acceptodds
Under review as a conference paper at ICLR 2027

PhysDEM: Physics-Defined Energy-Matching Diffusion for Spatiotemporal Field Generation under Scarce Measurements

Abstract

Generating and predicting spatiotemporal physical fields from scarce measurements is challenging, as observations are insufficient to characterize a distribution over complete fields. This limits conventional data-driven diffusion models that rely on full-field datasets. We introduce PhysDEM, a physics-defined diffusion framework that combines governing equations with spatially sparse observations to generate multiple plausible fields. First, we construct a Gibbs target by reweighting a measurement-conditioned Gaussian reference with PDE residual energy. Second, we derive an exact conditional-mean identity that reduces denoising to supervised learning of the standardized energy-induced mean correction. Third, a physics-displacement probability flow cancels Gaussian reference terms and enables amortized sampling with changing measurements through Gaussian conditioning, without retraining. Experiments on synthetic PDE systems and real-world-informed applications demonstrate that PhysDEM supports coherent field recovery and efficient sampling while maintaining stable diagnostics under tested noise levels, illustrating its practical value for field assessment. To our knowledge, PhysDEM is the first physics-defined diffusion model enabling amortized spatiotemporal field inference without preassembled full-field datasets.

Primary area:
generative models
Keywords:
Physical Diffusion Sampler, Spatiotemporal Field Generation, Few-Shot Inferring, Energy Matching, Governing Equations

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

This work proposes Physics-Defined Energy-Matching Diffusion (PhysDEM) as a way of retrieving the complete map of coefficients and their associated solution fields from sparse and noisy measurements. The method was tested on several PDE problems and applications involving reservoir permeability and pressure reconstruction, as well as a contamination transport problem using the Bemidji dataset. The main contribution is a theoretical construction that proposes pairs of coefficients and their associated solutions through a linearized model and constructs a reference for the coefficients-solution probability distribution using a Gaussian energy function and learns, through a diffusion model, the correction residual between the reference and the complete distribution, using data generated by MCMC chains rather than a premade dataset, resulting in a model that can produce alternative reconstructions without the need for retraining.

Critical strengths
  1. I believe that the method and the proposal are an elegant solution to model coefficient and solution fields jointly. The formulation provides a way to avoid solving complex systems that may be easy to evaluate but hard to solve with standard methods, offering an alternative in such cases.
  2. The design choices for the method are theoretically well justified. The authors propose the reference distribution analytically and derive the necessary correction to be learned. They establish how the sampler should use the correction and the learning objective, connecting the original problem of learning the complete distribution to residual learning from this reference, giving the neural network a defined role rather than treating the model as a black box.
  3. The trained sampler can be used across multiple configurations without retraining.
  4. The paper is well written and easy to understand.
Critical weaknesses
  1. Despite the elegant formulation, I find the evaluation of this work flawed as it is mainly statistical and not consistent across all experiments. Furthermore, apart from the results in the appendix, the evaluation does not report reconstruction errors of the sampled coefficients and their respective solutions. The results in the appendix are also concerning; relative errors of 1.028 and 1.019 can be obtained with a near-zero prediction, showing that the method, in my opinion, fails in these problems. I find the authors' justification of weak constraints and that further examination should be done regarding this.
  2. Related to the previous point, I believe that the statistical metrics are not enough to support the claims about the generated fields. These metrics have substantial flaws. For instance, spatial summaries can hide misplacement issues and average rank can balance if you have the same amount of overprediction and underprediction. These can be useful to characterize the learned distribution but should not replace a full field recovery metric set.
  3. The benefit of the correction is unclear over the reference. The paper shows that the spread-skill improves substantially from the reference distribution by incorporating the correction. However, it does not specify the component responsible for this improvement as the skill can improve by either improving the reconstruction error because the spread increases or both.
  4. The coarseness of the grid used is also concerning. All the experiments use a 64×6464 \times 64 spatial grid with 48 time points, without a grid study. Training depends on generating good initial solutions using pCN across multiple configurations. The appendix describes all configuration parameters but does not examine how many chains are necessary to train the models on a finer grid, as the model's applicability is currently limited in its current settings.
  5. The method cites as one of its advantages the lack of pre-computed solutions to train the model. This is appealing, as I said before; however, they fail to quantify the difference between computational cost and performance degradation. Avoiding forward solves is potentially useful only if it produces more accurate solutions, or the cost of solving the system is high enough that the time saved justifies the error. None of this is quantified in the paper.
Questions for authors
  1. Could you provide a systematic comparison between the proposed method and the baselines on the full field coefficients and solution reconstructions against a FEM reference? Use normalized L2L_2 error.
  2. Could you compare the reconstruction errors before and after the proposed correction?
  3. Could you provide for one of the systems in the paper an ablation on grid size, reporting how the training cost and accuracy change?
  4. Could you report how many pCN chains you require to achieve your results and also how much it increases as the grid becomes finer?
Reviewer_Wy1GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper models physics fields using diffusion models.It can be applied to sparse observation. In my understanding, it first uses PCN to generate clean samples as training data. It defines a gaussian prior for clean samples. Since the observation here is gaussian linear, the condition distribution (prior given observation) is also a gaussian. Then, the first order in physical energy is absorbed in the gaussian distribution. This guassian distribution is the starting noise of diffusion model. To calculate diffusion score, the physical energy (without first-order) is trained via score matching (or its variant).

Critical strengths

This paper is noval regarding the gaussian/diffusion modeling of physics data. It does not use vanilla diffusion model but applies something new, like diffusion prior, score obsorption.

Critical weaknesses

See questions.

Questions for authors

I would like to thank authors for the dedicated work. I am sorry that the paper is generally hard for me to follow. I am not familiar with some techniques and I hope the author can help me to understand the paper.

  1. In the pCN MCMC, it has a gaussian diffusion term. The data should lie in a low-dimension manifold. The diffusion term may deviate trajectory out of that manifold. Will this make the algorithm inefficient?
  2. If pCN can already model the data distribution, why we need a post-hoc diffusion training?
  3. Will the error of data generated by pCN propagate to diffusion model?
  4. Why we can model the clean sample distribution by \pho_0 ? Does this mean the clean sample itself is already a gaussian distribution? Will it be a strong assumption?

I would raise my score if the author could help me to understand the paper.

Reviewer_A2PsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes PhysDEM, a physics-defined diffusion framework for spatiotemporal field recovery from sparse measurements. It constructs a Gibbs target distribution using a Gaussian prior and PDE residual energy, generates training samples through MCMC, and learns energy-induced denoising corrections for amortized sampling. Experiments on five physical systems evaluate physical consistency, uncertainty calibration, and sampling efficiency.

Critical strengths

The paper presents a well-structured framework that integrates PDE constraints, conditional-mean denoising, and probability-flow sampling. The mathematical derivations are relatively comprehensive, and the ability to accommodate different sparse measurement configurations without retraining is practically appealing.

Critical weaknesses
  1. The necessity and advantages of MCMC-based training data generation are insufficiently justified. Numerical PDE solvers are readily available for many of the considered systems, yet the paper does not compare MCMC sampling with directly generating training fields through PDE simulations. It remains unclear whether the proposed strategy offers advantages in computational efficiency or data quality.
  2. The task is formulated as field recovery, but the experiments primarily emphasize observation consistency, PDE residuals, and uncertainty calibration. These metrics do not directly quantify reconstruction accuracy (e.g., RMSE) in unobserved regions, making it difficult to assess the method's effectiveness for field recovery.
Questions for authors
  1. What motivates the use of MCMC rather than numerical PDE solvers to generate training fields? Could the authors clarify its specific advantages and provide a comparison in terms of computational cost and downstream reconstruction performance?
  2. Given that field recovery is a primary objective, could the authors report RMSE for both static parameter fields and dynamic state fields, with consistent comparisons against the baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.