acceptodds
Under review as a conference paper at ICLR 2027

Learning Continuous-time Dynamics Guidance for Diffusion-based Spatiotemporal Reconstruction

Abstract

We address the ill-posed inverse problem of reconstructing complete spatiotemporal fields from sparse measurements in systems governed by unknown PDEs. Diffusion models have emerged as powerful generative models for such inverse problems. However, conditioning on sparse measurements alone can produce reconstructions that match the observations but remain inconsistent with the underlying dynamics, limiting their reliability in unobserved regions of the domain. We propose DiffSINDy, a guided diffusion posterior sampling framework in which a diffusion model provides a prior over full spatiotemporal trajectories, while a learned surrogate provides dynamics guidance without requiring the ground-truth governing equations. We instantiate the surrogate with Sparse Identification of Nonlinear Dynamics (SINDy) and couple it with a differentiable method-of-lines solver, yielding an interpretable continuous-time guidance model that can be integrated between consecutive snapshots even when the data are temporally sparse. We evaluate the method on ODE and PDE benchmarks against guidance based on neural surrogates, two oracle variants with true PDEs, and observation-only DPS. Across benchmarks, SINDy-guided diffusion achieves competitive reconstruction error and improves physical consistency over learned neural surrogates and observation-only guidance.

Primary area:
learning on time series and dynamical systems
Keywords:
Diffusion Models, Inverse Problems, Spatiotemporal Reconstruction, PDE, SINDy

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_P69VOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This is paper adds diffusionPDE with an additional dynamics guidance.

Critical strengths

The paper is easy to follow. The method is intuitive. The targeting problem is important for diffusion model.

Critical weaknesses

See questions.

Questions for authors
  1. If we are already able to train a dynamics model, why not just use this model to do generation/inpainting. Why using diffusion model?
  2. The method is slower than baselines methods because additional per-step gradient calculation. It also needs to train an additional model.
  3. The novalty is limited. It adds an extra guidance term in DiffusionPDE.
1 comment
AuthorsOpenReview ↗

Response to Reviewer P69V

Thank you for raising the roles of the diffusion prior, computational overhead, and the relationship to DiffusionPDE.

1. Why use diffusion if a dynamics model is available?

A dynamics model predicts evolution from an initial state. Recovering that state from sparse observations across space and time is still an inverse problem that may be ambiguous, ill-conditioned, or difficult to optimize. The diffusion model supplies a learned prior over complete trajectories for this reconstruction problem.

We added a dynamics-only baseline with access to the true equations (Appendix D.3). A GP conditioned on the available context initializes the initial field, which is optimized through differentiable rollout to match the observations, with a spatial smoothness penalty.

In Table S10, the errors of this baseline are approximately 1.6 and 2.3 times those of DiffSINDy on Burgers with 2.5% and 5% context, respectively, and 3.0 times on KS with 5% context. However, the baseline fits the observed context points less closely on KS than on Burgers' equation. The results support the value of a learned trajectory prior relative to this dynamics-only baseline.

2. Computational overhead

We agree that DiffSINDy requires additional surrogate training and inference-time integration and differentiation. It is slower than observation-only DPS and the cheaper discrete-time guidance variants.

The diffusion model is trained once per dataset and reused across guidance methods; neural-surrogate methods also require an additional trained dynamics model, whereas DPS and the oracle variants do not. The method is most appropriate when improved physical consistency justifies this cost.

3. Relationship to DiffusionPDE and novelty

DiffusionPDE is an important related work, and we build on the established idea of guiding diffusion sampling with observations and dynamics constraints. However, we would like to point out that the distinction concerns both the information required to construct the dynamics guidance and how its consistency is evaluated.

DiffusionPDE constructs its physics guidance from a supplied PDE form, even if quantities like coefficient fields or initial conditions may remain unknown. In practice, DiffusionPDE typically assumes that the underlying PDE equations are known. In our setting, the governing operator is not supplied. Instead, we identify a (shared) PDE from training trajectories using SINDy within a predefined candidate library. The latter provides structural assumptions, but the active terms and their coefficients are learned from data.

Our contribution is to learn continuous-time dynamics using SINDy and guide reconstruction through a differentiable MoL solver. This enforces consistency between reconstructed snapshots without requiring temporal derivative estimates from sparsely sampled data.

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

This paper proposes DiffSINDy, a diffusion-based framework for sparse spatiotemporal reconstruction under unknown governing dynamics. The method combines SINDy with differentiable continuous-time integration to provide dynamics-consistency guidance during diffusion posterior sampling. Experiments on ODE and PDE benchmarks demonstrate improved physical consistency, while gains in reconstruction accuracy remain relatively modest.

Critical strengths

The paper addresses a well-motivated problem, and the experimental evaluation is reasonably comprehensive. The results demonstrate the potential of learned dynamics guidance for improving the physical consistency of reconstructed fields.

Critical weaknesses
  1. The methodological novelty is limited. The proposed framework primarily combines established techniques, including SINDy, differentiable ODE solvers, and diffusion posterior sampling, without introducing a substantial new algorithmic mechanism. Moreover, the comparison with only discrete-time neural surrogates does not sufficiently establish the distinctive contribution of SINDy-based continuous-time guidance.

  2. The practical advantages and generalizability of the method remain insufficiently demonstrated. The experiments rely on clean, fully observed training trajectories and suitably designed SINDy libraries, without adequately examining robustness to dynamics identification errors or model mismatch. Furthermore, the improvements in reconstruction accuracy are relatively modest despite the increased computational cost.

Questions for authors
  1. What is the fundamental methodological contribution of this work beyond combining established techniques such as SINDy, differentiable ODE solvers, and diffusion posterior sampling? How does the proposed approach differ substantially from existing surrogate-guided diffusion methods?

  2. The experiments assume clean, fully observed training trajectories and appropriately constructed SINDy libraries. How robust is the proposed method when the identified dynamics are inaccurate, the training data are noisy, or the candidate library does not adequately represent the true governing equations?

  3. The proposed method achieves substantially greater improvements in physical consistency than in reconstruction accuracy. Given that spatiotemporal reconstruction is the primary task, what is the practical significance of improving physical consistency without a corresponding reduction in reconstruction error? Could the authors demonstrate whether these improvements lead to meaningful benefits in downstream applications, such as temporal interpolation or forecasting, particularly considering the additional computational cost?

1 comment
AuthorsOpenReview ↗

Response to Reviewer YWxx

Thank you for the detailed questions. We address the contribution, the continuous-time neural comparison, robustness, and the practical meaning of the reported gains below.

1. Contribution beyond established components

We agree that SINDy, differentiable integration, and DPS are established methods. Our contribution is a reconstruction framework in which interpretable dynamics identified from training trajectories provide an integration-based consistency constraint on complete generated spatiotemporal trajectories.

Unlike equation-supplied physics guidance, our governing operator is learned within a candidate library. Rather than estimating temporal derivatives from widely separated snapshots, we integrate the learned vector field between reconstructed snapshots. SINDy provides an interpretable instantiation of this mechanism.

Oracle-C and Oracle-D compare integration-based and differential-residual guidance when both receive the true equation. The temporal-resolution experiment further examines their behavior as the interval between snapshots grows. The comparisons support integration-based guidance for the reported inter-snapshot consistency criterion.

2. Continuous-time neural comparison

We added lNODE, a shared neural network that predicts temporal derivatives from a local spatial stencil and is numerically integrated between snapshots (Figures 3–4; Appendix C.2 and Table S6). It uses trajectory-prediction training and the same observation-plus-dynamics guidance framework. This supplies the continuous-time neural alternative missing from the original comparison.

The results indicate that the continuous-time neural surrogate generally does not match the performance of SINDy on our datasets (Figures 3 and 4). This suggests that the performance gains are a strength of the SINDy method with an appropriately chosen library. However, the 16 and 12 resolution reaction-diffusion lNODE experiments use smaller BO budgets because of lNODE's cost (Appendix C.4). For the same reason, a different ODE solver from SINDy was used across all datasets.

3. Complete training data, noise, and identification errors

All models are trained on clean, complete trajectories. We have not evaluated noisy or incomplete training data. Noise-aware identification, including weak-form approaches, is a possible extension that would need evaluation together with the diffusion prior.

The new Burgers experiment instead tests noisy conditioning observations: Gaussian noise with standard deviations 0.001, 0.0025, and 0.005 is added to the context while training data remain clean (Figure 4, column 6). DiffSINDy achieves the lowest reported mean reconstruction error among the tested non-oracle methods at these noise levels and retains low physical consistency loss. This supports robustness to the tested observation noise, not to noisy dynamics identification.

For model mismatch, we now report a missing-term experiment on the reaction-diffusion dataset. Figure 6 shows that this substantially increases surrogate prediction error and reduces the benefits of SINDy guidance. Nevertheless, reconstruction error remains comparable to observation-only DPS, while physical consistency is improved. This illustrates the importance of an adequately specified candidate library and motivates validating the learned dynamics before using them for guidance.

4. Physical consistency and computational cost

Two fields can have similar pointwise errors while differing in whether their successive snapshots are compatible with the governing evolution. Our results show competitive reconstruction error, with modest differences between guidance methods; the larger improvement is in inter-snapshot physical consistency (LPDE\mathcal{L}_{\mathrm{PDE}}).

Pointwise field accuracy is not sufficient for derivative-sensitive tasks. As an example of a derivative-sensitive method, SINDy/PDE identification uses estimated time derivatives and spatial-derivative features, so small oscillatory reconstruction errors can distort the inferred dynamics. This motivates assessing physical consistency alongside field RRMSE. Our results demonstrate improved consistency under the ground-truth evolution operator.

Continuous-time guidance incurs additional simulation and differentiation cost. Our supported claim is therefore a cost–consistency trade-off.

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

DiffSINDy recovers complete spatiotemporal fields from sparse measurements. It uses a diffusion model as a prior over entire trajectories and steers sampling with two gradient signals: one pulls the sample toward the observed data, and the other enforces consistency with the dynamics predicted by a learned surrogate. The surrogate is a SINDy model, a sparse combination of polynomial and derivative terms, trained in the style of a neural ODE.

Across four benchmarks, the method leaves reconstruction error largely unchanged but substantially lowers PDE residuals relative to standard DPS and to guidance from a neural surrogate, coming close to an oracle that has access to the true governing equation.

Critical strengths

Solving an important problem where the observations are sparse. The formulation of unknown equations are also practically useful and the use of a surrogate model is novel and interesting.

Critical weaknesses

The main limitation is a rather contrived experimental setup. The dynamics are “unknown” only within a hand-designed library, and no experiment tests a library missing a required term. The neural baselines receive no comparable structural prior. Moreover, all trajectories within a dataset share identical PDE coefficients, so a single SINDy model suffices. It is unclear how the approach extends to the more realistic case where coefficients vary across instances (e.g., unknown viscosity per trajectory), which would require inferring the dynamics from the sparse test observations themselves, perhaps with an amortized model that predicts coefficients.

Another big limitation is that the physical-consistency metric uses the same method-of-lines discretization and integrator that generate the data and that DiffSINDy sues for guidance. Since SINDy recovers almost exact coefficients, it is effectively optimizing the evaluation metric. An evaluation with an independent discretization, or data generated at higher resolution would prove their point.

Questions for authors

In what specific use cases do you think your setup is useful for?

1 comment
AuthorsOpenReview ↗

Response to Reviewer PeWc

Thank you for identifying the structural assumptions and potential coupling between guidance and evaluation. We added targeted experiments and clarify the intended setting below.

1. Intended use cases and shared dynamics

Our target setting is repeated reconstruction of trajectories from the same system, where a historical archive of complete trajectories is available but only sparse measurements are observed for a new trajectory. The governing operator is not supplied to the reconstruction method; its active terms and coefficients are learned within a specified candidate library. Boundary conditions are assumed known.

Potential use cases include reconstructing sparsely measured runs of a fixed experimental system using a densely recorded archive, or learning from archived outputs of a legacy simulator whose implementation is inaccessible. These motivate the setting; the present evaluation uses synthetic benchmarks.

We agree that trajectory-dependent coefficients define a broader problem. We will state this assumption explicitly as a limitation and identify trajectory-dependent coefficients as a direction for future work that could broaden the method’s applicability.

2. Coupling between numerical guidance and evaluation

We agree that using the same spatial discretization for data generation, guidance, and evaluation can favor a method that accurately identifies the corresponding semi-discrete dynamics.

To test this, we added two periodic Burgers datasets with matched periodicized initial conditions: one generated using finite-difference MoL and one using a dealiased pseudo-spectral spatial discretization (Figure 7; Appendix B). Models are trained separately on each dataset. On the pseudo-spectral data, DiffSINDy retains finite-difference guidance, whereas the reference consistency metric and oracle methods use the pseudo-spectral dynamics.

Despite increased SINDy surrogate error under this mismatch, Figure 7 shows that DiffSINDy retains improved reference evolution consistency relative to observation-only DPS, without a comparable deterioration in reconstruction error. This addresses the concern that the benefit depends entirely on sharing the guidance discretization.

3. Structural prior and library misspecification

The SINDy library is a substantive structural prior. To improve the comparison, we added a local neural vector-field model that also uses continuous-time integration but does not explicitly encode polynomial derivative terms (Figures 3–4; Appendix C.2). This makes the comparison more informative about the benefit of the SINDy representation.

We also include a deliberately misspecified library for the reaction-diffusion dataset by removing the cubic terms. Figure 6 shows that this substantially increases surrogate prediction error and reduces the benefits of SINDy guidance. Nevertheless, reconstruction accuracy remains comparable to observation-only DPS, while physical consistency (LPDE\mathcal{L}_{\mathrm{PDE}}) improves, suggesting that the diffusion prior continues to provide useful regularization even with inaccurate learned dynamics. These results highlight the importance of selecting a sufficiently expressive library and validating the learned surrogate before applying dynamics guidance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers for their constructive feedback. The revision adds experiments addressing the numerical evaluation, the role of continuous-time modeling and structural priors, and the value of the diffusion prior:

  1. Spatial-discretization mismatch: periodic Burgers data generated and evaluated with a dealiased pseudo-spectral spatial operator, while DiffSINDy retains finite-difference guidance (Section 5.2, Figure 7).
  2. A continuous-time neural baseline: a local neural vector field integrated between snapshots (Figures 3–4; Appendix C.2).
  3. Library misspecification: reaction-diffusion reconstruction with cubic terms removed from the candidate library (Figure 6).
  4. Noisy context observations: Burgers reconstruction with three levels of test-observation noise, retaining clean training data (Figure 4, column 6).
  5. Dynamics-only reconstruction: GP-initialized, regularized optimization through the true dynamics, without a diffusion prior (Appendix D.3).

Our contribution is learned inter-snapshot evolution-consistency guidance: an operator identified from training trajectories is integrated between reconstructed snapshots, and the resulting mismatch guides diffusion sampling. The comparisons examine the guidance mechanism and the value of the SINDy structural prior. The spectral Burgers experiment shows that the benefit is not confined to the same spatial discretization used by DiffSINDy, while the dynamics-only comparison supports the value of the learned trajectory prior relative to the evaluated optimization baseline.

We also clarify the limitations. The method assumes clean, complete training trajectories and shared dynamics. Missing terms can weaken the benefit of the learned model, and continuous-time guidance increases computational cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.