-conditioned cell Cψ\mathcal C_\psi of Eq. (9). We will replace Table 3 with this table.

We thank the reviewer again for the constructive comments. The multi-seed results, the clarified claim on Navier–Stokes, the clean ablation table and the explicit scope aim to make the evaluation more rigorous, and we would be glad to address any remaining questions.

[1] G. I. Taylor and A. E. Green. Mechanism of the production of small eddies from large ones. Proceedings of the Royal Society of London. Series A, 158(895):499–521, 1937. [2] M. C. Cross and P. C. Hohenberg. Pattern formation outside of equilibrium. Reviews of Modern Physics, 65(3):851–1112, 1993.

Reviewer_Qk4gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents CURVO, that predicts the average change of the PDE for a time interval h. In order to perform predictions for a whole trajectory, it uses attention and FNO blocks to predict the output query time steps in one go, allowing information to flow between them. Additionally, the model include a refinement option, where it splits all query times into L segements, and predicts the next segment given the previous one. The experiments compare to a variety of models, showing that CURVO outperperforms on most tasks, and include ablations e.g. about the refinement.

Critical strengths
  1. The core setup, with the interval-average response and the refinement seems interesting and novel.
  2. The model performs empirically well on SWE, Gray–Scott and Active Matter PDEs. The experiments contain many baselines and some relevant ablations.
Critical weaknesses
  1. The main weakness is that the method description is hard to follow. For example:
  • It seems unorganized and mixes the architecture (e.g. using attention and Fourier blocks) with the main model setup (coupled predictions, refinement, ...)
  • The motivation for the scale-dependent horizon structure (Eq. 4) seems not really relevant for this model in specific, as the U-shaped hierarchy is pretty standard. Hence this part just interrupts the reader from following the method.
  • Some parts are not explained at all, e.g. the decoder.
  1. The paper mixes the change to predicting the interval-average response, which is one of the claimed central contributions, with a lot of architectural choices. It remains unclear if this is now in general a better way to formulate learning a PDE.
Questions for authors
  1. Did you also compare to a model, that evolves the PDE in latent space (e.g. AROMA, CALM-PDE)? The refinement in latent space appears to be related to these approaches.
  2. Did you measure an advantage of predicting the average change v to just predicting the next state, independtly of the coupled predictions?
  3. Did you also ablate the architecture choices (FFT blocks, attention)?
1 comment

We thank the reviewer for the careful assessment and appreciate the recognition of our core setup and refinement, the results on SWE, Gray–Scott and Active Matter, and the breadth of baselines and ablations. Below we address the presentation of the method, the role of the target, and the requested comparisons and ablations, supported by the new Tables R1 and R2.

W1. We sincerely thank the reviewer for the careful reading and the valuable comments on the presentation. We will reorganize Sec.3 so that the formulation, namely interval-average responses, their joint prediction and refinement, comes first and the architecture follows. Architectural details, including a full description of the decoder, which follows the standard U-Net design, will move to the appendix. For Eq. (4), Sec.3.2 will keep only its consequence for the design, that horizons are coupled within each spatial scale (Proposition 1), and the derivation will move to the appendix.

W2. We agree that the target and the architecture should be distinguished. The interval-average target builds on MeanFlow's average velocity, as cited in Sec.3, and is not among our stated contributions. Our contribution is to learn the full response curve of an observed state over continuous horizons, with a multiscale spatial–horizon architecture that reconstructs an inference-chosen query set jointly in one forward pass. MeanFlow uses the same target for single intervals, and CURVO is more accurate on all four PDE datasets with fewer parameters (4.9M vs.\ 16.4M), so the gains do not come from the target alone; Q2 below isolates its effect.

Q1. We thank the reviewer for this suggestion and agree that the refinement is related to latent-space evolution. AROMA and CALM-PDE are closely related latent-space models, and we added the more recent CALM-PDE (NeurIPS 2025) to Table R1; CURVO is more accurate on all five configurations over the trajectory. Two differences remain: the refinement uses a fixed number of stages (L=2L=2) for any horizon instead of one step per Δt\Delta t, and each stage updates the latent features of all queried horizons jointly, at subinterval coordinates, without decoding intermediate states.

Q2. We thank the reviewer for this question. Table R2(a) compares the two targets with the same backbone and the same loss on the state, without coupling and refinement. Predicting vˉ\bar v gives lower error than predicting the state directly on NS, Gray–Scott and Active Matter, while direct state prediction is better on SWE. Since u(t+h)=ut+hvˉu(t+h)=u_t+h\bar v, errors in vˉ\bar v are scaled by hh, which keeps predictions near the observed time accurate. We attribute the SWE exception to the moving dam-break front, whose change between two times is sharper than the state itself. We will add this comparison to Sec.4.3.

Q3. We thank the reviewer for this question. Table R2(c) ablates the two horizon branches while keeping coupling and refinement: the spectral transform mixes horizons at the intermediate resolutions, and the attention uses the physical horizon offsets at the bottleneck. Using either branch alone raises the error, most on NS and Active Matter. Table 3(b) further controls refinement depth and interval conditioning. We will include Table R2 in Sec.4.3.

We thank the reviewer again for the constructive comments. The restructured Sec.~3, the separate analysis of the target, and the new comparisons and ablations aim to clarify the method and the source of its gains, and we would be glad to address any remaining questions.

Reviewer_KE1QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CURVO, a neural operator that predicts multiple future PDE states from one snapshot in a single pass, using interval-average responses at query horizons chosen at inference. Experiments on four PDE families show lower error than one-pass baselines and transfer to unseen query sets, plus an extension to one-step CIFAR-10 generation.

Critical strengths

• Consistent gains over one-pass baselines with a small model (4.9M parameters). • The generalization experiments (query cardinality, layout, off-grid horizons, 128² resolution) are fairly thorough.

Critical weaknesses

• The novelty is mostly architectural. The interval-average target is essentially MeanFlow's average velocity, and both lead-time conditioning and joint multi-time prediction already exist. The closest methods (VCNeF, CViT, P-STMAE) are discussed but not compared. • The horizon FFT in Eq. (6) runs over the query index rather than physical time, so the prediction at a given horizon depends on the other queries. Figure 5(a) shows this clearly: the SWE endpoint error is 0.0095 with the end layout vs. 0.0022 with uniform at K=4. This weakens the main claim that temporal resolution is a free inference-time choice. • Baselines look undertrained (CFO above 1.0 relative error on NS ν=10⁻⁴ and Active Matter, TANTE diverging on Gray–Scott), and all results come from a single seed while some margins are only 5–7%.

Questions for authors

• Gray–Scott default config is 0.11991 in Table 3(a) but 0.12991 in Table 3(c), and Active Matter at H_max is 0.9703 in Fig. 5(a) vs. 0.8712 in Table 1. Which numbers are correct? • What is the "DR" dataset in Figure 5? • Is Table 1 computed with one pass over all horizons or with K=8 contexts? • On Active Matter all methods are above 0.70 error. What does persistence give? • For CIFAR-10, what is the training target, and why is your own MeanFlow run (FID 3.81) left out of the main-text comparison?

2 comments

We thank the reviewer for the attention given to our work and the rigorous assessment. We appreciate your recognition of CURVO’s parameter efficiency and of our generalization experiments. Your critique touches on central questions of multi-horizon operator learning, and we welcome the opportunity to clarify the novelty, query dependence and reliability of our comparisons.

W1. We agree on the connection to MeanFlow: the interval-average target builds on its average velocity, as cited in Sec.~3, and is not among our stated contributions. Our contribution is to learn the full response curve of an observed state over continuous horizons, with a multiscale spatial–horizon architecture that reconstructs an inference-chosen query set jointly in one forward pass. Trained with eight queries, CURVO reads from four horizons up to the complete 100-state Gray–Scott trajectory, and is more accurate than MeanFlow on all four PDE datasets with fewer parameters (4.9M vs.\ 16.4M).

Lead-time conditioning and joint multi-time prediction have precedents (Sec.~2), but Poseidon, RIGNO, DeepONet and VCNeF answer each horizon independently, FNO3D requires a regular grid, and AR-FNO2D and CViT need h/Δth/\Delta t autoregressive calls. CURVO couples an inference-chosen set of horizons in one evaluation from a single state. We added VCNeF, CViT and CALM-PDE to Table R1 under the same protocol, and CURVO is more accurate than all three on all five configurations over the trajectory.

W2. We agree that Eq. (6) mixes horizons along the query index, so a prediction depends on its companion queries. This dependence comes with explicit cross-horizon interaction, and Sec. 5 reports its effect for concentrated layouts on NS and SWE. To isolate the influence of the companions, Appendix 7.3 holds a target horizon fixed and changes only its companion queries. Across 16 random contexts with KK from 4 to 32, the target prediction changes by at most 1.1%1.1\% on NS and 1.0%1.0\% on Gray–Scott, below its own error, and more companions lower its error (NS endpoint 0.0154 at K=8K=8, 0.0144 at K=32K=32). We also replaced the index transform with the physical-offset attention of Eq. (7) and retrained, and the dependence remains on SWE (Table R2) and NS, so it stems from the coupling itself. We thank the reviewer for highlighting the effect of query placement and will discuss it in Sec. 5.

W3. Both baselines converge in training. TANTE's losses on Gray–Scott decrease throughout, and its NaN arises only in 20 of 120 test trajectories, which diverge between rollout steps 67 and 100 (Fig.~4c). CFO has no published configuration for these datasets; retrained with the optimizer settings of its published 2D recipe, it improves from 1.85 to 1.20 on Active Matter and reaches 0.93 on NS (ν=10−4\nu=10^{-4}), still well above the other methods. We trained CURVO and its closest baselines with seeds 42, 123 and 2024 (Table R1): the margins on Gray–Scott and Active Matter exceed the seed variation, while on SWE the gap to U-NO lies within it. We will update Table 1 accordingly.

Q1. We thank the reviewer for catching these. The Active Matter endpoint in Table 1 should read 0.9712 (single seed 42); 0.8712 is a typo, and the value 0.9703 in Fig. 5(a) is obtained with eight queries rather than the full trajectory. The two Gray–Scott values come from the same checkpoint, 0.11991 on the full trajectory as in Table 1 and 0.12991 on the eight fixed test horizons used in Table 3(c), which matches the K=8K=8 value in Fig. 5(b). We now report Table 3(c) on the full trajectory and state the scoring in every caption.

Q2. We thank the reviewer for pointing this out. DR is the 2D diffusion–reaction dataset of PDEBench (Takamoto et al., 2022), a FitzHugh–Nagumo activator–inhibitor system. We evaluated it as well, but because of the page limit and its overlap with Gray–Scott as a reaction–diffusion system, it appears only in Fig. 5. We will add its description and full results to the appendix.

Q3. We thank the reviewer for this question. Table 1 uses a single forward pass per anchor, with every recorded step from the anchor to Hmax⁡H_{\max} as the query context (K=49K=49, 29, 100, 100 and 40 for the five configurations), i.e.\ 3.6 to 12.5 times the training cardinality Ktr=8K_{\mathrm{tr}}=8. We will state this clearly in the revision.

Q4. We thank the reviewer for the suggestion. Persistence (u^=ut\hat u=u_t) gives 1.0418 on the full trajectory (1.1797 at the endpoint). CURVO reduces this error by about 30%30\%, and all one-pass methods improve on it, so the errors above 0.70 reflect the difficulty of predicting Active Matter from a single state. We will report the persistence values in Sec. 4.1.

Q5. We thank the reviewer for the question. The CIFAR-10 target is the interval-average velocity of the noise–data transport, obtained from the MeanFlow identity at Ktr=4K_{\mathrm{tr}}=4 jointly queried intervals, without a teacher. Our MeanFlow run reaches a better FID; it was kept in Table 5 because the main text focuses on joint trajectory readout, and we will add it to the main-text comparison.

We sincerely appreciate your thorough review, particularly on the relation to MeanFlow and the query dependence. With the new comparisons and multi-seed experiments, the revised manuscript reflects the strengths and current limitations of CURVO more accurately. We thank you for guiding us toward a more rigorous presentation of our work.

Comments on the paper

2 comments

Table R1. Relative L2L^2 error over the trajectory (Rel.) and at Hmax⁡H_{\max}, mean ± std over seeds 42, 123 and 2024 where reported. Scales: NS ν=1e-3 and SWE ×1e-3; NS ν=1e-4, Gray-Scott and Active Matter ×1e-2. Best in bold, second best in italics.

MethodNS 1e-3 Rel.NS 1e-3 @HmaxNS 1e-4 Rel.NS 1e-4 @HmaxSWE Rel.SWE @HmaxGS Rel.GS @HmaxAM Rel.AM @Hmax
Spatial-backbone direct predictors
ResNet481.1 ± 10.3583.8n/an/a4.4 ± 0.13.9 ± 0.025.8 ± 0.032.0 ± 0.073.7 ± 0.286.8 ± 0.4
UNet26.8 ± 2.652.7 ± 5.210.3 ± 0.628.8 ± 1.03.1 ± 0.32.3 ± 0.613.2 ± 0.419.3 ± 0.375.897.3
UNO11.2 ± 0.334.2 ± 2.37.9 ± 0.228.6 ± 0.43.0 ± 0.22.3 ± 0.118.6 ± 1.324.9 ± 1.377.7 ± 0.9102.4 ± 0.3
CNO89.0212.626.4 ± 16.355.8 ± 22.87.0 ± 0.67.1 ± 0.229.445.476.4 ± 0.885.6 ± 0.3
Spectral space-time operators
FFNO14.044.27.227.43.12.125.932.277.898.1
FNO3D362.3 ± 32.7806.4 ± 104.943.5 ± 7.894.0 ± 15.650.3 ± 1.010.6 ± 0.225.8 ± 0.032.0 ± 0.093.9 ± 4.499.6 ± 1.3
Single-interval response prediction
MeanFlow65.1 ± 6.5120.0 ± 12.519.1 ± 0.441.5 ± 0.94.3 ± 0.23.0 ± 0.224.4 ± 2.133.2 ± 4.679.3 ± 0.395.6 ± 0.1
Continuous-time neural fields
VCNeF†34.7 ± 1.249.8 ± 1.814.5 ± 0.329.9 ± 0.94.4 ± 0.15.0 ± 0.214.8 ± 0.120.5 ± 0.176.7 ± 0.390.8 ± 0.5
Sequential evolution
AR-FNO2D6.5 ± 0.311.6 ± 1.310.9 ± 6.620.3 ± 8.68.3 ± 0.29.1 ± 0.525.0 ± 0.439.4 ± 1.090.5 ± 4.2172.9 ± 47.3
Flow Matching25.7 ± 2.344.2 ± 5.110.7 ± 1.533.0 ± 4.26.1 ± 0.86.3 ± 0.627.6 ± 2.839.0 ± 1.077.3 ± 1.6113.3 ± 2.4
TANTE30.9 ± 2.352.1n/an/a38.0 ± 4.149.1 ± 7.3n/an/a87.3 ± 1.3118.9 ± 0.7
CFO109.4 ± 52.0277.2 ± 157.793.3278.938.1 ± 15.5115.1 ± 39.426.1 ± 3.946.6 ± 14.9119.7 ± 15.2214.3 ± 70.1
CALM-PDE†29.0 ± 8.044.8 ± 12.87.8 ± 1.419.7 ± 2.94.6 ± 1.44.3 ± 1.518.823.974.2 ± 0.389.1 ± 1.1
CViT†108.5 ± 33.1190.9 ± 56.525.9 ± 1.563.0 ± 1.87.0 ± 0.68.1 ± 3.129.0 ± 1.438.0 ± 1.096.0 ± 1.7124.6 ± 3.4
CURVO (ours)7.5 ± 0.415.3 ± 1.26.1 ± 0.119.1 ± 0.32.7 ± 0.62.0 ± 0.411.9 ± 0.118.0 ± 0.170.8 ± 0.798.0 ± 1.2

Table R2. Ablations, full-trajectory relative L2L^2, single seed (42), scaled as in Table R1. Best in bold.

VariantNS 1e-3 (×1e-3)SWE (×1e-3)Gray-Scott (×1e-2)Active Matter (×1e-2)
(a) Prediction target, without coupling and refinement
vˉ\bar v, loss on state7.962.6712.0672.95
state10.171.8512.6579.19
(b) Coupling and refinement
neither10.573.8113.8581.35
w/o refinement8.692.8512.4078.44
w/o coupling9.953.4313.6481.76
(c) Horizon mixing
w/o attention8.302.9412.3375.16
w/o FFT8.173.1612.0075.47
CURVO (ours)7.222.8011.9970.15

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their careful reading and constructive feedback. We are glad that the reviewers found the formulation original and well motivated, the generalization experiments thorough, and the model small with competitive latency and memory. We summarize the main changes below; Tables R1 and R2 are attached to this comment.

  1. Reliability of the evaluation. CURVO and most baselines are retrained with seeds 42, 123 and 2024, and Table R1 reports mean ± std. The evaluation protocol, including the one-pass readout of all recorded horizons in Table 1, will be stated in every caption.

  2. Additional comparisons. VCNeF, CViT and the latent-space model CALM-PDE are added to Table R1, and CURVO is more accurate than all three on all five configurations over the trajectory.

  3. Ablations. Table R2 separates the prediction target, coupling and refinement, and the two horizon branches, all scored on the full trajectory.

  4. Presentation and scope. Sec. 3 will present the formulation before the architecture, with architectural details and the derivation of Eq. (4) moved to the appendix. The relation of the target to MeanFlow, the effect of query placement, and the current scope of 2D regular grids within the trained horizon window will be stated explicitly.

We hope these additions address the main concerns and are happy to clarify any remaining questions.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.