acceptodds
Under review as a conference paper at ICLR 2027

Numerical Calibration for Physics-Informed Diffusion Posterior Sampling

Abstract

Physics-informed diffusion posterior sampling aims to improve the physical consistency of reconstructions by incorporating physical constraints into the reverse process of a pretrained diffusion model. These constraints are imposed through physics guidance, which acts alongside the prior score during posterior sampling to steer the reverse process. However, physics guidance requires an explicit numerical realization of the PDE at inference time, whereas the numerical realization underlying the pretrained prior is typically overlooked when constructing the guidance and may even be entirely unknown. We show that this numerical asymmetry matters at finite resolution: different numerical realizations of the same continuous PDE can produce substantially different residuals and physics gradients for the same solution estimate, especially near the resolution limit. The resulting numerical mismatch can cause physics guidance to conflict with the prior score and observation guidance, ultimately degrading posterior reconstruction. These findings suggest a simple principle: when the numerical realization underlying the prior is known, physics guidance should use a matched realization. When it is unknown, we introduce Numerically Calibrated Physics Guidance (NCPG), which learns the numerical relation implicit in paired training data without access to the original solver or knowledge of its discretization. Once learned, the calibration model is frozen and reused across posterior sampling tasks. Experiments on linear and nonlinear PDEs show that numerical mismatch can substantially degrade posterior sampling performance, while NCPG largely mitigates this degradation and, in some settings, even outperforms numerically matched physics guidance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Diffusion Posterior Sampling, Physics-Informed

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GjrrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Numerically Calibrated Physics Guidance (NCPG) to address the, possible, mismatch between a diffusion prior's training data and some physics guidance.

Critical strengths

None

Critical weaknesses

I have to emphasize that this paper is poorly written with unnecessary complications (e.g., "numerical realization" for "discretization", "numerical provenance" for "data-generating numerical method", etc.). This makes reading and assessing the paper and its contributions difficult.

Other major issues:

  • The setting is poorly motivated. The paper does not discuss when one might have access to a diffusion prior, its training data, the PDE, but not able to sample the entire observation during inference. Also, if the PDE is known, then one can generate their own training data and diffusion prior so there should never be a "mismatch."
  • The paper assumes that the training data is ground truth--this assumption does not hold in many cases.
  • There are many missing details of the method and the implementation, and without these details (or code), I cannot adequately evaluate some of the results. For example, how long was the diffusion prior trained for? Currently, the results point towards the diffusion prior being poorly trained (Table 1 shows the ""no-physics" model scores a ~30-70% error for the only non-linear PDE).
  • There are no baselines from prior work. I would expect that the paper would include at least one baseline like DiffusionPDE, showing the problem and their proposed solution.
Questions for authors

Can you include one baseline from prior work, e.g., DiffusionPDE, and show the mismatch problem and whether NCPG fixes it?

Reviewer_z8UgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a better approach to calculate the PDE residual loss in diffusion posterior sampling for inverse problems. Its main contribution is: in the linear case, using regression to fit the best-calibrated residual operators from data; and in the non-linear case, using Jacobian and score approximation method. It has moderate improvement among the cases tested.

Critical strengths
  • This paper systematically compares different PDE residual calculation methods, including FD, FEM, spectral, and their method.
Critical weaknesses
  • The preconditioner is not defined. In the text, it is said "we construct a fixed SVD-based preconditioner Pφ from the fitted calibration model ..." and there is no exact definition.
  • In the Bayesian settings of inverse problems, the "guidance strength" aka the confidence/sigma of the observation likelihood is pre-defined by the noise level of observation. In this case sweeping guidance level and compare with the same groundtruth as in Figure 4 doesn't make sense.
  • The paper claims "posterior reconstruction" but the actual task is a single reconstruction, which is flawed in the posterior sampling settings as the solution is not unique. For example, PosteriorBench (https://arxiv.org/abs/2609.20794) provides a systematic approach to evaluate posterior matching.
  • The performance gain is moderate, especially for the NS case. The comparisons are limited to one architecture. There are many physics-informed approaches for inverse problems, such as PIDM (https://arxiv.org/abs/2403.14404), PIDDM (https://arxiv.org/abs/2505.22391), and https://arxiv.org/abs/2504.04375. A more thorough comparison is necessary to justify the contribution.
  • The paper mentions that a problem with the numerical method is the sensitivity to resolution/discretization. It would be better to conduct an ablation study to show the effect of changing resolution on the baselines and NCPG, and even the continuum accuracy.
Questions for authors
  • It's hard to understand the non-linear approach. What is overlapping windows and what is \mathcal{O}? Why use 3 consecutive frames?
  • The calibration is based on data fitting. Could there be OOD issues? What if the calibration data is from a different distribution than the test data?
  • Appendix B.4 defines guidance normalization, which is a non-trivial technique used and also alters the output distribution. It needs more justification.
Reviewer_wA4GOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NCPG to reduce numerical mismatch in physics-informed diffusion posterior sampling by learning a low-rank linear operator or a nonlinear trajectory denoiser from the prior's training data and evaluates reconstruction on Poisson, Helmholtz and Navier–Stokes problems.

Critical strengths
  1. The paper studies a useful problem. Different numerical methods for the same PDE can produce different physics guidance. The linear analysis explains how this mismatch affects the guidance and why high-frequency components can be more sensitive.

  2. The experiments provide controlled comparisons across three PDEs and three numerical methods. The prior and observations are fixed when comparing guidance methods. The evaluation uses 24 test cases with three sampling seeds per case. The clustered bootstrap analysis in Appendix C also supports the improvement over mismatched guidance.

Critical weaknesses
  1. The mechanism behind the improvement is not fully clear. Linear NCPG includes a low-rank projection and an SVD preconditioner. Nonlinear NCPG adds a trajectory denoiser conditioned on the forcing. These components can improve reconstruction through better conditioning or an additional learned prior. The current experiments do not clearly separate these effects from numerical calibration. Normalizing the guidance norms does not resolve this question. More direct evidence is needed to support the mechanism described in Section 4.2.

  2. The matched FD Navier–Stokes baseline is not fully matched in time. Appendix B.1 says the data generator uses Crank–Nicolson for diffusion and forward Euler for the nonlinear update. However, the analytic guidance uses a centered three-frame temporal difference. The baseline therefore matches the spatial method but does not use the generator's full discrete time update. This makes the improvement from 8.7% to 6.2% in Table 1 difficult to interpret. Part of the gain could come from correcting the remaining temporal mismatch.

Questions for authors
  1. Could the authors provide a more direct check of the calibration mechanism? For the linear method, please isolate the effect of the SVD preconditioner through an ablation or an analytic baseline with comparable preconditioning. For the nonlinear method, please show whether the learned correction reduces the reference discrete residual at intermediate sampling states. This would help explain how much of the improvement comes from numerical calibration.

  2. Could the authors repeat the FD Navier–Stokes comparison using a physics residual consistent with the data generator's time update? Please report whether NCPG still improves over this fully matched baseline. If the current baseline is intended to match only the spatial method, this should be stated clearly in the paper and Table 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.