acceptodds
Under review as a conference paper at ICLR 2027

Aligning Exact-Likelihood Normalizing Flows with Physics-Aware Score Denoising

Abstract

Normalizing flows offer expressive conditional generation and exact likelihood evaluation, making them attractive for scientific field reconstruction. However, practical reconstruction often relies on score-denoised outputs, whose interpretation depends on the density learned during training. Incorporating physical constraints changes this density, creating a potential mismatch between the training objective and the denoising rule. We introduce PAIM-Flow, a framework for aligning exact-likelihood normalizing flows with physics-aware score denoising. PAIM-Flow combines maximum likelihood with physical risk evaluated on fields generated through the inverse flow, while reserving denoising for inference. At the conditional population level, we characterize the density tilt induced by physical supervision and derive an objective-dependent score correction. This analysis clarifies the assumptions required for matched denoising and distinguishes the exact likelihood of the underlying flow from the distribution of denoised outputs. Experiments on Darcy pressure reconstruction, conditional Poisson problems, and sparse turbulent field reconstruction demonstrate substantial improvements in reconstruction accuracy over the evaluated baselines. Controlled ablations disentangle the effects of conditional modeling, physical supervision, and inference-time denoising, separating overall reconstruction gains from the contribution of score correction. Our work provides a principled foundation for connecting physics-aware training objectives with score-based inference in exact-likelihood generative models.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
normalizing flows, exact likelihood estimation, physics-informed machine learning, score-based denoising, training-inference alignment, scientific inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_d7G4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PAIM-Flow to investigate the alignment between physics-aware training and inference-time score denoising in normalizing flows. The method augments maximum likelihood training with a physical risk evaluated on raw generated fields and derives the resulting density deformation and score correction at an idealized population optimum. Experiments on Darcy, Poisson, and sparse turbulent field reconstruction examine the respective effects of physical supervision, ordinary score denoising, and the proposed correction.

Critical strengths
  1. The paper addresses a theoretically meaningful question: physical supervision changes the learned density, requiring the statistical interpretation of the original denoising rule to be reconsidered. Deriving a reciprocal density tilt for the specified objective, rather than assuming an exponential tilt, provides a useful clarification.
  2. The authors distinguish the exact likelihood of the underlying flow from the distribution of denoised outputs. Comparisons using identical model weights and raw samples help isolate the contribution of score correction. The main text also discusses negative results and theoretical limitations candidly.
Critical weaknesses
  1. The performance claims in the abstract are inconsistent with the experimental results. The abstract claims substantial improvements in reconstruction accuracy. However, in Table 1, Darcy RMSE decreases only from 0.006496240 to 0.006495043, corresponding to an approximately 0.018% improvement; the improvement on Poisson is approximately 0.009%, while turbulent reconstruction becomes slightly worse. The physical residual metric also deteriorates on Darcy. Section 5 explicitly acknowledges that the current results do not establish a substantial advantage for PAIM-Flow. The abstract therefore overstates the empirical evidence and should be revised to reflect the actual findings.
  2. The central score correction does not demonstrate a clear practical benefit. In Table 2, with identical PAIM-Flow weights and raw samples, matched denoising changes Darcy RMSE only from 0.006495049 to 0.006495043, while increasing the physical residual from 16.02540 to 16.03952. Section 4.2 further states that matched denoising produces slightly worse RMSE than ordinary denoising at all nine endpoints in the earlier three-seed experiments. These ablations do not establish the practical value of the paper’s central correction. The authors should identify conditions under which the correction provides consistent benefits and validate them through repeated experiments. Otherwise, the paper should be positioned more explicitly as a theoretical analysis of training–inference alignment and an empirical investigation of its limitations.
  3. The baseline comparisons and benchmark coverage are insufficient. The main experiments primarily compare against TARFlow, making it difficult to assess competitiveness against other physical-field reconstruction methods. For sparse reconstruction, the authors should include PhySense, Senseiver, and relevant generative reconstruction baselines under identical observation locations, observation counts, and evaluation protocols. For Darcy and Poisson, comparisons with neural operators or other methods suited to conditional PDE mapping would also be informative. In addition, the authors should evaluate the method on the PhySense sea-temperature benchmark to examine its performance under varying sensor locations and counts, irregular valid domains, and temporal distribution shifts. The experiments should follow the official evaluation protocol and explicitly specify the physical risk or prior used for this task, together with its applicability. This would provide stronger evidence of the method’s effectiveness and generalization ability.
Questions for authors

See Weaknesses.

Reviewer_jHZ7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PAIM-Flow, a framework connecting physics-regularized normalizing-flow training with inference-time score denoising. Its training objective combines maximum likelihood on Gaussian-corrupted data with a bounded physical risk evaluated on generated fields. The authors characterize the conditional population optimum through a reciprocal density tilt and derive an associated score correction. Experiments on Darcy, Poisson, and sparse turbulent-field reconstruction examine the correction in finite trained models, alongside controls for conditional fitting and numerical postprocessing.

Critical strengths

1.The theoretical analysis explicitly connects the physical training objective to the resulting population density and score correction. It distinguishes the raw flow likelihood, the denoising estimator, and the distribution of final outputs. 2.The experiments separate physical training, inference-time matching, and numerical postprocessing. Comparisons using identical checkpoints and raw samples isolate the matching rule’s effect, and the paper reports its empirical limitations.

Critical weaknesses

1.Darcy RMSE in Table 2 changes very little after matching. It would help to connect this result more directly to the reported correction-magnitude and direction diagnostics, so readers can understand what mainly limits the correction’s effect. 2.Darcy and Poisson use PDE residuals, whereas the turbulence task uses wall and smoothness constraints. I would like a clearer discussion of how this difference affects the interpretation of the matching results across tasks. 3.The current normalizer estimate requires eight additional generated samples per condition. Reporting this calibration cost separately would make the practical inference trade-off easier to assess.

Questions for authors

1.Taken together, what do the existing diagnostics suggest is the main reason for the limited matching effect in the Darcy experiment? 2.Do the authors expect matching to behave differently with a PDE residual versus a wall/smoothness proxy? How should readers interpret the three tasks in this respect? 3.How much inference time is spent generating the eight-sample calibration bank, compared with raw generation and score evaluation?

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

This paper proposed the PAIM-Flow to target to incorporate physical constraints into the conditional generative model training and inference. PAIM-Flow is a normalizing flow method built on the top of TARFlow. More precisely, aside from the NLL loss for normalizing flow training, the author also adapted a physics-informed sample regularization term based on the solution space to incorporate the physics constraints. They also provided an inference matching scheme and mathematically proved the benefits compared to the vanilla TARFlow method with joint law. For the experiment part they evaluated the PAIM-Flow with TARFlow across three benchmarks: Darcy flow, Poisson’s equation and Turbulent flow. Despite that experimental result does not show significance, they also tried to explain the cause of this discrepancy between the theoretical analysis and empirical performance. Overall this is a very interesting work that is rooted in a special family of generative model training.

Critical strengths

This is a nice work and the authors involved in this paper have a very strong mathematical background. Especially for the method part, the author introduced their physics-aware objective and the inference matching mechanism in a clean and tidy way. They also provided a very rigorous analysis based on the potential performance gain after they adapted the new scheme to the vanilla TARFlow. Despite that the experiment result is not ideal, they also tried their best to provide the reasoning behind this. Overall, this paper is well written and rigorous in math.

Critical weaknesses

My comments will be majorly based on the experiments. First and most critically is the improvement of PAIM-Flow compared to the baseline TARFlow showing in the validation results is trivial on all the three Darcy, Poisson, Turbulent tasks. This is insufficient to show the effectiveness of their proposed physics-aware training objectives and the score-based inference either in the reconstruction quality or in reducing physics violations. Despite that the author provided several explanations for this failure, it will still be considered as the failure of the whole method. Besides that, for the three physics tasks provided, there are also other methods that can perform steady-state snapshot prediction, for example the optimization based families and PINNs. I think it is also important to compare the current method with those methods to show the advantages and disadvantages. Also based on my quick search, I also noticed that the ICLR 2026 paper https://github.com/tum-pbs/PBFM is doing a Physics-Based Flow Matching which could serve as a baseline for your paper. Another thing is about the evaluation metrics, aside from the regular metrics of RMSE, SSIM, RMAE, I would also recommend you to consider wall clock time and NFE.

Questions for authors
  1. Could you explain at a high level why TARFlow—an autoregressive normalizing flow—was selected as the base architecture for this task? What is the advantage of using this backbone for your task compared to other generative models, for example, flow matching and diffusion models?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.