acceptodds
Under review as a conference paper at ICLR 2027

Learning PDE Dynamics between Submanifolds Using Green's Observation Operators

Abstract

Many physical systems are driven and observed only on lower-dimensional submanifolds of a larger spatial domain, while their dynamics are governed by the ambient medium occupying that domain. Examples include laser-heated parts imaged by an infrared camera, and ground-level emissions measured on a sensor plane. Full-domain solvers, however, compute the entire volume for every new source although only the observation submanifold is needed, and black-box surrogates do not exploit that the ambient medium remains fixed. We introduce the Green's Observation Operator (GObO), which maps the ambient medium once to the Green's kernel of a linear PDE restricted to the source and observation submanifolds. New sources then cost one lower-dimensional integral and no network evaluation. Exponential rates in the kernel yield an exact finite streaming state with horizon-independent memory; we prove its stability and an approximation rate for the restricted heat kernel. On three-dimensional heat conduction and advection–diffusion with collocated and distinct source and observation geometries, GObO trained on static sources predicts responses to moving sources zero-shot with 4–8 lower error than black-box surrogates, at 1.4 ms per query after a single conditioning pass. The same kernel transfers across resolutions and admits corrections for mild nonlinearities, including radiative losses and temperature-dependent conductivity, without retraining, at the cost of lower in-distribution accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, operator learning, Green's functions, partial differential equations, source–observation submanifolds, scientific machine learning

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_ggM3OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Green’s Observation Operator (GObO) for predicting PDE responses between lower-dimensional source and observation submanifolds coupled through an ambient medium. The method learns a medium-conditioned Green kernel represented by low-rank spatial factors and exponential temporal modes. This representation preserves linearity in the source, enables efficient reuse across different sources, and supports streaming prediction through a fixed-dimensional state. Experiments on three-dimensional heat conduction and advection–diffusion demonstrate generalization from stationary training sources to moving test sources and efficient repeated source evaluation, with a trade-off in in-distribution accuracy. Additional experiments investigate resolution transfer and extensions to selected nonlinear effects.

Critical strengths
  • A well-motivated perspective on a practically PDE problem: In many applications, observations are indeed available only on a lower-dimensional submanifold, while need to be predicted under varying source configurations. Focusing on the source to observation mapping directly addresses these practical requirements. This perspective provides a compelling motivation for learning a reusable operator that captures the influence of the ambient dynamics.

  • Generalization from stationary training sources to moving test sources: Models trained exclusively on stationary sources retain good predictive accuracy on moving sources without retraining, with substantial improvements over the evaluated black-box baselines. The results support the method’s potential utility for predicting responses to source configurations beyond those encountered during training.

Critical weaknesses
  • Limited baseline comparisons: I recognize that predicting between submanifolds under varying sources presents challenges that many existing methods are not directly designed to address. Nevertheless, it would be valuable to explore additional comparisons with methods that can be reasonably adapted to this setting. Such comparisons would strengthen the empirical evidence for its advantages.

  • Lack of a complete method architecture diagram: Although Figure 4 illustrates the conditioning network, it does not show the full prediction pipeline. A complete architecture diagram, even if included in the appendix, would help readers understand how the individual components work together.

Questions for authors
  • Can GObO generalize to more spatially localized sources or sources with rapid temporal variation? In particular, does the model trained on the current stationary-source distribution retain its predictive accuracy under these changes without retraining?

  • The experiments focus on heat conduction and advection–diffusion. How applicable is the proposed framework to other propagation kernels? Could the authors discuss its expected performance and whether such settings would require changes to the kernel parametrization?

1 comment

Dear reviewer, we thank you for your thoughtful feedback and positive assessment of our work, especially your recognition of the practical relevance of our setting.

To summarize, we addressed your concerns through:

  • Additional DeepONet baseline and discussion of related DMD/Koopman methods.
  • A complete diagram illustrating the full GObO pipeline.
  • New experiments with spatially localized and rapidly varying sources, evaluating zero-shot generalization.
  • Discussion of applicability to other propagation kernels and the corresponding parametrization requirements.

Details are provided below. Changes to the revised paper are marked in blue. Further additions include expanded theory and ablations, mixed-source training, and curved-manifold experiments.

We believe these additions address your concerns and further strengthen the evidence for our method. We hope they will reinforce your positive assessment and support a higher rating. If any relevant concerns remain, we would greatly appreciate the opportunity to address them.

W1: Baseline Comparisons

The original paper already considered several complementary baselines: FNO-2D, FNO-3D, a backbone-matched spectral predictor, and NGO, together with a one-step variant using GObO's own backbone (Appendices E.1.1, E.3, F.2.1, and F.3.2). We now additionally include DeepONet, extending the comparison to another major neural operator architecture (Appendix F.2.1, Table 17). It also performs poorly on unseen moving sources.

Furthermore, we recognize the connection to DMD/Koopman methods and have added a discussion in Appendix A.1. In particular, our best linear one-step map experiment is closely related to DMDc and delay-embedded variants, as now discussed in Appendix F.3.2.

Together, these comparisons cover several neural operator architectures, a direct Green's operator competitor, and connections to classical data-driven dynamical models.

W2: Architecture Diagram

We appreciate this suggestion and agree that the complete prediction pipeline should be more clearly illustrated. We therefore added a full GObO pipeline diagram (Figure 6), showing how the medium is encoded, the kernel is constructed, and new sources are evaluated through batch or streaming prediction. Together with the existing detailed diagram of the conditioning network (Figure 5) and minor text clarifications, this now illustrates both the overall method and its individual components more clearly, improving readability (Appendix D.2–D.3).

Q1: Generalization to Other Sources

To directly investigate this question, we added zero-shot experiments with narrower spatial pulses, short temporal bursts, and pulse trains (Appendix F.2.1, Table 17). All models are evaluated without retraining on these new source distributions.

Narrow pulses cause severe failures for all models, demonstrating a limitation when sources are substantially more localized than in training. However, GObO retains low errors on short bursts (0.015) and pulse trains (0.010), compared with 0.261 and 0.084 for FNO-2D, respectively. Thus, GObO substantially outperforms the baselines on both types of rapid temporal variation.

Furthermore, we now show that GObO's generalization advantage persists even when training includes moving sources (Appendix F.2.5, Table 22). Together, these experiments strengthen the evidence that GObO generalizes beyond its training source distribution, while identifying extreme spatial localization as a remaining limitation for all models.

Q2: Other Propagation Kernels

GObO is not inherently restricted to heat conduction. The main requirements are linear, time-invariant dynamics admitting a suitable Green's kernel expansion with stable temporal rates. This includes other dissipative PDEs and potentially oscillatory systems such as damped wave equations.

Importantly, we already demonstrate such an adaptation for advection–diffusion, where complex conjugate rates accommodate oscillatory components while retaining the same kernel parametrization and streaming construction (Appendix D.4). This illustrates how different propagation mechanisms can be handled within the framework.

The expected accuracy, however, depends on the kernel spectrum. Diffusion strongly suppresses high-frequency modes, whereas wave propagation generally exhibits weaker spectral decay and may require more modes for comparable accuracy. Vector-valued systems may additionally require adapted spatial factors, while time-dependent media would require more substantial changes. Thus, the main expected limitation is approximation efficiency rather than applicability of the framework, provided the required spectral structure holds.

Further, added an experiment on nontrivial curved source and observation manifolds (Appendix F.2.3, Table 20). This provides further evidence that the framework applies beyond the original geometric settings.

Reviewer_1HABOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a Green's Observation Operator (GObO) for predicting the dynamical evolution from a source manifold to an observation manifold in partial differential equations (PDEs). The model first learns a low-rank, time-dependent representation of the Green's kernel from the medium configuration, then projects a new source term onto the source manifold and computes its response on the observation manifold through the kernel function. In this way, the medium configuration needs to be encoded only once, after which the model can efficiently process source terms with different locations, trajectories, and temporal profiles in the same medium. The core of the model can be summarized by the following low-rank representation of the Green's kernel Kα(ξ,η,τ)≈∑r=1Re−μrτAr(α)Br(α)K_{\alpha}(\xi, \eta, \tau) \approx \sum_{r=1}^{R} e^{-\mu_r \tau} A_r(\alpha) B_r(\alpha)

Here, the medium configuration α determines the coefficients of the kernel, while A_r and B_r can be interpreted as low-dimensional factors defined on the source and observation manifolds, respectively. The authors conduct experiments on heat-conduction and advection-diffusion problems. They train the model using static sources and evaluate it on previously unseen moving sources, including linearly moving and serpentine trajectories. The experiments also compare GObO with FNO, a one-step autoregressive model, and different observation-manifold configurations. The results indicate that GObO is relatively stable for moving-source and long-horizon predictions, while the inference cost for each new source is low.

Critical strengths
  1. GObO has a clear physical motivation in its architectural design. The model treats the medium configuration and the source term separately and exploits the linear structure of the source response, allowing a model trained for one medium to be reused for multiple new sources. This design is more interpretable than directly learning a black-box source-to-observation mapping and addresses the repeated-computation problem targeted by the paper.

  2. The experiments address the practical issue that low-dimensional observations do not provide a closed state representation. The authors evaluate moving sources, long-horizon prediction, and different observation surfaces, showing that the model is not merely fitted to a single fixed trajectory. The heat-conduction and advection-diffusion experiments also provide reasonably broad evidence for the applicability of the method.

Critical weaknesses
  1. The FNO comparisons are not yet sufficient to support the paper's main claims about the generalization of black-box models. Both GObO and FNO are trained primarily on static sources, while linearly moving and serpentine sources are used as unseen moving-source test cases. Thus, the experiments show that GObO performs well in the specific setting of "static-source training with zero-shot evaluation on moving sources," but they do not directly establish that GObO outperforms FNO under general training conditions. In particular, for the heat-conduction experiment, FNO has lower error on static sources, whereas GObO's advantage appears mainly on the moving-source tests. Moreover, FNO-2D uses a two-dimensional encoding of the medium, while GObO receives a normalized three-dimensional medium as its conditioning input. The two methods therefore do not receive exactly the same environmental information. The current results cannot clearly disentangle whether the performance difference arises from the model architecture or from differences in the input information and task setup. Comparisons involving mixed static-and-dynamic-source training, identical medium inputs, and comparable training budgets are the most important additions needed in the experiments.

  2. There remains a substantial gap between the theoretical analysis and the model used in practice. Theorem 1 concerns the spectral expansion of the exact Green's operator for the heat equation and requires the time lag to satisfy τ ≥ τ_0 > 0 However, the practical model operates on heterogeneous three-dimensional media and uses a two-dimensional cosine basis, fixed analytic temporal decay rates, a finite number of temporal modes, a low-rank factorization, and coefficients predicted by a neural network. The paper does not provide an error bound for the combined effect of these approximations, nor does it sufficiently characterize how the number of spatial basis functions, the number of temporal modes, and the rank affect the results. Therefore, the current theory explains the motivation for the method, but it does not yet fully account for the error or stability of the practical model.

Questions for authors
  1. Could you add a fairer comparison in which GObO and FNO use the same three-dimensional medium input, comparable parameter counts, and similar training budgets, and are evaluated under (i) static-source training, (ii) mixed static-and-dynamic-source training, and (iii) previously unseen moving trajectories? This would help determine whether GObO's advantage primarily comes from its source-term decomposition or from the current zero-shot evaluation setup.

  2. For the two-dimensional cosine basis and the fixed temporal decay rates used in practice, could you report their discrepancy relative to the spectral decay rates of the true operator, together with error curves as the number of spatial basis functions, the number of temporal modes, and the low-rank order are varied? Please also clarify how the model handles the short-time singular behavior of the Green's kernel when the source and observation manifolds coincide or are very close to each other.

1 comment

Dear reviewer, we thank you for your thoughtful feedback and positive assessment of our work.

To summarize, we addressed your concerns through:

  • Clarification that both models receive the same 3D medium information.
  • Mixed-source training and additional source-transfer tests, confirming GObO's generalization advantage.
  • Corollary 1 controlling the short-time singularity, even for coincident manifolds.
  • Proposition 3 separating parametrization errors, with rate, basis, and rank ablations.

These changes appear in the revised paper, marked in blue. Further additions include a DeepONet baseline, curved-manifold experiments, a full pipeline diagram, and minor clarifications to improve readability.

We believe these additions address your concerns and strengthen our evidence. We hope they will reinforce your positive assessment and support a higher rating. If any relevant concerns remain, we would greatly appreciate the opportunity to address them.

W1/Q1: Fair Baselines and Moving-Source Training

We clarify a point that may have been ambiguous: FNO-2D uses exactly the same encoder of the full 3D medium as GObO, while FNO-3D operates directly on the volume using 3D convolutions. We have made this explicit in the revision. All models share splits, loss, and 30,000 training steps (Appendix D). Even a higher-capacity FNO-2D (21.11M parameters, close to GObO's 22.16M) fails to transfer (Appendices F.1.1, F.2.2). Thus, medium information and model size do not explain the gap alone.

To test whether static-only training caused FNO's weak transfer, we now train both models on static pulses and straight scans, without changing architectures or training steps (Appendix F.2.5, Table 22). FNO-2D improves on straight scans, but GObO still achieves 27× and 7× lower errors on held-out serpentine scans and two-pulse sums:

Mixed trainingStaticStraightSerpentineTwo-pulse
GObO0.0070.0070.0030.006
FNO-2D0.0010.0050.0820.042

We also added zero-shot tests with narrower pulses, short bursts, and pulse trains (Appendix F.2.1). GObO achieves errors of 0.015 and 0.010 on bursts and trains, versus 0.261 and 0.084 for FNO-2D. Narrow pulses remain challenging for all models when training lacks localized sources.

Together, these results show that GObO's transfer advantage extends beyond the original setting and persists after moving-source training, supporting separation of sources from the medium-dependent kernel. Moreover, on curved manifolds GObO closes the in-distribution gap, slightly outperforming FNO-2D on static sources (0.041 vs. 0.046), while retaining its moving-source advantage (0.040 vs. 0.196; Appendix F.2.3, Table 20).

W2/Q2: Practical Theory and Approximation

Singularity

We added Corollary 1: for bounded sources, lags below τ0\tau_0 contribute only O(τ0)O(\sqrt{\tau_0}) to the observed response, even for coincident manifolds, with a stronger bound when separated. Thus, the short-time singularity has a controllably small effect on predictions.

Importantly, the singularity concerns the exact Green's kernel, whereas GObO's finite exponential parametrization is bounded even at zero lag. The discrete prediction therefore never requires evaluating the singular exact kernel. Consistently, halving the test-time step changes error only from 0.011 to 0.010 (Appendix F.4.2). Hence, the singularity is theoretically controllable and does not present a practical numerical problem.

Parametrization Error

The original Theorem 1 bounded spectral truncation. We now add Proposition 3 (Appendix B.3), separating errors from discarded modes, fixed decay rates, and finite spatial bases, and showing how the parametrization controls them.

Figure 13 (Appendix F.4.6) provides network-free error curves varying distinct decay rates, cosine basis size, and factor rank, isolating representation error. At lag τ=0.05\tau=0.05, increasing the basis from 424^2 to 32232^2 reduces kernel error by 56×. Additional rates and rank also help before saturation; rank 8 matches rank 16. This shows high approximation accuracy at modest sizes.

Complementary end-to-end retraining of the full kernel model (Tables 28, 27, and 19) yields:

ChoiceStatic relative error
Rates M=4/8/16/32M=4/8/16/320.029 / 0.013 / 0.011 / 0.011
Basis 82/162/3228^2/16^2/32^20.019 / 0.009 / 0.011
Factor rank L=1/16L=1/160.130 / 0.011

Fixed and learned rates yield the same static error (0.011, Table 19). Because true heterogeneous-medium rates are not available in closed form, we did not measure their discrepancy directly, but Proposition 3 bounds the fixed-rate approximation error.

Together, theory and experiments show high accuracy with modest representation sizes. Further complexity brings diminishing returns, supporting the practical parametrization.

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

This paper proposed a method to learn linear responses through a field that is partially observable at the surfaces. They proposed to use a neural network on the 3D material field, and output a set of kernels with exponential time decay to model the Green's function. Trained only on static pulses, the model handles moving sources with roughly 1% error on 3D heat conduction, where surface-to-surface baselines fail to learn the temporal relation.

Critical strengths
  1. The problem is well-motivated and well posed. The authors identified a useful scenario where partial observation is limited to surfaces on a 3D media.
  2. The method is evaluated honestly and throughly on test cases. The authors report the improvement of the method and also acknowledge its in-distribution accuracy is worse than standard ML baselines.
Critical weaknesses
  1. The baselines are only evaluated on a constrained setting. If GoBO is able to see the media, FNO could also use that information (either by directly taking it as input, or use a solver to generate fully-observable data). It seems to me that FNO's failure to handling moving sources could be it has never seen a moving target.
  2. The framing is more general than the experiments. The authors speak of manifolds and write around general notations, but the bases and experiments are implemented on rectangular domains. I'd ask the authors to either tone the framing down to shapes and surfaces, or demonstrate the method on nontrivial manifold geometries.
Questions for authors
  1. Is this paper related to DMD and Koopman operator methods? These are also common methods in learning time-dependent but this paper didn't discuss their relationship.
  2. How much error in Gobo comes from representation error (spectral truncation) and how much from the network capacity / optimization? Can we expect gobo's accuracy to exponentially improve as we increase basis?
1 comment

Dear reviewer, we thank you for your feedback and for the positive overall assessment of our paper, especially regarding its motivation and presentation.

To summarize, we addressed your concerns through:

  • Clarification that all baselines receive the medium and are compared fairly.
  • Mixed-source training, retaining GObO's advantage.
  • Experiments on nontrivial curved manifolds, where GObO outperforms FNO even in-distribution.
  • Discussion of DMD/Koopman and its connection to our experiments.
  • Proposition 3 on representation errors, with further ablations.

Details are provided below and incorporated into our revised paper, with changes marked in blue. Further additions include expanded theory and ablations, a DeepONet baseline, a full pipeline diagram, and minor clarifications to improve readability.

We believe the additional results and clarifications directly address the main concerns underlying your initial assessment. We hope this will allow you to reconsider your rating accordingly. If any relevant concerns remain, we would greatly appreciate the opportunity to address them.

W1: Baselines

All baselines receive the medium. In fact, FNO-2D uses the same medium encoder as GObO, while FNO-3D directly uses 3D convolutions. Furthermore, we provide a one-step variant using GObO's backbone instead of NGO's original direct basis projection of the medium. We have emphasized this in the paper.

To test whether FNO's limitation comes from never seeing moving sources, we now train on mixed static and moving sources (Appendix F.2.5, Table 22). Even then, GObO achieves 27× and 7× lower errors than FNO-2D on held-out serpentine scans and two-pulse sums, respectively. Thus, moving-source training alone does not remove the performance gap, supporting the structural advantage of GObO.

W2: Nontrivial Manifolds

The original manifold configurations were inspired by practical applications, but our method is not limited to them. The submission already included a controlled experiment on nontrivial curved manifolds, which we have now extended to a full experiment with trained models and heterogeneous medium data (Appendix F.2.3, Table 20). Crucially, GObO slightly outperforms FNO-2D on these manifolds even for in-distribution static sources, while retaining its strong advantage on unseen moving sources:

Curved surfacesStaticMoving
GObO0.0410.040
FNO-2D0.0460.196

Still, we would like to emphasize that our mathematical framework is more general and applies to nontrivial manifolds beyond those tested here.

Q1: DMD

We thank the reviewer for pointing out this relation. We agree that DMD/Koopman methods belong in the related work and have added a paragraph in Appendix A.1.

Their main limitation in standard form is the assumption of closed-state evolution, which fails for dynamics restricted to observation manifolds. However, DMD approaches can incorporate history (Arbabi & Mezić, 2017; Peitz et al., 2025). In fact, our best linear one-step map experiment (Appendices E.3.3 and F.3.2) is closely related to DMDc (Proctor et al., 2016) with one frame, and to delay-embedded methods with longer histories. History reduces the fit residual by approximately 3× from 1 to 16 frames, whereas GObO evolves a source-driven internal state without needing past output observations. We now also discuss this connection alongside the experiment (Appendix F.3.2).

Q2: Error Disentanglement

First, we added Proposition 3, separating three sources of representation error: spectral truncation, fixed decay rates, and finite spatial bases. It shows how the parametrization controls these errors.

Furthermore, we added a network-free ablation varying the distinct decay rates, cosine basis size, and factor rank (Appendix F.4.6, Figure 13), studying representation error independently of network learning. At lag 0.05, increasing the basis from 424^2 to 32232^2 reduces kernel error by approximately 56×. More rates and higher rank also improve approximation before gains saturate. Thus, the parametrization achieves high approximation accuracy with modest representation sizes.

Complementary full-model ablations with retraining (Tables 28, 27, and 19) show:

ChoiceRel. error
Rates M=4/8/16/32M=4/8/16/320.029 / 0.013 / 0.011 / 0.011
Basis 82/162/3228^2/16^2/32^20.019 / 0.009 / 0.011
Factor rank L=1/16L=1/160.130 / 0.011

Together, these results show that representation capacity matters, while final accuracy also depends on training data, identifiability, network capacity, and optimization.

Regarding exponential convergence, Theorem 1 establishes this for the true heat eigenspectrum, not the learned cosine basis. Nevertheless, we observe rapid initial improvement followed by saturation, indicating that high accuracy does not require excessively large representations.

Comments on the paper

1 comment

Dear Area Chair and reviewers,

we thank you for your thoughtful reviews and positive initial assessment of our work. In response to your feedback, we have strengthened the paper in three main areas:

  • Experiments: We clarified that baselines are compared fairly, added DeepONet, and discussed the connection to DMD/Koopman methods. New experiments on curved manifolds, mixed-source training, and further ablations strengthen our empirical evidence. Across settings and training configurations, GObO's generalization advantage persists.
  • Theory: We clarified that the short-time singularity is both theoretically controllable and not a concern in practical implementation (Corollary 1). We also provided a decomposition of representation error (Proposition 3).
  • Clarity and presentation: We added some clarifications in the main text and provided a complete pipeline overview figure to make the paper easier to understand.

We believe these additions address the main concerns raised in the reviews and further strengthen the paper. We would be happy to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.