acceptodds
Under review as a conference paper at ICLR 2027

Physics-Informed MeanFlow for PDE-Constrained Inverse Problems

Abstract

Generative models have demonstrated promising results for inverse problems governed by Partial Differential Equations (PDEs). However, current diffusion-based methods make limited progress at early sampling stages: constraints are evaluated on the Tweedie posterior-mean estimate, which at high noise is close to the dataset mean and therefore carries little information about the specific solution. Moreover, accurate reconstruction can require thousands of guided denoising steps, each depending on the previous one. Both effects slow convergence. To address them, we introduce Physics-Informed MeanFlow (PIMF), which fine-tunes a MeanFlow prior with a PDE-residual loss and, at inference, evaluates observation and PDE constraints on the MeanFlow one-step estimate of the clean state rather than on the posterior mean, so that guidance is informative from the first step. PIMF (i) improves accuracy and computational efficiency over diffusion-based solvers on most benchmark tasks, (ii) makes ensemble-based uncertainty estimation affordable, and (iii) recovers alternative physically valid solutions consistent with the same sparse observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative AI, Flow Matching, Mean Flow, Diffusion, Physics-Informed AI, Efficiency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_94zLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Physics-Informed MeanFlow (PIMF) to reconstruct PDE coefficients and solution fields from sparse observations. It evaluates guidance losses on a MeanFlow endpoint estimate. This aims to provide more informative guidance early in reconstruction than a diffusion posterior-mean estimate. Training combines SplitMeanFlow pretraining with PDE-residual fine-tuning. Experiments are conducted over five PDE families and nine reconstruction tasks. PIMF-FMinit achieves the lowest mean target error on six tasks. In addition, ablation studies are provided to examine component contributions, runtime, and the consistency of ensemble candidates.

Critical strengths
  • The paper is well-written and easy to follow.
  • The guidance design addresses a concrete limitation of diffusion-based reconstruction. To evaluate constraints on a MeanFlow endpoint is well motivated. The reduced-budget results suggest that this choice can make sparse-reconstruction more efficient.
  • Since the comparison include diffusion, physics-informed training, flow matching, and distillation baselines, the empirical results demonstrate both benefits and failure cases of the proposed algorithm. The authors also separate training-time physics from inference-time guidance.
Critical weaknesses
  • The contribution of physics-informed training remains difficult to isolate. In particular, Table 2 shows that observation-guided MeanFlow already provides strong performance. Adding PDE guidance worsens both Helmholtz and Poisson inverse problems. The matched-duration controls in Table 8 also show mixed effects. Residual fine-tuning improves Darcy forward reconstruction but worsens Darcy inverse error from 7.86% to 9.33%. Therefore, these results support MeanFlow-based guidance more clearly than the proposed physics-informed training process, which limits the novelty of the approach.
  • The reduced-budget comparison does not separate sampling efficiency from guidance settings. The experiments retain weights selected for 2000 iterations. Algorithm 1 applies guidance updates without explicit time-step scaling. Thus, changing the iteration count affects both the transport discretization and the accumulated guidance. Runtime measurements provide useful insights into Helmholtz forward reconstruction. However, coverage across other tasks is limited, and PIDDM is missing from the corresponding comparison in terms of runtime. These gaps are significant, as computational efficiency is one of the main motivations.
  • The physical constraints do not consistently validate the reconstructed coefficient-solution pairs. For Darcy, the residual pattern differs from that of the solver used to generate the data. This raises concerns that residual minimization may, in some cases, contradict the data distribution. Table 11 also shows that 42% of the Darcy ensemble candidates exhibit a relative inconsistency in the forward solution of more than 10%. In NSNB, the proxy residual value neither enforces incompressibility nor verifies the governing dynamics. It essentially provides no guidance. The authors acknowledge these limitations but significantly restrict the evidence for physical validity across the entire benchmark suite. The most rigorous checks are limited to Poisson and Helmholtz.
Questions for authors
  • Can the authors provide matched-duration, physics-free MeanFlow controls for the remaining PDE families? Do the reduced-budget gains persist without residual fine-tuning? This would clarify which improvements require the proposed physics component.
  • Are guidance updates scaled by the time increment anywhere in the implementation? If not, do the reduced-budget rankings persist after tuning guidance separately for each budget?
  • Do the main rankings remain stable when guidance weights are selected on a disjoint validation subset? In addition, please clarify whether any results were repeated across independent training seeds.
2 comments

We thank the reviewer for the evaluation of our training objective, guidance, and evaluation protocol. We address the questions and weaknesses as follows.

W1+Q1. Isolating the contribution of physics-informed training on Helmholtz and Burgers

The ablations now cover all non-proxy datasets (App. E.4, Table 11). The physics loss is a bit beneficial on Poisson, Darcy forward and Burgers, degrades Darcy inverse and Helmholtz forward, and irrelevant for Helmholtz inverse. The models were retrained for this rebuttal; within each row, the two runs differ only in the physics weight. They are separate runs from the Table 1, so absolute values differ from it. For Burgers, the submitted PIMF model was under-tuned. We keep the submitted Table 1 unchanged. The update would only improve PIMF trained from scratch relative to the baselines, so Table 1 is conservative for our method, and PIMF-FMinit, on which our main claims rest, is not affected. Helmholtz has a higher variance in results.

TaskWithout physics lossWith physics loss
Poisson fwd1.721.66
Poisson inv14.3813.99
Darcy fwd4.444.18
Darcy inv7.869.33
Burgers (from scratch)3.503.43
Helmholtz fwd (from scratch)2.753.02
Helmholtz inv (from scratch)15.2115.20

On Darcy inverse, most of the weakness comes from the guidance estimate rather than from training (Section 7; App. F). With the same weights, the posterior-mean estimate achieves a 4.15% instead of 9.33% on the 1000 test instances (we explain the motivation in our response to Reviewer 1VMP, W1+Q2).

The reduced-budget gains are still there without residual fine-tuning. The table below is Poisson forward, with weights re-tuned per budget, on 50 instances.

N=100N = 100N=200N = 200N=500N = 500
Without physics loss2.152.091.77
PIMF-FMinit2.202.051.86

These gains therefore come mainly from the MeanFlow model, not from the physics loss.

W2+Q2. Are guidance updates scaled by the time increment? Do rankings persist after per-budget tuning?

To the first question, no. Guidance updates are not scaled in any method depending on the time. Every sampler uses a constant guidance weight, and only the transport step is multiplied by the step length (Algorithm 1). Fewer iterations therefore mean less total guidance. The only rescaling is applied when the PDE guidance is added, as in DiffusionPDE, and that value is not fine-tuned (like them, we multiply the observation guidance by 0.1; App. C.1).

We re-tuned all five methods (App. H, Table 16). On Helmholtz forward at N=100N = 100, DiffusionPDE reaches 19.2, PIDM 17.5, FmPDE 3.57, PIMF from scratch 2.20 and PIMF-FMinit 2.22. PIMF-FMinit is first or tied at every budget. The full table is given in our response to Reviewer 1VMP (W2+Q1).

On runtime, PIDDM is now measured (see our response to Reviewer qnSJ, W2+Q2). Timing on Helmholtz is the same in both directions (within 1%), and all 2D tasks share the network and sampler (App. G, Tables 14,15).

Q3. Do rankings hold with weights selected on a disjoint validation subset? Were results repeated across training seeds?

Validation subset. The weights were searched on one test instance (Appendix C.1). Without the selection instances (0 and 1), no entry of Table 1 changes by more than 0.04 percentage points (App. D.1, Table 7). As the answer to reviewer 1VMP, DiffusionPDE (which we use as the baseline) doesn't clearly explain how they choose the hyperparameters.

Seeds. The submission used a single training seed. We re-trained all five PIMF-FMinit models from the end of Stage 1 and tested them on the same 1000 samples (App. D.2, Table 8).

Helm. fwdHelm. invNSNB fwdNSNB invDarcy fwdDarcy invPoisson fwdPoisson invBurgers
Submitted1.7014.512.958.274.189.331.6613.992.36
Re-trained1.7013.762.948.294.209.291.6714.332.34

Seven tasks reproduce, Helmholtz inverse improves, and Poisson inverse worsens, which might need more hyperparameter tuning. We will upload the checkpoints to our GitHub repository.

W3. The physical constraints do not consistently validate the reconstructed pairs

These are the benchmark's residuals, shared with the baselines. On Poisson and Helmholtz, where they are actually a rigorous check, 0–2% of PIMF candidates are inconsistent with the forward solve, against 60–72% for DiffusionPDE. We restrict our physical-validity claims to these two PDEs.

The residual on Darcy, which also guides DiffusionPDE and FmPDE, separates true from mismatched pairs (6.2×6.2\times) but does not directly imply that they are valid.

On NSNB, it is a proxy (App. B), so it does not say much this sense.

In the revised manuscript, we state the Darcy and NSNB limitations (Section 7).

I'd like to thank the authors for their response. The authors address several of my experimental concerns, such as per-budget hyper-parameter tuning, runtime comparisons, and reproducibility. In addition, I appreciate the stronger physical-consistency evidence on Poisson and Helmholtz.

However, the new ablation studies indicate that most of the performance gains may come from MeanFlow rather than physics-informed tuning, which reinforces my concern about the contribution of the physics-informed component. Furthermore, the physical effectiveness of reconstructions on Darcy and NSNB remains insufficiently investigated. Since my major concern about the main contribution of the proposed work remains unresolved, I maintain my original score.

Reviewer_qnSJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Physics-Informed MeanFlow (PIMF) for reconstructing PDE coefficient and solution fields from sparse observations. It combines physics-informed fine-tuning with iterative guidance evaluated on MeanFlow’s one-step clean-state predictions. The method is evaluated on five PDE families through reconstruction comparisons, component ablations, efficiency analysis, and checks of alternative candidate solutions.

Critical strengths
  1. Evaluating constraints on MeanFlow’s clean-state predictions is well motivated. The results show earlier observation-error reduction and good reconstruction accuracy with fewer iterations, particularly on Helmholtz and Poisson. This addresses a practical limitation of diffusion-based reconstruction.
  2. The main comparison covers nine tasks with 1000 test instances per task. PIMF-FMinit achieves the lowest mean target-field error on six tasks. The improvement on Burgers is particularly substantial: the error decreases from 4.84% for the strongest baseline to 2.36%. These results provide useful evidence that the overall method works.
Critical weaknesses
  1. The benefit of the physics-informed component is not fully established. The matched-duration controls in Table 8 are useful, but cover only Poisson and Darcy, not Helmholtz or Burgers. Burgers shows a large improvement but is excluded from the physics ablation. Table 2 also shows that adding PDE guidance can worsen inverse reconstruction. Overall, the evidence supports guided MeanFlow more clearly than it establishes the added value of its physics components.
  2. The speed comparison is incomplete. Appendix F provides useful runtime measurements, but excludes PIDDM, a closely related baseline with a different cost per iteration. In addition, the reduced-budget experiments retain the settings used for 2,000 steps rather than tuning them for smaller budgets. Since efficiency is a central contribution, a runtime–accuracy comparison including PIDDM and budget-appropriate settings is needed to establish the advantage more convincingly.
Questions for authors
  1. Can you provide an equal-training-duration control without the physics loss for Helmholtz or Burgers, keeping the inference procedure unchanged?
  2. Does PIMF retain its advantage over PIDDM when compared using measured end-to-end runtime at similar reconstruction accuracy?
2 comments

We thank the reviewer for the thoughtful evaluation, and in particular for the points regarding guidance and runtime comparisons. We address the concerns.

W1+Q1. Equal-training-duration control without the physics loss for Helmholtz or Burgers

The ablations now cover every PDE except NSNB, whose residual is a proxy (App. E.4, Table 11). The physics loss translates into small gains in Poisson, Darcy forward, and Burgers, hurts on Darcy inverse and Helmholtz forward, and stays irrelevant for Helmholtz inverse. The models were retrained for this rebuttal; within each row, the two runs differ only in the physics weight. They are separate runs from the Table 1, so absolute values differ from it. For Burgers, the submitted PIMF model was under-tuned. We keep the submitted Table 1 unchanged. The update would only improve PIMF trained from scratch relative to the baselines, so Table 1 is conservative for our method, and PIMF-FMinit, on which our main claims rest, is not affected. Helmholtz has a higher variance in results.

TaskWithout physics lossWith physics losspp (Wilcoxon)
Poisson fwd1.721.6610−1110^{-11}
Poisson inv14.3813.9910−11210^{-112}
Darcy fwd4.444.188×10−68\times10^{-6}
Darcy inv7.869.333×10−63\times10^{-6}
Burgers (from scratch)3.503.439×10−59\times10^{-5}
Helmholtz fwd (from scratch)2.753.027×10−237\times10^{-23}
Helmholtz inv (from scratch)15.2115.200.040.04

The Poisson and Darcy rows are PIMF-FMinit, which we retrained without the physics loss for this rebuttal. The Burgers and Helmholtz rows are PIMF trained from scratch, where the control differs from the physics run only in the physics weight.

We agree with the reviewer. The evidence supports guided MeanFlow more clearly than the physics component. The revised manuscript reports the effect of the physics loss per task, including where it is not beneficial.

W1. Table 2: adding PDE guidance can worsen inverse reconstruction

We agree. We report this effect in Section 5.4, while keeping the physics guidance to ensure comparability with the baselines. In inverse recovery, the residual applies a differential operator to an imperfect uu and amplifies its high-frequency error (App. B).

The effect is not specific to PIMF (App. E.2, Table 10). On Helmholtz inverse, removing PDE guidance improves PIMF-FMinit from 14.56 to 12.50, FmPDE from 15.60 to 12.92, and PIDM from 14.11 to 13.45. In the revised manuscript, inference-time PDE guidance is optional and turned off by default on inverse tasks (Section 4.2).

W2+Q2. The runtime comparison excludes PIDDM; does PIMF keep its advantage at similar accuracy?

Good point. Yes, from about 40 s per instance. PIDDM is six times cheaper per step, but it saturates at about 2.5%, and PIMF-FMinit surpasses that accuracy in less wall-clock time. We measured both methods end-to-end on Helmholtz forward (App. G, Table 15). Entries are relative errors in %, with the wall-clock time in parentheses.

Methods / stepError at 30–40 sError at 48–60 sError at 80–122 s
PIDDM0.0343.34 (31 s)2.79 (48 s)2.54 (82 s)
PIMF-FMinit0.2062.22 (39 s)1.95 (60 s)1.73 (122 s)

PIDDM's best result, 2.54% at its full 2000 steps, takes 82 s, whereas PIMF-FMinit reaches 2.22% in 39 s and keeps improving. Below about 30 s, PIDDM is the better choice, because PIMF-FMinit needs at least 100 iterations. In the revised manuscript, PIDDM is added to Tables 14,15, and we explain the trade-off.

PIDDM is run with its paper settings (App. J) on 100 instances and was not re-tuned per budget; on 1000 instances, its full-iteration error is 2.70% (Table 1).

W2. Reduced-budget experiments retain the settings used for 2000 steps

We thank the reviewer for raising this point. We re-tuned every guided method for each budget on held-out instances (App. H, Table 16). Re-tuning helps the baselines the most. The ranking still holds, with the flow-based methods at 2–4% and PIMF-FMinit lowest or tied at every budget (2.22% at 100 iterations, against 3.57% for FmPDE). The full table is given in our response to Reviewer 1VMP (W2+Q1).

Reviewer_qnSJOpenReview ↗

response to author

Thank you for the detailed response and the additional experiments! The runtime concern is largely addressed. The new controls answer my question, but they confirm that the benefit of the physics-informed component is small and task-dependent. Base on this, I would like to raise the rating to 3.

Reviewer_1VMPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Physics-Informed MeanFlow (PIMF) for PDE-constrained inverse problems. Instead of evaluating observation and PDE constraints on the Tweedie posterior-mean estimate, the authors use a MeanFlow projection to obtain a clean-state estimate for guidance. The method also incorporates physics-informed fine-tuning and a staged guidance procedure. Experiments across five PDE families demonstrate improvements in reconstruction accuracy and convergence speed on several tasks, along with applications to uncertainty estimation and alternative solution discovery.

Critical strengths
  1. I find the idea of using the MeanFlow clean-state projection for inverse-problem guidance interesting and well motivated. In particular, exploring an alternative to the Tweedie posterior mean at early sampling stages is an intuitive direction that could potentially benefit other inverse problems as well.

  2. I appreciate the relatively extensive experimental evaluation across five PDE families and both forward and inverse problems. The comparisons against diffusion- and flow-based methods, together with the component ablations and convergence analyses, provide useful evidence of the proposed method's performance. I also appreciate that the authors discuss settings where the physics-informed components do not improve reconstruction.

Critical weaknesses
  1. I think the central motivation is not sufficiently substantiated. PIMF builds on DiffusionPDE's guidance framework by introducing a MeanFlow-based generative prior and physics-informed training. A key claim is that the MeanFlow projection provides a more informative guidance target than the Tweedie posterior mean, particularly at high noise levels. However, beyond the conceptual illustration in Figure 1b, I could not find a direct quantitative comparison of the two estimates or their guidance gradients. I would have liked stronger evidence that this difference is responsible for the improved convergence. Table 2 further shows that observation-only MeanFlow guidance already performs well, while inference-time PDE guidance can degrade inverse reconstruction.

  2. I am not fully convinced by the broader accuracy and computational efficiency claims. In Table 1, PIMF trained from scratch underperforms PIDM on six of the nine reconstruction tasks, while most of the reported improvements come from PIMF-FMinit, which requires additional FM pretraining. Although the proposed method achieves faster convergence on several tasks, Table 9 shows that it is nearly twice as expensive as FmPDE at the same iteration count. Furthermore, the reduced-budget experiments reuse guidance hyperparameters selected for 2,000 iterations, making it unclear whether the same conclusions would hold after budget-specific tuning. I also find the use of a test-set instance for guidance hyperparameter selection concerning.

Questions for authors
  1. Could the authors compare against more recent PDE inverse-problem solvers, particularly FunDPS (Yao et al., NeurIPS 2025) and DDIS (Lin et al., ICLR-W 2026), or clarify the differences that prevent direct comparison? A DiffusionPDE baseline using fewer sampling steps with appropriately retuned guidance weights would also help substantiate the efficiency claims.

  2. Could the authors provide quantitative evidence for the central MeanFlow projection hypothesis? For example, how do the MeanFlow and Tweedie clean-state estimates compare across noise levels in terms of reconstruction error, distance to the dataset mean, and alignment of the resulting measurement-guidance gradients?

2 comments

We thank the reviewer for the thorough evaluation of our guidance estimate and experimental protocol. Below we address the weaknesses and questions.

Q1. Comparison with other baselines

We ran FunDPS as an additional baseline, using the authors' released checkpoints and their settings. We show relative errors in % on 1000 samples.

PDEPIMF-FMinit fwd / invFunDPS fwd / inv
Poisson1.66 / 13.991.95 / 20.30
Helmholtz1.70 / 14.512.23 / 16.30
NSNB2.95 / 8.273.12 / 8.71
Darcy4.18 / 9.332.57 / 5.01

PIMF-FMinit is better on 6 of the 8 tasks. FunDPS does not cover Burgers as it has only one channel. FunDPS uses its default 500 iterations and PIMF-FMinit 2000 (we take the setup from the paper). At 500 iterations, with weights re-tuned for that budget (see W2+Q1), PIMF-FMinit reaches 1.73% on Helmholtz forward and 1.86% on Poisson forward (50 samples), against 2.23% and 1.95% for FunDPS.

For Darcy, we find that with the posterior-mean estimate (we explain the motivation in W1+Q2), PIMF-FMinit reaches 4.15% on Darcy inverse, below FunDPS.

We did not run DDIS for the following reason. Its code was released in April, but it has no checkpoints of trained models, so a comparison would require training, for every PDE, both its coefficient prior and its neural-operator surrogate. It also addresses inverse problems only, with a prior over the coefficient alone, whereas we treat forward and inverse reconstruction with a joint prior. We now discuss DDIS in the related work (Section 6). PIDDM, a recent baseline, is already included in the paper.

W1+Q2. Comparison of the MeanFlow and Tweedie estimates

We thank the reviewer for raising this point, which also improved our own understanding of the method. Both columns below use the same network and only the estimate changes (tested on 100 samples, 2000 iterations; App. F, Table 12). The MeanFlow estimate is significantly better on 6 of the 8 tasks, not significantly different on Darcy forward, and worse on Darcy inverse.

TaskMeanFlow estimatePosterior-mean estimatepp (Wilcoxon)
Poisson fwd1.771.963×10−43\times10^{-4}
Poisson inv14.4016.208×10−178\times10^{-17}
Helmholtz fwd1.702.072×10−92\times10^{-9}
Helmholtz inv14.5615.754×10−184\times10^{-18}
NSNB fwd2.993.283×10−83\times10^{-8}
NSNB inv8.138.374×10−184\times10^{-18}
Darcy fwd4.074.370.640.64
Darcy inv8.784.097×10−187\times10^{-18}

We also measured the quantities requested by the reviewer, using 200 samples and 12 noise levels on Helmholtz, Poisson, and Darcy (App. F).

  • Distance to the dataset mean. The MeanFlow estimate keeps around 0.98 to 1.15 of the true distance at every noise level, whereas the posterior mean shrinks to 0.26 to 0.55 at σ=49\sigma = 49. This is a sign of collapse, which is expected in diffusion models.
  • Reconstruction error. The MeanFlow estimate is worse in L2L^2, as expected, because the posterior mean is L2L^2-optimal.
  • Gradient alignment. The MeanFlow estimate is not better at high noise (cosine similarities of 0.55 versus 0.69 at σ=49\sigma = 49) and is equal at lower σ\sigma.

What differs is the step size. With the MeanFlow estimate, the guidance gradient projected on the true direction is more than two times larger at σ=49\sigma = 49 on Helmholtz and Poisson, and equal below σ=4\sigma = 4. The MeanFlow estimate therefore pushes harder exactly where the posterior mean has collapsed. We will replace "more informative target" with this statement in the revised manuscript.

Darcy inverse is the exception (Section 7). The MeanFlow estimate gives stronger guidance but is a less accurate estimate, so it helps only where the stronger guidance is needed. On Helmholtz and Poisson the posterior mean collapses towards the dataset mean at high noise, and the MeanFlow estimate gives a guidance step more than two times larger. On Darcy the posterior mean collapses much less so the step is at most 1.3×1.3\times larger and only the lower accuracy remains. Darcy is also the only task with a two-valued target (two physical materials) that is computed after applying a threshold, where we believe an accurate estimate matters most.

A MeanFlow network provides both estimates from the same weights and at the same cost, so the estimate can be chosen per task without retraining. To avoid choosing it on test data, we selected it on ten held-out instances. We then evaluated it on all 1000 test instances: 4.15% against 9.33% (better on 988 of 1000). This is below every method in Table 1 (PIDM: 6.47%) and below FunDPS (5.01%).

We also ran the swap at reduced budgets, since the reviewer asks whether the estimate is responsible for the faster convergence. We use the same network on Helmholtz and Poisson forward, with 50 test instances and weights re-tuned per budget and per estimate (App. F).

AuthorsOpenReview ↗

Part 2 (continues right away from Part 1)

Helm. N=100N{=}100Helm. N=200N{=}200Helm. N=500N{=}500Poisson N=100N{=}100Poisson N=200N{=}200Poisson N=500N{=}500
MeanFlow estimate2.221.951.732.202.051.86
Posterior-mean estimate2.202.081.962.151.981.94
pp (Wilcoxon)0.860.032×10−42\times10^{-4}0.340.210.06

Without re-tuning, the MeanFlow estimate degrades far less. At N=200N = 200 we obtain 3.33% against 17.05% on Helmholtz and 3.67% against 11.34% on Poisson. After re-tuning, the posterior-mean estimate recovers most of this gap, and the MeanFlow estimate remains better only on Helmholtz at N=200N = 200 and 500500. We now narrow the claim. The MeanFlow estimate needs far less re-tuning of the guidance weights at reduced budgets, and its advantage after re-tuning is small.

W1.2. Table 2: observation-only guidance is already strong, and PDE guidance can hurt inverse tasks

We mention this effect in Section 5.4 and elaborate on it here. The baselines behave in the same way, so the effect is not specific to PIMF (App. E.2, Table 10). For each method, we selected the schedule on 20 held-out instances and tested on 100 samples. We report relative errors in % on the inverse tasks, with and without PDE guidance.

MethodHelmholtz inv, with PDEHelmholtz inv, withoutPoisson inv, with PDEPoisson inv, without
PIMF-FMinit14.5612.5014.4013.27
FmPDE15.6012.9216.1913.52
PIDM14.1113.4519.5416.05

Removing PDE guidance lowers the error of all three methods on both tasks. We include this comparison in the revised manuscript. In the original submission, we kept PDE guidance active for a fair comparison with DiffusionPDE.

W2. PIMF from scratch loses to PIDM on 6 of 9 tasks, and the gains might come from FM pretraining

The reviewer is correct that PIMF gains considerably from FM pretraining. However, PIDM is also pretrained from the DiffusionPDE checkpoints.

  • Pretrained vs. pretrained: PIDM starts from the DiffusionPDE checkpoint. PIMF-FMinit outperforms it on 7 of 9 tasks.
  • Same FM model, without MeanFlow training: FmPDE is our FM model with guidance and without MeanFlow training. PIMF-FMinit outperforms it on 7 of 9 tasks.
  • From scratch vs. from scratch: PIMF outperforms DiffusionPDE on 7 of 9 tasks.

In terms of cost (A100 GPU hours), PIMF-FMinit requires around 840 to 1000, of which FM pretraining accounts for about 330, while PIDM requires 383 in addition to DiffusionPDE's pretraining. We acknowledge that MeanFlow models take more time to train, particularly with the extra physics loss and step, which we report in the compute cost.

W2. PIMF is nearly twice as expensive as FmPDE per iteration (Table 9)

We thank the reviewer for pointing this out; the reviewer is correct. On average, PIMF takes 0.21 s per iteration and FmPDE 0.09 s. At equal wall-clock time, however, PIMF-FMinit is still more accurate on Helmholtz forward (App. G, Table 15). Entries are relative errors in %, with weights re-tuned per budget on 10 held-out instances and evaluated on 50 test instances ("it." denotes iterations).

Wall-clock per instanceFmPDEPIMF-FMinit
37–39 s3.22 (200 it.)2.22 (100 it.)
60–64 s3.13 (500 it.)1.95 (200 it.)

PIMF-FMinit is better on 46 of the 50 instances in both rows. On Poisson forward the gap is small: 2.20 vs. 2.31 (not significant) and 2.05 vs. 2.20 (p=0.04p = 0.04).

W2+Q1. Reduced budgets reuse the 2000-iteration weights

We re-tuned all five guided methods for each budget (App. H, Table 16). At N=100N = 100 and 200200, the flow-based methods (2–4%) stay far ahead of the diffusion baselines (9–20%), although the gap is smaller than we reported. At N=500N = 500, PIDM closes most of it (3.02 and 3.00). For each budget NN, we scaled the weights by 11, 2000/N\sqrt{2000/N} or 2000/N2000/N, selected the scale on 10 held-out instances, and evaluated on 50 others. We report the forward tasks:

MethodHelm. N=100N{=}100Helm. N=200N{=}200Helm. N=500N{=}500Poisson N=100N{=}100Poisson N=200N{=}200Poisson N=500N{=}500
DiffusionPDE19.213.210.617.910.85.79
PIDM17.59.173.0219.69.723.00
FmPDE3.573.223.132.332.312.20
PIMF (from scratch)2.202.222.282.612.462.63
PIMF-FMinit2.221.951.732.202.051.86

Re-tuning helps DiffusionPDE the most, as expected.

W2. Use of a test-set instance for guidance hyperparameter selection

We recomputed Table 1 without the first two test instances, which include the tuning instance. No entry changes by more than 0.04 percentage points, and the ranking is identical on all nine tasks. DiffusionPDE describes no validation split. The revised manuscript reports the 998-instance table (App. D.1, Table 7).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary to AC

We want to thank all the reviewers and the AC for their time and constructive feedback. All three reviewers recognized the main idea of the paper. We ran extensive new experiments to clarify and address every empirical and methodological concern raised:

  • Reviewer 1VMP: We benchmarked against FunDPS, outperforming it on 6 of 8 tasks. We also tackled one of our main limitations with Darcy, reaching 4.15% on its inverse problem (beating FunDPS's 5.01% and all Table 1 baselines) by leveraging a unique property of MeanFlow: extracting the posterior-mean estimate from the same network without retraining. Measuring both estimates across 12 noise levels reveals some insights from MeanFlow. It avoids collapse toward the dataset mean and yields over 2×2\times larger projected guidance steps at high noise. Separately, PIMF-FMinit is more accurate than FmPDE at equal wall-clock time on Helmholtz forward.
  • Reviewer 94zL: We completed equal-duration physics-loss ablations across all non-proxy PDEs (adding Helmholtz and Burgers), retrained all five PIMF-FMinit models with new seeds to confirm reproducibility, and verified on held-out splits.
  • Reviewer qnSJ (already raised score to 3): We added the requested Helmholtz and Burgers controls and showed that inference-time PDE guidance hurts inverse tasks in, for example, FmPDE and PIDM baselines, not just PIMF. We also added wall-clock comparisons against PIDDM, showing PIMF-FMinit beats PIDDM's saturation accuracy in half the time (2.22% in 39 s vs. 2.54% in 82 s), and finally re-tuned every method per iteration budget.

Summary for the Area Chair. With FunDPS added, multi-seed verification, end-to-end wall-clock benchmarks, and per-budget baseline re-tuning, PIMF-FMinit achieves state-of-the-art accuracy on 6 of 9 tasks out-of-the-box (7 of 9 using its dual-estimate capability) and leads or ties at every reduced budget (N=100,200,500N = 100, 200, 500). Our rebuttal disentangles the mechanisms by importance, establishing MeanFlow endpoint guidance as a fast, tuning-resilient solver for PDE inverse problems while transparently characterizing when physics losses help. We hope the Area Chair and reviewers will consider this strengthened empirical and analytical contribution in their final assessment.

Please find all revisions highlighted in blue in the updated manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.