acceptodds
Under review as a conference paper at ICLR 2027

OCL-PDE: A Generative Framework for PDE Inverse Problems with Observation-Complementary Latents

Abstract

Partial differential equation (PDE) inverse problems are often ill-posed, making fine-scale details difficult to recover. We address this problem by introducing a learned observation-complementary latent representation that preserves reconstruction-relevant information and is combined with the observation to reconstruct the unknown field. Building on this representation, we propose OCL-PDE, a generative framework that encourages the observation to guide large-scale structure and the latent to supply complementary fine-scale details. OCL-PDE is built on a physics-aware autoencoder (AE) and conditional Flow Matching, supporting inverse reconstruction as well as forward PDE prediction. Experiments demonstrate improved reconstruction accuracy and fine-detail recovery compared with the evaluated baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE inverse problems, scientific machine learning, latent representation learning, conditional generative models, flow matching

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_9fXHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces an approach for solving ill-posed partial differential equation inverse problems. The authors propose to learn an observation-complementary latent representation that preserves reconstruction relevant information for inverse reconstruction. The setup consists of a physics-aware autoencoder and a conditional Flow Matching part. The physics-aware autoencoder is trained to learn a forward response predictor. The decoder part combines the observations with the latent representation. After training of the autoencoder, conditional Flow Matching learns to generate latent codes from Gaussian noise conditioned on the observations. Once trained, the autoencoder and Flow Matching can be combined for fine-scale reconstruction.

Critical strengths

The paper proposes a combination of autoencoders and flow matching for solving ill-posed PDE inverse problems.

Critical weaknesses
  • It is not quite clear to me why the authors did not use a conditional variational autoencoder rather than the proposed autoencoder. Forcing the encoder to predict both an observation yy and a latent representation zz seems unnecessary if the observation can be passed as a condition to both the encoder and decoder like in the standard conditional variational autoencoder architecture.
  • In the related work section, the authors state quite a few related methods for solving PDE inverse problems. In the experimental section, however, their approach is only compared to two previous works for the Darcy flow problem with a continuous coefficient and coarse observations.
Questions for authors
  • Please can you specify how the known query in Figure 1 is incorporated into the autoencoder? The textual description does not match up with the Figure here.
  • The loss function in Equation 5 is only partly defined. Please can you specify how Lreg\mathcal{L}_{\text{reg}} is defined similar to how you defined the other three loss terms?
  • In your experiments you empirically show the importance of the latent. Can you show why learning both the observation and latent is required rather than just training a standard conditional variational autoencoder where the encoder maps the unknown field x observation to a latent space?
3 comments

We thank the reviewer for reviewing our manuscript and for providing insightful and constructive comments.

W1 / Q3. The proposed autoencoder and the CVAE alternative

W1

It is not quite clear to me why the authors did not use a conditional variational autoencoder rather than the proposed autoencoder. Forcing the encoder to predict both an observation and a latent representation seems unnecessary if the observation can be passed as a condition to both the encoder and decoder like in the standard conditional variational autoencoder architecture.

Q3

In your experiments you empirically show the importance of the latent. Can you show why learning both the observation and latent is required rather than just training a standard conditional variational autoencoder where the encoder maps the unknown field x observation to a latent space?

Response

In OCL-PDE, the autoencoder learns an observation-complementary latent representation zz for reconstruction. The encoder-produced latent zz do not generally follow a standard Gaussian distribution, so we learn a separate model to sample from it. For this separate model, we use conditional Flow Matching to transform standard Gaussian noise into latent samples suited to the observation yy.

We interpret the standard conditional variational autoencoder mentioned by the reviewer as the model introduced by Sohn et al. (2015) [1] in Learning Structured Output Representation using Deep Conditional Generative Models, hereafter referred to as CVAE.

First, we use two encoder outputs to encourage the latnt zz to complement the observation yy. The latent zz is trained to capture features of cc that are difficult to recover from yy alone. As in residual learning [2], starting from an observation-based reconstruction lets the latent focus on correcting the details that the initial reconstruction misses, which can improve reconstruction accuracy. A CVAE encoder only outputs the latent, and do not have this explicit incentive to learn a complementary latent representation. In terms of encoder inputs and outputs, the distinction is

\qquad E_{\mathrm{CVAE}}(c,y)=z,$$ Second, we separate learning the latent representation from learning how to generate it. We first train the autoencoder to learn a latent representation $z$ of the unknown field $c$. We then freeze the autoencoder and train a conditional flow on the encoder-produced latent. This allows the flow to learn how to generate $z$ given $y$ while preserving the representation of $c$ learned in the first stage. A standard CVAE instead jointly learns the recognition model $q_\phi(z\mid c,y)$ and the conditional prior $p_\eta(z\mid y)$ through $D_{\rm KL}\\!\left(q_\phi(z\mid c,y)\\,\\|\\,p_\eta(z\mid y)\right)$ term. CVAE use this coupling to train the model, which may encourage $z$ to depend mainly on $y$, retaining less information about $c$ beyond what the observation already provides. Third, OCL-PDE is a framework which solves both the forward and inverse problems. Our encoder serves as a forward solver, predicting the PDE response from the input field $c$. This is also useful when numerical field–response pairs are scarce. We can sample coefficient or initial-state fields and use the encoder's predicted PDE responses to train the decoder and conditional flow without additional numerical solves. The Binary Darcy topology figure shows distinct Binary Darcy reconstructions generated by OCL-PDE for the same pressure observations. **Binary Darcy topology figure (for the revised manuscript).** Binary Darcy at 5% pressure observations. Each row shows the reference and the first four draws in generation order. The titles report relative pressure observation errors and Betti numbers $(\beta_0,\beta_1)$, which count connected components and holes in the high-permeability phase. **References** [1] [Sohn et al., Learning Structured Output Representation using Deep Conditional Generative Models, NeurIPS 2015](https://proceedings.neurips.cc/paper_files/paper/2015/hash/8d55a249e6baa5c06772297520da2051-Abstract.html). [2] He et al., [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385), CVPR 2016.

W2. Baseline comparisons for continuous Darcy flow

Reviewer comment

In the related work section, the authors state quite a few related methods for solving PDE inverse problems. In the experimental section, however, their approach is only compared to two previous works for the Darcy flow problem with a continuous coefficient and coarse observations.

Response

Our experiments focus on the backward Allen–Cahn problem and also evaluate coefficient recovery in Darcy flow with both binary and continuous coefficients. We added DPS [1], DiffusionPDE [2], and FunDPS [3] as baselines for recovering continuous coefficient from coarse pressure observations. DiffusionPDE and FunDPS are supervised baselines, both are trained with 1,000 numerical coefficient–pressure pairs. OCL-PDE achieved the smallest reconstruction error among the four methods at both noise levels (Table 1). We will add these results to the Experiments section of the revised manuscript.

Table 1. GRF Darcy flow reconstruction before refinement. Entries are mean relative L2L_2 errors (%) of the permeability estimates, averaged over 64 test fields. Each estimate averages 16 posterior samples. Half of the pressure observations are clean anchors.

Method1% pressure noise3% pressure noise
DPS11.2911.58
DiffusionPDE39.9739.21
FunDPS14.8015.05
OCL-PDE (ours)3.445.21

Q1. The known query in Figure 1

Reviewer comment

Please can you specify how the known query in Figure 1 is incorporated into the autoencoder? The textual description does not match up with the Figure here.

Response

The known query (x,t)(x,t) enters the autoencoder as coordinate channels. In Allen–Cahn, x=(x1,x2)x=(x_1,x_2) gives the spatial coordinates on the 128×128128\times128 grid, and tt is a single physical time shared by all grid points. We repeat tt across the grid to form the three-channel query map (x1,x2,t)(x_1,x_2,t). The encoder uses the initial field ϕ0\phi_0 and this query map to predict the PDE response ϕt\phi_t. The decoder uses the response, the same query map, and the latent variable zz to reconstruct ϕ0\phi_0. Equations (2) and (4) omit the query inputs for brevity.

Q2. The latent regularizer in Equation 5

Reviewer comment

The loss function in Equation 5 is only partly defined. Please can you specify how LregL_{\rm reg} is defined similar to how you defined the other three loss terms?

Response

Following the VAE formulation of Kingma and Welling (2014) [4], LregL_{\rm reg} is the KL divergence from the encoder's diagonal Gaussian to N(0,I)\mathcal N(0,I), averaged over latent components and training samples. Here latent variable z∈Rdz\in\mathbb{R}^d. Appendix A.2, Eq. (16) of the submitted paper gives its explicit form:

\bigl(\mu_i^2+\sigma_i^2-\log\sigma_i^2-1\bigr)\right],$$ where $\mu_i$ and $\sigma_i$ are the encoder output mean and standard deviation for latent component $i$ given the training field $c$. Note that, in OCL-PDE, the KL term, weighted by $\lambda_{\rm reg}=10^{-4}$, serves only as a regularizer on the means and variances of the latents distributions. Thus, the latent distribution generally does not equal a standard Gaussian. This term helps keep the latent on a consistent scale. We therefore use a separate conditional flow matching [5] to learn the latent distribution. In the revised Methodology section, we will give the explicit formula for $L_{\rm reg}$ alongside the other loss terms in Eq(5). **References** [1] Chung et al., [Diffusion Posterior Sampling for General Noisy Inverse Problems](https://arxiv.org/abs/2209.14687), ICLR 2023. [2] Huang et al., [DiffusionPDE: Generative PDE-Solving under Partial Observation](https://proceedings.neurips.cc/paper_files/paper/2024/hash/eb3878c1dcbfff9ee95d5d033e5f5942-Abstract-Conference.html), NeurIPS 2024. [3] Yao et al., [Guided Diffusion Sampling on Function Spaces with Applications to PDEs](https://proceedings.neurips.cc/paper_files/paper/2025/hash/b8a28baaa25cebfa9345a8a2edf7f19d-Abstract-Conference.html), NeurIPS 2025. [4] Kingma and Welling, [Auto-Encoding Variational Bayes](https://arxiv.org/abs/1312.6114), ICLR 2014. [5] Lipman et al., [Flow Matching for Generative Modeling](https://arxiv.org/abs/2210.02747), ICLR 2023. Liu et al., [Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow](https://arxiv.org/abs/2209.03003), arXiv 2022.

Dear authors,

thank you very much for the rebuttal. Because of the very limited time until the end of the author-reviewer discussion period, I unfortunately won't be able to go through the rebuttal in detail before the deadline. I will however consider the rebuttal in the upcoming discussion period with the AC.

Reviewer_r3bGOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OCL-PDE combines a physics-aware auto-encoder with conditional flow matching to recover unknown fields from PDE observations. The decoder uses an observation-based reconstruction together with a learned latent correction, intended to recover details that the observation pathway misses. The encoder also provides a forward PDE predictor. Experiments cover backward Allen–Cahn recovery and Darcy coefficient inversion with continuous and binary coefficients, including sparse and noisy observations.

Critical strengths
  1. The reconstruction improvements are encouraging. The comparison with the parameter-matched No-Z model is particularly useful, since it provides evidence that adding the latent pathway helps beyond simply increasing model size. For Allen–Cahn, the relative reconstruction error decreases from approximately 0.098 to 0.059. Improvements also appear across all three observation densities in binary Darcy.
Critical weaknesses
  1. The need for the generative stage is not fully established. The reported point estimates aggregate multiple generated reconstructions. While the No-Z comparison shows that the latent pathway is useful, it does not establish whether generating latents with flow matching is better than predicting a latent deterministically from the observation using the same decoder. This distinction matters because some of the improvement could come from the learned representation and decoder rather than the generative model itself.

  2. The advantage under noisy observations is less consistent. Without clean anchors, OCL-PDE and No-Z have almost identical coefficient errors at 3% noise in continuous Darcy. The appendix also shows that pressure refinement increases OCL-PDE’s reconstruction error when only 0–5% of observations are clean anchors. These results are reported transparently, but they leave uncertainty about when the additional complexity provides a worthwhile benefit under noisy measurements.

Questions for authors
  1. How does flow-based latent generation compare with a deterministic observation-to-latent predictor using the same frozen decoder? How much does reconstruction accuracy change when using one sample instead of the reported sample average?

  2. How sensitive is the pressure-refinement procedure to small errors in the supposedly clean anchors? Is there an observable criterion for deciding when refinement is likely to help rather than worsen reconstruction?

2 comments

We thank the reviewer for reviewing our manuscript and for providing insightful and constructive comments.

W1/Q1. The need for the generative stage

W1

The need for the generative stage is not fully established. The reported point estimates aggregate multiple generated reconstructions. While the No-Z comparison shows that the latent pathway is useful, it does not establish whether generating latents with flow matching is better than predicting a latent deterministically from the observation using the same decoder. This distinction matters because some of the improvement could come from the learned representation and decoder rather than the generative model itself.

Q1

How does flow-based latent generation compare with a deterministic observation-to-latent predictor using the same frozen decoder? How much does reconstruction accuracy change when using one sample instead of the reported sample average?

Response

PDE inverse problems are inherently ill-posed: many fields can match the same observations. Fine-scale structures are difficult to recover. To bridge this gap, we introduce an observation-complementary latent (OCL) that specifically captures the fine-scale information.

To capture fine-scale details, OCL-PDE couples an autoencoder with a conditional Flow Matching sampler. The encoder extracts both the observation (PDE response) and the observation-complementary latent zz from the input field, while the decoder takes both to reconstruct the field. Similar to residual learning [1], starting from an observation-based estimate lets zz focus solely on details, thereby improving reconstruction accuracy.

Because the observation-complementary latent zz carries specific physical meaning governed by the PDE operator, its distribution is inherently non-standard Gaussian. We cannot simply draw samples from a standard normal distribution as decoder input at inference. Therefore, we train a conditional Flow Matching model to transform standard Gaussian noise into valid, observation-conditioned latent samples.

A generative sampler is essential here because a deterministic mapping cannot capture the full solution space of an ill-posed problem. A deterministic map from yy to zz yields only a single point estimate, failing to represent other physically plausible fields. By contrast, drawing different noise samples in conditional Flow Matching generates diverse latents for the same yy, enabling the decoder to produce a set of distinct, plausible candidate reconstructions.

Quantitative and visual validations confirm that OCL-PDE generates diverse yet physically consistent solutions. We preselected two test fields per task, drew 16 candidates per field, and checked each candidate with an independent PDE solver. All 96 candidates had observation errors below 2.392.39\\% (Table 1). Figure 13 in Appendix G of the revised manuscript shows topologically distinct Binary Darcy reconstructions that all accurately reproduce the identical pressure measurements.

Table 1. Independent-solver observation errors for 16 candidates per test field.

TaskField 0 meanField 0 maximumField 1 meanField 1 maximum
Allen–Cahn0.721%1.120%1.664%2.386%
GRF Darcy1.305%1.466%1.415%1.690%
Binary Darcy0.832%1.210%1.145%1.965%

Binary Darcy topology figure (for the revised manuscript). Binary Darcy at 5% pressure observations. Each row shows the reference and the first four draws in generation order. The titles report relative pressure observation errors and Betti numbers (β0,β1)(\beta_0,\beta_1), which count connected components and holes in the high-permeability phase.

The generated samples also help identify where reconstruction errors are larger. Across 64 test fields with 16 draws per field, the mean Pearson correlations between sample standard deviation and absolute error were 0.4840.484 for Allen–Cahn, 0.7510.751 for Binary Darcy at 5% observations, and 0.4510.451 for GRF Darcy at 3% noise. The positive correlations show that regions with greater variation across generated samples tend to have larger reconstruction errors.

In the revised Methodology section, we will explain why we use a generative stage in ill-posed inverse problems. We will add these examples to the Experiments section and report their observation errors computed using an independent PDE solver.

References

[1] He et al., Deep Residual Learning for Image Recognition, CVPR 2016.

W2/Q2. Noisy observations and pressure refinement

W2

The advantage under noisy observations is less consistent. Without clean anchors, OCL-PDE and No-Z have almost identical coefficient errors at 3% noise in continuous Darcy. The appendix also shows that pressure refinement increases OCL-PDE's reconstruction error when only 0–5% of observations are clean anchors. These results are reported transparently, but they leave uncertainty about when the additional complexity provides a worthwhile benefit under noisy measurements.

Q2

How sensitive is the pressure-refinement procedure to small errors in the supposedly clean anchors? Is there an observable criterion for deciding when refinement is likely to help rather than worsen reconstruction?

Response

At 1% pressure noise, refinement increases the coefficient error for all methods in the anchor sweep at 0–5% clean-anchor coverage. With even fewer clean anchors, the pressure update may lack sufficient reliable constraints, this problem exists for all baselines. But among the tested methods, OCL-PDE benefits from refinement at the lowest clean-anchor coverage (10%). At 3% pressure noise, OCL-PDE also starts to benefit at 10% coverage.

Here, 10% means that only 90 of the 900 interior pressure measurements are clean and the other measurements remain noisy. Our goal is to use these 900 measurements to reconstruct the 128×128128\times128 coefficient field. The number of clean anchors is extremely small relative to the number of pixels of the unknown fields which we need to reconstruct. OCL-PDE uses the clean anchors more effectively, which is why OCL-PDE benefits from refinement at the lowest clean-anchor coverage compared to the other methods. We will add a discussion of this point in the revised Experiments section.

Reviewer_R3MnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper introduces OCL-PDE, a generative framework for PDE inverse problems. The central idea is to learn an observation-complementary latent representation that provides additional information for reconstructing unknown coefficient or initial-condition fields. The framework combines a physics-aware autoencoder with conditional Flow Matching to generate these latents from observations.
  • The method is evaluated on backward Allen–Cahn recovery and continuous and binary Darcy coefficient inversion. The experiments demonstrate substantial improvements in reconstruction accuracy over the evaluated baselines, particularly for Allen–Cahn and binary Darcy.

Minor note:

  • Mathematical notation The mathematical notation is occasionally imprecise. For example, the union Dp∪Du\mathcal D_p\cup\mathcal D_u in Eq. (5) is not well defined, since Dp\mathcal D_p contains field-response pairs while Du\mathcal D_u contains only fields. Also, Bphys\mathcal B_{\mathrm{phys}} in the physics loss is not clearly defined.
Critical strengths
  • S1 Motivation and presentation The problem of recovering unknown physical fields from incomplete observations of PDE-governed systems is a highly relevant problem in scientific machine learning. The motivation is clear, and the introduction and overview provide a good understanding of the proposed approach.
  • S2 Observation-complementary latents The central idea of separating information provided by the observation from the remaining information captured by a learned latent is interesting and natural. Particularly for ill-posed inverse problems, where observations might determine the large-scale structure but leave fine-scale details ambiguous, this seems like a promising approach.
  • S3 Strong reconstruction results The reported improvements over the evaluated baselines are substantial. The authors also provide useful ablation studies, including parameter-matched No-Z baselines, which demonstrate that the added observation-complementary latent contributes to reconstruction accuracy.
Critical weaknesses
  • W1 Limited methodological novelty and justification of design choices

    • The proposed framework combines several established components, including a physics-informed autoencoder with an observation-conditioned decoder, VAE-style latent regularization, and conditional Flow Matching. The main contribution appears to be the particular combination and architecture rather than a fundamentally new methodological idea.
    • Furthermore, the framework is relatively complicated, with many architectural and training decisions whose necessity is not sufficiently investigated. For instance, the authors use a weak-form physics loss, which is mathematically reasonable, but do not investigate its benefit over a strong-form formulation or training without physics regularization.
    • Most importantly, the motivation for using Flow Matching remains unclear in light of the evaluation. The main focus is on producing accurate point estimates, rather than capturing the distribution of plausible inverse solutions. The No-Z ablation demonstrates the benefit of the latent pathway, but does not establish whether conditional generative modeling is actually necessary. A deterministic predictor from observation to latent might achieve similar results if applied in a similarly sophisticated framework.
    • Consequently, I find the justification for the overall complexity somewhat limited. A clearer understanding of which components are essential, and why, would strengthen the methodological contribution.
  • W2 Averaging of samples and evaluation of generative inverse modeling

    • A central concern is the averaging of generated reconstructions before evaluation. For continuous fields, the authors literally average the reconstructed coefficient or initial-condition fields. For binary Darcy, they average phase probabilities before thresholding. While this is reasonable for obtaining a point estimate, it can be problematic for genuinely ambiguous inverse problems.
    • For instance, consider a multimodal distribution of plausible coefficient fields, where two structurally different coefficients produce essentially identical observations under the PDE solution operator. Averaging samples may reduce the expected reconstruction error while producing a coefficient field that is itself unlikely under the underlying distribution. However, a generative model correctly representing both possibilities would be penalized for sampling a valid alternative to the particular ground truth.
    • This relates to a more general limitation of the evaluation. The authors focus predominantly on reconstruction accuracy against ground-truth fields. While they additionally report errors of independently computed PDE solutions against the full reference solutions, they do not explicitly evaluate the consistency of individual generated reconstructions with the available observations under the PDE solution and observation operators.
    • More importantly, the evaluation does not establish whether the learned conditional distribution captures the ambiguity inherent in ill-posed inverse problems, despite specifically choosing to use a generative model in the framework. Recent work increasingly emphasizes evaluation of the posterior, including uncertainty calibration and coverage (e.g., Falkiewicz et al., NeurIPS 2023; Zong et al., CMAME 2025; Zach et al., ICLR 2026).
    • I would therefore expect an evaluation of individual samples, their consistency with the observations under the governing PDE, and whether their variability meaningfully reflects the uncertainty of the inverse problem. This seems particularly important for justifying the generative formulation in the engineered framework.
  • W3 Missing related work and generative inverse baselines

    • I find the selection of generative baselines somewhat incomplete. In particular, FunDPS (Yao et al., NeurIPS 2025) appears to be a highly relevant missing comparison. It addresses similar PDE inverse problems, including binary Darcy coefficient reconstruction from sparse observations, using unconditional diffusion models with inference-time guidance.
    • More generally, there is an established line of work on Bayesian inverse problems with learned generative priors. These approaches learn an unconditional distribution over unknown coefficient fields. Then, they perform inference using the PDE solution operator, either through a numerical solver or a learned surrogate. Examples include Mosser et al. (2020) for seismic waveform inversion, Patel et al. (2022) for physics-based Bayesian inverse problems with GAN priors, or Meng et al. (2022) for learned functional priors combined with physics-informed models.
    • Such methods appear particularly relevant here, as they explicitly use generative models to represent plausible coefficient structures and enforce consistency with the observations during inference. I would expect a more thorough discussion and, ideally, an experimental comparison.
  • W4 Questions regarding the empirical evaluation and sample efficiency

    • I was surprised by the relatively poor performance of some generative baselines, particularly DiffusionPDE. In the binary Darcy experiment, DiffusionPDE performs substantially worse than PINO. Interestingly, the opposite is reported in FunDPS, where DiffusionPDE significantly outperforms PINO on Darcy inversion. While the experimental setups are not identical, this large reversal raises questions regarding the baseline implementations, training procedures, and general comparability of the results.
    • Furthermore, the relatively small number of numerically paired training examples might be an interesting advantage of the proposed approach. However, the authors also use additional dynamically generated fields during training. The actual sample efficiency is therefore difficult to assess. There is no systematic investigation of performance as a function of the number of available numerical training pairs. Therefore, as currently presented, it is unclear whether 1) the full-field generative baselines (e.g. DiffusionPDE) are performing poorly due to insufficient training/data and 2) whether the proposed method in fact is more sample efficient.
    • In combination with the missing evaluation of individual generated samples discussed in W2, this makes it difficult to determine whether the observed improvements arise from a more effective generative method, the architectural design, or differences in training and evaluation procedures.
Questions for authors
  • Q1 (W1) Could the authors provide justifications for the main design choices, particularly the use of Flow Matching and weak-form physics regularization? Would a deterministic observation-to-latent predictor, using the same encoder and decoder, achieve comparable reconstruction accuracy?

  • Q2 (W2) Could the authors evaluate individual reconstructions rather than only their averages? In particular, how well do individual samples reproduce the available observations under the PDE solution operator? Furthermore, does the learned conditional distribution meaningfully capture the uncertainty of the inverse problem?

  • Q3 (W3) Could the authors discuss and ideally compare against FunDPS and Bayesian inverse approaches based on unconditional generative priors for the coefficient fields? These appear particularly relevant for evaluating the benefits of the proposed observation-complementary latent representation.

  • Q4 (W4) Could the authors clarify the substantial differences in relative baseline performance compared with previously published results, particularly the reversal between PINO and DiffusionPDE on Darcy inversion? Furthermore, could the authors provide a more detailed analysis of sample efficiency, distinguishing the number of numerically paired examples from the additional generated training data?

9 comments

We thank the reviewer for reviewing our manuscript and for providing insightful and constructive comments.

S1. Motivation and presentation

Reviewer comment

The problem of recovering unknown physical fields from incomplete observations of PDE-governed systems is a highly relevant problem in scientific machine learning. The motivation is clear, and the introduction and overview provide a good understanding of the proposed approach.

Response

We prefer to use "sparse" to describe the observation instead of "incomplete", because there is no "complete" observation.

W1a. The role of the observation-complementary latent

Reviewer comment

The proposed framework combines several established components, including a physics-informed autoencoder with an observation-conditioned decoder, VAE-style latent regularization, and conditional Flow Matching. The main contribution appears to be the particular combination and architecture rather than a fundamentally new methodological idea.

Response

PDE inverse problems are inherently ill-posed: many fields can match the same observations. Fine-scale structures are difficult to recover. To bridge this gap, we introduce an observation-complementary latent (OCL) that specifically captures the fine-scale information.

In OCL-PDE, the autoencoder learns an observation-complementary latent representation zz for reconstruction. The encoder-produced latent zz generally do not follow a standard Gaussian distribution, so we train a separate model to sample from it. For this separate model, we use conditional Flow Matching to transform standard Gaussian noise into latent samples suited to the observation yy.

To address the reviewer's concern, we clarify the distinct roles within OCL-PDE: the observation guides large-scale structures, while the latent supplies complementary fine-scale details. This decomposition motivates our architecture: the encoder generates both measurements and the latent, while the decoder takes them as inputs. This is similar to residual learning [1], starting from an observation-based estimate allows the latent to focus solely on the fine-scale details, thereby improving overall reconstruction accuracy and recovering fine-scale details more faithfully.

Crucially, the latent distribution is not a standard Gaussian and the VAE-style KL term acts solely as a regularizer to stabilize the latent scale. Consequently, we cannot simply sample the latent zz following a standard Gaussian at inference. To resolve this, we employ conditional flow matching to transform standard noise into valid, observation-conditioned latent samples. In the revised Methodology section, we will explain how OCL-PDE learns the observation-complementary latent, uses conditional Flow Matching to sample it.

W1a / Q1. Weak-form and strong-form physics losses

Reviewer comment

Furthermore, the framework is relatively complicated, with many architectural and training decisions whose necessity is not sufficiently investigated. For instance, the authors use a weak-form physics loss, which is mathematically reasonable, but do not investigate its benefit over a strong-form formulation or training without physics regularization.

Response

We chose the weak-form PDE loss because integration by parts reduces the required spatial derivative order by one and avoids differentiating discontinuous coefficients, relaxing regularity requirements. This is a common practice in related methods [2].

We tested this choice by changing the training PDE residual to the strong form. OCL-PDE achieves similar reconstruction accuracy with both physics-loss formulations. The weak and strong forms yield errors of 5.94% and 6.32% on Allen–Cahn, and 2.43% and 2.27% on Binary Darcy at 5% observations, respectively; GRF errors are also close. These results shows the flexibility of our framework: the complementary latent can be combined with either formulation.

References

[1] He et al., Deep Residual Learning for Image Recognition, CVPR 2016.

[2] Zang and Koutsourelakis, DGenNO, Section 3.1 and Appendix A. Bao et al., PFWNN: A deep learning method for solving forward and inverse problems of phase-field models, Journal of Computational Physics, 2025.

W1b / Q1 / Q2. The generative stage and plausible individual-candidates

Reviewer comment

Most importantly, the motivation for using Flow Matching remains unclear in light of the evaluation. The main focus is on producing accurate point estimates, rather than capturing the distribution of plausible inverse solutions. The No-Z ablation demonstrates the benefit of the latent pathway, but does not establish whether conditional generative modeling is actually necessary. A deterministic predictor from observation to latent might achieve similar results if applied in a similarly sophisticated framework.

Q1

Could the authors provide justifications for the main design choices, particularly the use of Flow Matching and weak-form physics regularization? Would a deterministic observation-to-latent predictor, using the same encoder and decoder, achieve comparable reconstruction accuracy?

Q2

Could the authors evaluate individual reconstructions rather than only their averages? In particular, how well do individual samples reproduce the available observations under the PDE solution operator? Furthermore, does the learned conditional distribution meaningfully capture the uncertainty of the inverse problem?

Response

A generative sampler is essential here because a deterministic mapping cannot capture the full solution space of an ill-posed problem. A deterministic map from yy to zz yields only a single point estimate, failing to represent other physically plausible fields. By contrast, drawing different noise samples in conditional Flow Matching generates diverse latents for the same yy, enabling the decoder to produce a set of distinct, plausible candidate reconstructions.

We preselected two test fields per task, drew 16 candidates per field, and checked each candidate with an independent PDE solver. All 96 candidates had observation errors below 2.392.39\\% (Table 1). The Binary Darcy topology figure for the revised paper shows topologically distinct Binary Darcy reconstructions whose simulated pressures closely match the same observations.

Table 1. Independent-solver observation errors for 16 candidates per test field.

TaskField 0 meanField 0 maximumField 1 meanField 1 maximum
Allen–Cahn0.721%1.120%1.664%2.386%
GRF Darcy1.305%1.466%1.415%1.690%
Binary Darcy0.832%1.210%1.145%1.965%

Binary Darcy topology figure (for the revised paper). Binary Darcy at 5% pressure observations. Each row shows the reference and the first four draws in generation order. The titles report relative pressure observation errors and Betti numbers (β0,β1)(\beta_0,\beta_1), which count connected components and holes in the high-permeability phase.

In the revised Methodology section, we will explain why we use a generative stage in ill-posed inverse problems. The submitted manuscript omitted examples of distinct reconstructions. We will add these examples to the Experiments section and report their observation errors computed using an independent PDE solver.

W2a / Q2. Point estimates and uncertainty information

Reviewer comment

A central concern is the averaging of generated reconstructions before evaluation. For continuous fields, the authors literally average the reconstructed coefficient or initial-condition fields. For binary Darcy, they average phase probabilities before thresholding. While this is reasonable for obtaining a point estimate, it can be problematic for genuinely ambiguous inverse problems.

Q2

Could the authors evaluate individual reconstructions rather than only their averages? In particular, how well do individual samples reproduce the available observations under the PDE solution operator? Furthermore, does the learned conditional distribution meaningfully capture the uncertainty of the inverse problem?

Response

we evaluate OCL-PDE across two complementary dimensions: point-estimate accuracy (via ensemble averaging) and individual-sample validity (without averaging).

  1. Ensemble averaging is only used for point-estimate comparison against a single ground truth. In standard inverse problem benchmarks, each test observation is generated from one reference field, making a point estimate necessary for direct pixel-wise comparison.

  2. Crucially, we evaluate individual, unaveraged samples to verify physical consistency and uncertainty quantification (UQ). Rather than relying solely on the mean, we directly pass individual generated fields through the forward PDE solver. This confirms that distinct reconstructions accurately reproduce the same observations, such as topologically different Binary Darcy fields (the Binary Darcy topology figure for the revised paper). Furthermore, the spread across individual samples provides reliable spatial UQ: as shown in Table 2 and the field 0 error-localization figure for the revised paper (consistent with analyses in DGenNO [1]), regions with higher sample standard deviation strongly correlate with larger reconstruction errors (positive Pearson correlation).

We will incorporate uncertainty quantification metrics into the main text of the revised manuscript.

Table 2. Pearson correlations for error localization over 64 fields (K=16K=16).

MethodACBinary, 5% observedGRF, 3% noise
OCL-PDE.484.751.451

Error-localization figure for field 0 (for the revised paper). Reference, mean of 64 draws, absolute error and sample standard deviation for field 0. Error and standard deviation share a scale within each row.

W2a / Q2. Multiple plausible solutions

Reviewer comment

For instance, consider a multimodal distribution of plausible coefficient fields, where two structurally different coefficients produce essentially identical observations under the PDE solution operator. Averaging samples may reduce the expected reconstruction error while producing a coefficient field that is itself unlikely under the underlying distribution. However, a generative model correctly representing both possibilities would be penalized for sampling a valid alternative to the particular ground truth.

Response

As the individual draws in our response to W1 and the illustration (the Binary Darcy topology figure for the revised paper), OCL-PDE can generate multiple plausible solutions for one observation.

The Binary Darcy samples in the Binary Darcy topology figure for the revised paper have different topologies, as indicated by their Betti numbers, while their simulated pressures closely match the same observations.

References

[1] DGenNO, Section 4.4 and Figure 12.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.