acceptodds
Under review as a conference paper at ICLR 2027

ChronoSync: Recovering Long-Horizon Coupled Dynamics via Cross-Field Synchronization of Decoupled Diffusion Models

26 views

Abstract

Long-horizon prediction of coupled multiphysics PDEs is challenging even when fully coupled trajectories are available for training. We study a harder setting in which the target joint dynamics are unavailable during training and only decoupled conditional trajectories for individual physical fields are observed. We introduce ChronoSync, an inference algorithm that recovers coupled dynamics by synchronizing independently trained conditional Elucidated Rolling Diffusion Models. ChronoSync sequentially composes field-wise probability flows within each denoising step using operator splitting, while a rolling temporal window propagates the synchronized state over long PDE time horizons. We characterize the local composition perturbation through Lie brackets and derive a recursive bound for its propagation across rollouts. Under a mixed regime which allows for varying degrees of local rollout stability in the perturbations, we show that the long-horizon splitting error can remain bounded. We evaluate ChronoSync on FitzHugh–Nagumo, viscous Burgers, Darcy transport, and Rayleigh–B'enard convection, spanning increasingly complex multiphysics interactions from bidirectional linear coupling to nonlinear transport, density-driven flow, and buoyancy–advection feedback under incompressibility. Across these case studies, ChronoSync recovers close to 90% of the conditional to joint oracle error gap on average and is generally the best decoupled data method, substantially approaching models trained directly on coupled trajectories. ChronoSync additionally recovers physically meaningful coupled observables, closely tracking solute transport in Darcy flow and convective heat transport in Rayleigh-B'enard convection. These results demonstrate that long-horizon coupled dynamics can be recovered from decoupled generative models by synchronizing their probability flows across diffusion time and propagating that synchronization coherently across PDE time.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
conditional generative models, probability flow ODEs, operator splitting, multiphysics PDEs, long-horizon forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rgmWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ChronoSync targets long-horizon prediction of coupled multiphysics PDEs when training data come only from decoupled simulations, where each field is solved with the other fields prescribed. One conditional Elucidated Rolling Diffusion Model (ERDM) is trained per field. At inference, inside each Heun step the field models are swept in order, each conditioned on the latest under-relaxed (κ=0.25\kappa = 0.25) estimate of the other fields, and the rolling window is advanced in PDE time with one bootstrap refinement. The paper characterizes the splitting error through Lie brackets and a rollout recurrence. On FitzHugh–Nagumo, Burgers, Darcy transport and Rayleigh–Bénard, ChronoSync cuts error by 53–74% relative to the best decoupled baseline on the three harder systems, recovers about 90% of the conditional-to-oracle gap, and tracks solute centre of mass and the Nusselt number.

Critical strengths
  1. The gains over decoupled baselines are large and consistent on the harder systems (Rayleigh–Bénard velocity 0.663 → 0.217, Darcy q 0.141 → 0.036 vs GenCP-AR), and on Darcy they hold at every window length tested (Fig. 22). The rolling design also avoids the memory blow-up of full-trajectory M2PDE. Composing separately trained field models over long horizons is a useful problem.
  2. The evaluation goes beyond field error: Darcy solute centre of mass, Rayleigh–Bénard Nusselt number, and extrapolation from 300 to 399 frames, where ChronoSync stays close to FNO-AR-Joint.
Critical weaknesses
  1. The gains are not attributed to the synchronization itself, and the existing ablations do not support the extra components. The baselines use different backbones and samplers, and there is no same-ERDM control (window-level alternation, Jacobi update). Where κ\kappa and bootstrap are ablated (FHN and Burgers only), they do not help: κ=1\kappa = 1 matches or beats the default 0.25 (Figs. 16, 19), and on Burgers no bootstrap beats nboot=1n_{\mathrm{boot}} = 1 (c≈0.020c \approx 0.020 vs 0.025, Fig. 21), yet both are used in all experiments. The comparison with the coupled oracles also depends on the window. Darcy uses W=10W = 10, which is ChronoSync's best window and the oracles' worst; at W=48W = 48, ERDM-Joint reaches c≈0.03c \approx 0.03 against ChronoSync's best 0.075 (Fig. 22). The "90% of the gap recovered" figure inherits this choice.
  2. The theory does not cover the implemented algorithm. Eq. (3) sums local Lie-bracket defects without propagating them through later denoising steps, the rollout recurrence drops the O(∣δ∣2)O(\\|\delta\\|^2) terms, and the mixed-regime contraction in Corollary 1 is assumed but never measured. The analysis is for plain Lie–Trotter, not the relaxed sweep with bootstrap used in the experiments. It explains the idea but does not support the long-horizon stability claim.
Questions for authors
  1. Can you add same-backbone ablations (ERDM with window-level alternation, Jacobi update, κ=1\kappa = 1), κ\kappa/bootstrap sensitivity on Darcy and Rayleigh–Bénard, and inference cost (NFE, wall-clock) against the baselines?
  2. Can you report Table 1 and the gap-recovered numbers with each method at its own best window length? Can you also report estimates of LrL_r or of the Lie-bracket terms along real rollouts?
  3. In the released code (FHN/Burgers only), "lie" mode conditions the second field on the first field's unrelaxed output, while Algorithm 1 uses the relaxed estimate. Which one produced Table 1? Also, the conditional data use one pre-iteration, which already carries some coupling information, and for Rayleigh–Bénard it costs more than the coupled simulation (Table 7). Which practical setting is this method meant for?
5 comments
AuthorsOpenReview ↗

Response to Reviewer rgmW

We thank the reviewer for the careful reading of our paper and for providing valuable feedback. All references below are to the revised PDF, where changes are marked in blue.

W1 / Q1. Same-backbone ablations, kappa/bootstrap sensitivity, and inference cost

Window-level alternation (new Table 18, Appendix K.5). We keep the conditional ERDMs, rolling procedure, 5 Heun steps and number of network evaluations per committed frame fixed. Only the cross-field exchange frequency changes.

Cross-field exchangeDarcy ccDarcy qqRBC TTRBC (ux,uy)(u_x,u_y)
Every Heun step (ChronoSync)0.0800.0380.0710.229
Once per window0.0870.0480.1170.339
Error change+8.75%+26.3%+64.7%+48.03%

Exchanging once per window raises the error on every field. The increase is largest on Rayleigh–B'enard (RBC), where the coupling is strongest and the window is longest. Five Heun steps match the 32 step accuracy of Table 1 (Table 17), so this comparison applies directly to the main results.

κ\kappa and bootstrap on Darcy and RBC (new Figures 24, 25, 27, 28).

  • κ\kappa. Varying κ∈{0.25,0.5,1}\kappa \in \{0.25, 0.5, 1\} changes the errors little. κ=1\kappa=1, the exact sequential Lie–Trotter sweep we analyse, is within 6.7% of the reported κ=0.25\kappa=0.25 on every field. Relaxation is therefore an optional damping, not the source of the accuracy. We therefore kept κ=0.25\kappa=0.25 in Table 1.
  • Bootstrap. We agree that the bootstrap does not help on FitzHugh–Nagumo and Burgers. On the strongly coupled systems, however, one rolling bootstrap pass is helpful. Removing it raises the Darcy qq error from 0.036 to 0.096, the RBC TT error from 0.074 to 0.088, and the RBC u\mathbf{u} error from 0.217 to 0.276. A second pass does not help further.

Inference cost (new Appendix I, Table 14). RBC, W=55W=55, batch of 5 trajectories, 55 committed frames, one GPU. NFE is the total number of network evaluations for the run.

MethodStepsNFEWall-clock (s)Peak memory (GiB)
M2PDE-AR10004,000862.269.82
GenCP-AR (CNO backbone)102041.6734.13
ChronoSync3213,8602989.7810.64
ChronoSync51,980428.910.64
ERDM-Joint323,465749.99.90

How the NFE counts arise:

  • ChronoSync: 55 rollouts × (2S−1)(2S-1) Heun evaluations × 2 fields × 2 passes (one bootstrap).
  • ERDM-Joint: 55 rollouts × (2S−1)(2S-1) evaluations of one joint model.
  • M2PDE-AR: one 55-frame block × 1000 DDPM steps × 2 fields × 2 outer iterations.
  • GenCP-AR: one block × 10 flow steps × 2 fields.

With 5 Heun steps, ChronoSync is 7 times faster than with 32 steps, at comparable accuracy (Tables 14, 17).

W1 / Q2. Window length and gap recovered at each method's best window

On Darcy, ChronoSync prefers short windows and the oracles prefer long ones (Figure 23). On RBC, every method improves with WW (Figure 26). Table 1 uses the same window per case study for all methods for consistency. Whereas, window ablations are now provided for all the case studies.

  • RBC (Figure 26). ChronoSync is the best decoupled method and lies below the same-backbone ERDM-Joint (coupled data) at every W∈{10,25,55,75}W \in \{10, 25, 55, 75\}.
    • At its best window (W=75W=75), ChronoSync reaches TT 0.058 and u\mathbf{u} 0.188.
    • At its best window (W=75W=75), ERDM-Joint reaches 0.097 and 0.333.
    • The paper's W=55W=55 is therefore not ChronoSync's best case.
  • Darcy (Figure 23). On cc at W=48W=48, ERDM-Joint (c≈0.03c \approx 0.03) is better than ChronoSync at its best window (0.075). ChronoSync remains the best decoupled method at every Darcy window.

Gap recovered at each method's best window. The gap recovered is (econd−eCS)/(econd−eFNO)(e_{\text{cond}}-e_{\text{CS}})/(e_{\text{cond}}-e_{\text{FNO}}). The conditional error does not depend on WW, so the gap changes only through ChronoSync's best window (numerator) and FNO-AR-Joint's best window (denominator). Taking each at its own best window (except for FHN and Burgers because FNO-AR-Joint is already doing very well):

FieldTable 1Each method at its best WW
Darcy qq (CS W=10W=10, FNO W=48W=48)94.7%92.0%
Darcy cc (CS W=10W=10, FNO W=48W=48)70.8%58.1%
RBC TT (both W=75W=75)91.9%96.4%
RBC u\mathbf{u} (both W=75W=75)90.5%92.4%
Average over all seven fields91.9%90.6%

The Darcy gaps fall because FNO-AR-Joint improves at long windows. The RBC gaps rise because ChronoSync improves at long windows. For FitzHugh–Nagumo and Burgers, we keep the Table 1 values (96.5–99.7%), where the gaps are already close to 100%. The headline of about 90% therefore does not depend on the window choice.

AuthorsOpenReview ↗

Response to Reviewer rgmW (continued)

W2 / Q2. Theory and the implemented algorithm; estimates of LrL_r

  • Propagation of local defects (equations 2–3, Appendix C.1.2). The splitting analysis now covers the Heun type Lie–Trotter scheme that ChronoSync runs. Local approximation errors are propagated through the newly introduced Heun-type analysis scheme (Equations 2, 3, 34–35). Lemma 1 discusses bounds for the global error from per-rollout defect through ∥SW∥C3h2\|\mathcal{S}_W\|C_3h^2, so they enter the rollout bound (equation 57) as an explicit term rather than being dropped.
  • Measured LrL_r products, replacing the assumed contraction (new Section 5, Table 2, Figure 4). We inject the measured splitting defect into real rollouts and track the committed-frame response. This response is a proxy for ∏qLq\prod_q L_q. Both systems are in the mixed regime of Corollary 1:
    • Darcy contracts at ρ^=0.992\hat\rho = 0.992–0.9930.993 per rollout.
    • RBC amplifies at ρ^=1.011\hat\rho = 1.011–1.0201.020 while convection develops, then contracts at 0.9970.997–0.9980.998 after r≈220r\approx220.
  • The O(∥δ∥2)\mathcal{O}(\|\delta\|^2) terms. Dropping these is the standard linearized stability step for non-autonomous discrete-time systems (Potzsche, 2010). The measurement in Section 5 is made in this linear regime.
  • Relaxed sweep and bootstrap (new Remark 1). The analysis covers κ=1\kappa=1. With one bootstrap pass, the rollout map is a composition of two coupled Heun passes, so Theorem 2 applies unchanged. The relaxation κ<1\kappa<1 is empirically within 6.7% of κ=1\kappa=1 (see above).

Q3. Released code vs. Algorithm 1, pre-iteration, and the target setting

  • Code vs. Algorithm 1. Thank you for checking the code. The code is the reference implementation and produced all reported results, including Table 1. Each field is conditioned on its partner's freshest, unrelaxed estimate, and the κ\kappa-relaxation is applied once after the sweep. Algorithms 1 and 3 and the text of Section 3.4 and Appendix B are now corrected to match the code. The two orderings coincide at κ=1\kappa=1, so the theoretical analysis is unaffected.
  • Pre-iteration.
    • FitzHugh–Nagumo and Burgers use no pre-iteration. Each field is solved against a partner sampled independently from a Gaussian random field (Appendix E.1–E.2), and ChronoSync recovers 96.5–99.7% of the gap there.
    • For Darcy and RBC, a single pre-iteration keeps the prescribed fields physically admissible. The resulting conditional data remain far from the coupled dynamics. RBC conditional errors are 0.219 (TT) and 0.977 (u\mathbf{u}) in Table 1, and they diverge over time in Figure 2. Same is the case for Darcy too.
  • Target setting. We agree that on RBC the two conditional solves together cost more than one coupled solve (Table 8). Our target is the setting where coupled solves are unavailable and when conditional solves are grossly inaccurate (as in our setting; Figure 2). Separate single physics codes, legacy solvers that cannot be coupled, or observations of one field given the others. For Darcy, the conditional solves are also cheaper (46 s vs. 130 s).

We hope these additions address the reviewer's concerns, and we would be glad to clarify anything further.

Thank you for the additional experiments and clarifications. The same-backbone ablation, bootstrap tests, and window-length analysis substantially address my empirical concerns. The stability measurements are also useful, although the theoretical coverage of the relaxed implementation remains limited. These additions strengthen my support for the paper, and I will maintain my weak-accept rating.

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

ChronoSync composes separately trained conditional Elucidated Rolling Diffusion Models (ERDMs) at inference time: within each Heun step it applies the field-wise probability-flow updates in sequence (Lie–Trotter splitting), and a rolling window advances the synchronized state in PDE time. A Lie-bracket analysis bounds the local splitting error (Theorem 1) and its propagation across rollouts (Theorem 2). On FitzHugh–Nagumo, viscous Burgers, Darcy transport and Rayleigh–Bénard convection, ChronoSync improves on decoupled-data baselines and recovers about 90% of the gap between conditional models and coupled-data oracles, with additional checks on physical observables and time extrapolation.

Critical strengths
  1. The paper studies a useful and less explored setting in which coupled trajectories are unavailable for training, and gives a concrete inference-time procedure for coordinating field-wise generative models across both diffusion time and PDE time.
  2. The evaluation covers four coupled systems of increasing difficulty and includes both decoupled-data baselines and coupled-data oracles. Reporting solute transport and the Nusselt number alongside field-wise errors (Sec. 4.4) makes the Darcy and Rayleigh–Bénard results especially informative.
Critical weaknesses
  1. Theorem 1 (Sec. 3.3) and Theorem 2 (Appendix C.4–C.5) bound the error of composing the learned field-wise flows, but they do not show that independently trained conditional models are mutually consistent with a joint model, which is what recovering the coupled dynamics requires. The long-horizon bound further assumes that cumulative rollout sensitivities decay geometrically (Sec. 3.3), and this quantity is never measured in Sec. 4. The theory therefore explains splitting stability, but not when the method should succeed.
  2. For Darcy transport and Rayleigh–Bénard, the conditional training data are produced by PDE-specific windowed pre-iteration (Sec. 4.1, Appendix E.3–E.4), which presumes a solver for each sub-problem and is far more structured than the fragmented-observation setting that motivates the paper. Inference cost is also not reported: Table 7 gives only data-generation time, while ChronoSync runs 32 Heun steps per window plus bootstrap refinement (Table 12).
Questions for authors
  1. (Clarification) Can you report the rollout-sensitivity products assumed by Theorem 2, or another diagnostic relating local splitting error to long-horizon error? What evidence supports compatibility among the independently trained field-wise flows?
  2. (Clarification) What are the runtime and peak memory of a 300-frame rollout, compared with ERDM-Joint, and how do they change with fewer Heun steps or without bootstrap refinement? How much PDE-specific knowledge is needed to build the conditional data when no sub-problem solver is available?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Wwam

We thank the reviewer for the careful reading of our paper and for providing valuable feedback. All references below are to the revised PDF, where changes are marked in blue.

W1 / Q1. Rollout-sensitivity products and compatibility

Measured rollout sensitivities (new Section 5, Figure 4, Table 2, Appendix L). The geometric decay of Corollary 1 is now measured rather than assumed. At injection rollout r0r_0, we perturb the rolling state along the splitting defect that ChronoSync actually produces at r0r_0, estimated by step doubling (SS vs. 4S4S Heun steps). We then track the committed frame response cr0(ℓ)c_{r_0}(\ell) under fixed noise. To the first order, cr0(ℓ)c_{r_0}(\ell) is the loss/gain of the rollout propagator along the defect direction. Hence, it is a measured proxy for the sensitivity products ∏qLq\prod_q L_q in equation 5. Fitting the form of Corollary 1 gives the following.

Fitted ρ^\hat\rho per rolloutr0=5r_0=5r0=54r_0=54r0=144r_0=144
Darcy (W=10W=10)0.9930.9921.009
Rayleigh–Bénard (W=55W=55), r≤220r \le 2201.0111.0131.020
Rayleigh–B'enard (W=55W=55), r>220r > 2200.9970.9981.003

Both systems are in the mixed regime. The defect is amplified while the Darcy finger-like plume or the Rayleigh–B'enard convection develops and then contracted afterwards. For Darcy injections at r0≤54r_0 \leq 54, the response ends at 7–21% of its value at the first window boundary.

This diagnostic also speaks to when the method should succeed. It links the local splitting error to the long-horizon error (from Figure 2), and it requires no coupled reference data. Yet, the responses follow the shape of the relative ℓ2\ell_2 error curves (Figures 2 and 4).

Compatibility. We agree that the theory does not establish that independently trained conditionals are compatible with a joint model. It bounds the error of composing the learned flows relative to their summed flow, as stated in Section 3.3 ("Scope of the analysis"). Compatibility is an open question for compositional generative models in general (e.g., priors works GenCP ICLR 2026 and M2PDE ICML 2025) and we plan to address this important topic as part of future work (now clearly stated in Section 6 "Conclusions and Limitations"). In addition, our conditional training data has substantial errors with respect to the joint/coupled data as seen in Figure 2. Still, ChronoSync is able to recover close to 90% of the conditional to joint oracle gap on average. On Rayleigh–B'enard it is more accurate than the same-backbone ERDM-Joint trained on coupled data, both on the stress-test set (Table 1) and on 20 randomly drawn test trajectories (Table 16).

W2 / Q2. Inference cost

The new Appendix I (Table 14) reports wall-clock time and peak memory on Rayleigh–B'enard (W=55W=55, batch of 5 trajectories, 55 committed frames, one GPU):

MethodStepsTotal (s)Per frame (s)Peak memory (GiB)
M2PDE-AR1000862.2615.689.82
GenCP-AR1041.670.7534.13
ChronoSync322989.7854.3510.64
ChronoSync5428.97.7910.64
ERDM-Joint32749.913.639.90
  • 300-frame rollouts. Every rollout does the same work, so wall-clock time scales linearly with the number of committed frames. Peak memory is set by the window size and does not grow with the horizon.
  • Fewer Heun steps. Reducing ChronoSync from 32 to 5 Heun steps makes it about 7 times faster and changes the errors by at most 6.67% (Table 17).
  • Without bootstrap. Removing the rolling bootstrap halves the number of network evaluations. Since wall-clock time is proportional to the evaluation count (Table 14), this roughly halves the runtime. On the harder systems, however, one bootstrap pass contributes to accuracy (Figures 25, 28).
AuthorsOpenReview ↗

Response to Reviewer Wwam (continued)

W2 / Q2. Pre-iteration and PDE-specific knowledge

  • FitzHugh–Nagumo and Burgers use no pre-iteration. Their conditional data come from solving each field against a partner field sampled independently from a Gaussian random field (Appendix E.1–E.2). This is the clean decoupled setting. ChronoSync recovers 96.5–99.7% of the gap on these systems (Table 15).
  • Darcy and Rayleigh–B'enard use one windowed pre-iteration. This follows M2PDE and GenCP pre-iteration procedure so that the conditional data is at least physically admissible. Please note that we use only 1 pre-iteration and thus the conditional training data have substantial errors compared to the coupled/joint truth data (Figure 2).
  • Building conditional data without a sub-problem solver. ChronoSync assumes that conditional trajectories exist, either from single physics solvers (the common situation in multiphysics practice) or from observations of one field given the others. How to construct such data with no solver at all is outside the scope of this work. It relates to the future direction stated in our conclusion on conditional data that are weakly informative or even mutually incompatible.

We hope these additions address the reviewer's concerns, and we would be glad to clarify anything further.

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

This paper studies how to recover long-horizon coupled PDE dynamics when training uses only decoupled conditional trajectories. ChronoSync trains a conditional Elucidated Rolling Diffusion Model for each physical field, sequentially updates cross-field conditioning within each denoising step, and performs long-horizon prediction through overlapping temporal windows. The authors analyze the propagation of composition perturbations using operator splitting and sensitivity recurrences. They evaluate the method on four PDE systems against decoupled models and models trained on joint data, including physical observables and prediction beyond the training time horizon. The results show substantial improvements in recovering coupled dynamics on the more complex systems.

Critical strengths

1.The problem is important, and the composition procedure provides a meaningful extension to long-horizon prediction. M2PDE and GenCP already study recovering coupled systems from decoupled conditional models. This paper further embeds cross-field updates within rolling diffusion and addresses how local composition errors enter subsequent temporal windows. Its contribution lies in systematically studying the interaction between composition and long-horizon prediction, together with an algorithm and an error-analysis framework.

2.The experiments demonstrate substantial error reductions and evaluate coupled physical behavior. In Table 1, the Darcy concentration and pressure-anomaly errors decrease from 0.159/0.141 for GenCP-AR to 0.075/0.036 for ChronoSync. The Rayleigh–Bénard temperature and velocity errors decrease from 0.162/0.663 to 0.074/0.217. The solute center-of-mass, Nusselt-number, and temporal-extrapolation evaluations further support the recovery of meaningful coupled behavior. Models trained on joint data provide useful reference points for interpreting these results.

Critical weaknesses

1.The connection between the theory and the implemented algorithm needs additional justification. Appendix C.3–C.4 drops second-order terms after local linearization and then derives a long-horizon recurrence. Smoothness and finite perturbations alone do not automatically ensure that the omitted terms remain controlled over long rollouts. Furthermore, the main experiments use (\kappa=0.25) and bootstrap refinement, whereas the direct Lie–Trotter correspondence is initially established for (\kappa=1). The authors explicitly acknowledge that the learned conditionals need not be compatible and that their summed flow need not equal the true joint flow. The theory therefore currently provides a conditional perturbation analysis rather than a complete explanation of the deployed algorithm’s long-horizon accuracy.

2.The attribution of the main gains and the representativeness of the test results require stronger evidence. Appendix J provides parameter-sensitivity studies, but does not directly isolate per-denoising-step cross-field synchronization using the same ERDM backbone, rolling procedure, and computational budget. Existing cross-method comparisons change training and sampling mechanisms simultaneously. In addition, Table 6 reports only 24 Darcy and 16 Rayleigh–Bénard test trajectories, while Figure 1 describes a 16-sample stress-test subset selected for poor conditional-trajectory performance. The relationship between this subset and the main evaluation set must be clarified before the generality of the results can be assessed.

Questions for authors

1.Theoretical clarification: How are the nonlinear remainder terms in Appendix Equation 46 controlled, and how does the recurrence extend to the relaxation updates and bootstrap refinement used in practice? If the analysis currently applies only to an idealized variant, please state the corresponding algorithmic scope explicitly.

2.Mechanism validation: Using the same conditional ERDMs and a comparable function-evaluation budget, how much does updating cross-field conditioning at every denoising step improve over exchanging conditioning only after a complete denoising pass? This comparison would directly affect the assessment of the proposed mechanism.

3.Evaluation-protocol clarification: Are the 16 stress-test samples in Figure 1 the entire Rayleigh–Bénard test set used in Table 1, Figures 2–3, and the temporal-extrapolation experiments? If selection was involved, please specify the original sample pool and selection rule, and provide results on the unselected set or paired uncertainty analysis.

2 comments
AuthorsOpenReview ↗

Response to Reviewer fRwi

We thank the reviewer for the careful reading of our paper and for providing valuable feedback. All references below are to the revised PDF, where changes are marked in blue.

W1 / Q1. Remainder terms, relaxation, bootstrap, and algorithmic scope

Remainder terms. Our analysis is leading order, which is standard for splitting methods (Blanes et al., 2024). Two kinds of higher-order terms appear, and both are now accounted for.

  • Terms of higher order in the step size hh. The revised Appendix C.1.2 analyzes the Heun type Lie-Trotter scheme that ChronoSync actually runs. It keeps the local remainder Rn=O(hn3)\mathcal{R}_n=\mathcal{O}(h_n^3) explicitly (equation 30). Local approximation errors are propagated through the newly introduced Heun-type analysis scheme (Equations 2, 3, 34–35). Lemma 1 discusses bounds for the global error from per-rollout defect through ∥SW∥C3h2\|\mathcal{S}_W\|C_3h^2, so they enter the rollout bound (equation 57) as an explicit term rather than being dropped.
  • Terms of higher order in the perturbation δr\delta_r. Dropping these is the standard linearized stability step for non-autonomous discrete-time systems (Potzsche, 2010). Because δr\delta_r is driven by O(h)\mathcal{O}(h) splitting defects, the O(∥δr∥2)\mathcal{O}(\|\delta_r\|^2) terms are of higher order than the terms retained. We also no longer assume that the linearized propagator is stable and rather empirically estimate it. The new Section 5 measures it along ChronoSync rollouts. A perturbation of 2×10−32\times10^{-3} of the state norm, injected along the measured splitting defect, is amplified while the flow develops and then contracts. The fitted contraction rates are ρ^=0.992\hat\rho = 0.992–0.9930.993 per rollout on Darcy, and 0.9970.997–0.9980.998 on Rayleigh–Bénard (RBC) after r≈220r\approx220 (Table 2, Figure 4). This is the mixed regime of Corollary 1.

Relaxation and bootstrap (new Remark 1 in Appendix C.5).

  • The theoretical analysis covers the sequential Lie–Trotter sweep at κ=1\kappa=1 and this is now mentioned in Remark 1.
  • With nbootroll=1n^{\mathrm{roll}}_{\mathrm{boot}}=1, the rollout map Ψr\Psi_r is the composition of two coupled Heun passes. Theorem 2 therefore applies unchanged.
  • Relaxation with κ<1\kappa<1 is now covered in Appendix C.6, marked in blue text. Empirically, on Darcy and Rayleigh-B'enard, κ=1\kappa=1 is within 6.7% of the reported κ=0.25\kappa=0.25 on every field (Figures 24, 27), so the analysed scheme performs like the one we report.

Compatibility. The theory bounds the error due to composing the learned flows relative to their summed flow. Section 3.3, "Scope of the analysis" clarifies this. Model error enters as a separate per-rollout perturbation ζr\zeta_r (Appendix C.6). Its practical effect is measured empirically through the gap recovered relative to the coupled-data oracles. In addition, our conditional training data has substantial errors with respect to the joint/coupled data as seen in Figure 2. Still, ChronoSync is able to recover close to 90% of the conditional to joint oracle gap on average. On Rayleigh–B'enard it is more accurate than the same-backbone ERDM-Joint trained on coupled data, both on the stress-test set (Table 1) and on 20 randomly drawn test trajectories (Table 16).

W2 / Q2. Isolating per-step synchronization at a matched budget

The new Table 18 (Appendix K.5) keeps the conditional ERDMs, rolling procedure, 5 Heun steps and number of network evaluations per committed frame fixed. Only the cross-field exchange frequency changes.

Cross-field exchangeDarcy ccDarcy qqRBC TTRBC (ux,uy)(u_x,u_y)
Every Heun step (ChronoSync)0.0800.0380.0710.229
Once per window0.0870.0480.1170.339
Error change+8.75%+26.3%+64.7%+48.03%

Exchanging once per window raises the error on every field. The increase is largest on RBC, where the coupling is strongest and the window is longest (W=55W=55). Five Heun steps match the 32-step accuracy of Table 1 (new Table 17), so this comparison applies directly to the main results.

AuthorsOpenReview ↗

Response to Reviewer fRwi (continued)

W2 / Q3. RBC test set and selection

The 16 stress-test trajectories are the Rayleigh-B'enard test set used throughout the paper (Table 1, Figures 1–3, Section 4.5). We started from a held-out pool of 50 coupled RBC test trajectories, ranked them by the relative ℓ2\ell_2 error of their conditional trajectories against the coupled reference, and kept the 16 with the largest error. The selection uses only input data and never the output of any method/algorithm, so it raises the difficulty for every method alike. The rule is now stated in Table 7 and Appendix J.5.

To check that the selection does not drive the results, the new Table 16 repeats the evaluation on 20 randomly drawn RBC test trajectories.

MethodTT(ux,uy)(u_x,u_y)Gap recovered (TT / u\mathbf{u})
Conditional trajectories0.1720.6040 / 0%
GenCP-AR0.1580.67411.38 / −13.8%
ChronoSync0.0570.19293.4 / 81.6%
ERDM-Joint (coupled data)0.1080.38652.03 / 43.1%
FNO-AR-Joint (coupled data)0.0490.099100 / 100%

On the random set, ChronoSync recovers 93.4% and 81.6% of the gap, compared with 91.9% and 90.5% on the stress-test set. The ranking is unchanged, and ChronoSync again outperforms the same backbone ERDM-Joint trained on coupled data.

For the Darcy case, the 24 test trajectories are the full held-out Darcy test set, with no selection.

We hope these additions address the reviewer's concerns, and we would be glad to clarify anything further.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to the comments and suggestions

We thank the ICLR 2027 conference and the three reviewers for their careful and constructive feedback. All changes in the revised PDF are marked in blue.

What the reviewers found valuable. The reviewers agreed that recovering coupled dynamics from decoupled conditional data is an important and less-explored problem (fRwi, Wwam). They highlighted:

  • the large, consistent error reductions on the harder Darcy and Rayleigh–B'enard (RBC) systems, which hold at every Darcy window length (fRwi, rgmW)
  • the evaluation across four systems against both decoupled baselines and coupled-data oracles (Wwam)
  • the physical observables (solute transport, Nusselt number) and the time extrapolation beyond the training horizon (fRwi, Wwam, rgmW).

The new additions further strengthened the theory, introduces a new section connecting the theory to experiments, and several additional ablations.

Three main additions

1. Connecting theory to experiments: measured propagation of splitting defects (new Section 5, Figure 4, Table 2, Appendix L). (Wwam, rgmW) Corollary 1's mixed-regime condition is now measured rather than assumed. At injection rollout r0r_0, we perturb the rolling state along the splitting defect that ChronoSync actually produces at r0r_0, estimated by step doubling (SS vs. 4S4S Heun steps). We then track the committed frame response cr0(ℓ)c_{r_0}(\ell) under fixed noise. To the first order, cr0(ℓ)c_{r_0}(\ell) is the loss/gain of the rollout propagator along the defect direction and it is a measured proxy for the sensitivity weights ∏qLq\prod_q L_q in equation 5. Then, using Corollary 1 form, we fit ρ^\hat\rho in a log-linear space using least squares. Example results are shown below.

Fitted ρ^\hat\rho per rolloutr0=5r_0=5r0=54r_0=54r0=144r_0=144
Darcy (W=10W=10)0.9930.9921.009
RBC (W=55W=55), r≤220r \le 2201.0111.0131.020
RBC (W=55W=55), r>220r > 2200.9970.9981.003

Both systems are in the mixed regime. The defect is amplified while the Darcy finger-like plume or the RBC convection develops, and contracted afterwards. For Darcy injections at r0≤54r_0 \leq 54, the response ends at 7–21% of its value at the first window boundary. The responses require no ground truth coupled data, yet they follow the shape of the relative ℓ2\ell_2 error curves (Figure 4).

2. Same-backbone ablations that isolate synchronization (new Appendix K.5, Table 18, Figures 24–28). (fRwi, rgmW) Exchange frequency at a matched budget (Table 18). Both schedules use the same conditional ERDMs, 5 Heun steps, and the same number of network evaluations per committed frame. The only difference is whether cross-field estimates are exchanged at every Heun step (ChronoSync) or once per window.

Cross-field exchangeDarcy ccDarcy qqRBC TTRBC (ux,uy)(u_x,u_y)
Every Heun step (ChronoSync)0.0800.0380.0710.229
Once per window0.0870.0480.1170.339
Change+8.75%+26.3%+64.7%+48.03%

The benefit grows with window length and coupling strength. This attributes ChronoSync's gains to synchronization within diffusion time rather than to the backbone.

κ\kappa and rolling bootstrap on Darcy and Rayleigh-B'enard (Figures 24, 25, 27, 28).

  • Varying κ∈{0.25,0.5,1}\kappa \in \{0.25, 0.5, 1\} does not change the errors much, and that includes κ=1\kappa=1, the theoretically analysed scheme.
  • One rolling bootstrap pass does help: Darcy qq falls from 0.096 to 0.036, Rayleigh-B'enard TT from 0.088 to 0.074, and Rayleigh-B'enard u\mathbf{u} from 0.276 to 0.217. A second pass does not help further.

RBC window sweep for every method (Figure 26). ChronoSync is the best decoupled method and lies below ERDM-Joint.

3. Theory aligned with the algorithm that is run (Sections 3.3–3.4, Modified Appendix B, C.1.2, C.5). (fRwi, rgmW)

  • The splitting theoretical analysis now covers the Heun-type Lie–Trotter scheme that ChronoSync uses (Appendix C.1.2). Local approximation errors are propagated through the newly introduced Heun-type analysis scheme (Equations 2, 3, 34–35). Lemma 1 discusses bounds for the global error from per-rollout defect including the O(h)\mathcal{O}(h) from the Lie bracket plus O(h2)\mathcal{O}(h^2) from Heun stepping.
  • Algorithms 1 and 3 now match the implementation used for all reported results:
    • each field is conditioned on the unrelaxed estimates of the fields updated before it in the sweep
    • the κ\kappa-relaxation is applied once after the sweep
    • at κ=1\kappa=1, this is exactly the sequential Lie–Trotter composition
    • a STALE (Jacobi) mode is also specified.
  • Remark 1 in Appendix C.5 states the scope of the theoretical analysis. This analysis covers κ=1\kappa=1, and with nbootroll=1n^{\mathrm{roll}}_{\mathrm{boot}}=1 the rollout map is a composition of two coupled Heun passes, so the recurrence of Theorem 2 applies unchanged.

We answer each reviewer's remaining questions in the individual responses.

AuthorsOpenReview ↗

General response to the comments and suggestions (continued)

Other additions

  • Inference cost (new Appendix I, Table 14). All runs are on RBC with W=55W=55, 5 test trajectories, 55 committed frames, and one GPU:

    MethodWall-clockPeak memory
    ChronoSync, 5 Heun steps428.9 s10.64 GiB
    ERDM-Joint749.9 s9.90 GiB
    M2PDE-AR862.26 s9.82 GiB
    GenCP-AR41.67 s34.13 GiB

    Reducing ChronoSync from 32 to 5 Heun steps is about 7× faster and changes the errors by at most 6.67% (Table 17).

  • Randomly selected Rayleigh-B'enard test trajectories (new Appendix J.5, Table 16). On 20 random test trajectories, ChronoSync reaches errors of 0.057 (TT) and 0.192 (u\mathbf{u}). It recovers 93.4% and 81.6% of the conditional to joint oracle gap, compared with 91.9% and 90.5% on the stress-test set. The stress-test selection rule is now stated in Table 7.

We answer each reviewer's remaining questions in the individual responses.

For all reviewers: as the discussion period is approaching its end, please join the discussion and put acknowledge on authors' rebuttal, thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.