acceptodds
Under review as a conference paper at ICLR 2027

FROST: Stochastic Prediction-Space Adaptation for Sim-to-Real PDE Forecasting

Abstract

Simulation pretraining followed by real-data fine-tuning is a natural strategy for adapting learned PDE forecasters, such as neural operators, to physical experiments. Yet deterministic adaptation can leave a structured mismatch between forecasts and experimental futures, whose variability reflects unresolved dynamics and measurement noise. To this end, we introduce FROST (Frozen-Operator Stochastic Transport), which freezes an upstream forecaster during correction and learns the conditional distribution of its residual from real measurements. Conditioned on observed history and the base forecast, FROST corrects systematic bias while modeling structured variability around the forecast. Across five RealPDEBench systems, FROST improves long-horizon ensemble-mean accuracy and fair CRPS over the corresponding fine-tuned FNO on every system, with reductions of up to 61% and 75%, respectively, and matches or outperforms every deterministic and generative baseline in mean rollout accuracy on all five systems. Additionally, we find that when applying FROST, the base FNO need not be fine-tuned at all. Together, these results establish stochastic prediction-space correction as a powerful complementary axis of sim-to-real adaptation for PDE forecasting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Sim-to-Real Transfer, Generative Modeling

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_ro2vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the adaptation of neural PDE forecasters from simulation training to real experimental data. The authors propose FROST, an adaptation scheme that freezes the forecaster and trains a corrector to learn the conditional distribution of its residual, given the measured history and the forecast, using stochastic interpolants. On five RealPDEBench systems, the resulting ensemble improves both the relative L2 error of its mean and the fair CRPS over the fine-tuned forecaster.

Critical strengths

Extensive real-data evaluation. The method is evaluated on five real experimental systems, including parameter settings held out from the real training data. FROST improves both accuracy and fair CRPS on seen and held-out parameters, supporting the claim that it is an adaptation method that generalizes across different real-data shifts.

Backbone-agnostic correction. On the cylinder and public foil datasets, FROST is applied to six frozen forecasters, including task-specific architectures and one pretrained foundation model, and improves all of them. This supports FROST as a general post-hoc module that does not require a specific architecture.

Critical weaknesses

Weak calibration evidence. Headline results use only K=4 ensemble members, and the paper's own calibration metrics (spread/skill ratios and Gaussian coverage) indicate strong under-dispersion. Moreover, the main probabilistic metric, the fair CRPS, is a marginal score and therefore ignores the spatio-temporal dependence that the paper itself motivates. This makes it unclear whether FROST learns a calibrated residual distribution.

Weakly supported sim-to-real framing. A forecaster trained only on real data and a weak simulation-only forecaster both reach, once corrected, almost the same accuracy as the simulation-pretrained and fine-tuned one. The corrected error barely depends on the base forecaster, and no generator of equal capacity trained on the history alone is reported on the main systems. We therefore cannot assess whether the gains come from a large corrector that compensates for weak forecasts, rather than from correcting a sim-to-real mismatch.

Questions for authors
  1. Can the authors report results on the main systems for a generator of the same capacity, conditioned only on the history (without the forecast) and trained with the same budget as FROST?

  2. Can the authors compare FROST with other probabilistic correctors, such as score-based diffusion or a CRPS-trained corrector?

  3. Can the authors report evaluations with more ensemble members and with dependence-sensitive scores such as the Variogram Score?

  4. Can the authors clarify whether FROST is a general correction method, or what specifically distinguishes it for sim-to-real adaptation?

7 comments
AuthorsOpenReview ↗

Response to Reviewer ro2v

Thank you for highlighting our thorough experimentation, generalization to held-out physical parameters, and the ability to correct multiple forecasters!

W1. Weak calibration evidence. Headline results use only K=4 ensemble members, and the paper's own calibration metrics (spread/skill ratios and Gaussian coverage) indicate strong under-dispersion. Moreover, the main probabilistic metric, the fair CRPS, is a marginal score and therefore ignores the spatio-temporal dependence that the paper itself motivates. This makes it unclear whether FROST learns a calibrated residual distribution.

We believe the under-dispersion comes from the sim-to-real setting itself. The corrector is trained on a few hundred measured trajectories, and with this little data and few sampling steps the flow sampler places its members close to the conditional mean. Consistent with this, increasing the steps from six to thirty raises FROST's spread/skill on all five systems without retraining.

Because the mismatch is mostly one of scale, FROST is compatible with a common post-hoc calibration technique (Johnson and Bowler, 2009; Van Schaeybroeck and Vannitsem, 2015), which rescales each ensemble member's deviation from the ensemble mean using a validation-set-derived factor, fitted for each system and seed on validation windows of the training recordings. Once calibrated, six-step FROST's spread/skill is 0.75–0.98 on controlled cylinder, FSI, public foil and combustion and 0.96 on cylinder without the outlier recording (Table R1). The result is robust to the ensemble size: from 4 to 16 members, rollout error improves by 1.1–4.5% at six and thirty steps, and the calibrated spread/skill is the same (Table R1).

Table R1. Rollout relative L2 / spread/skill (ideal 1), six steps. Cylinder on all 320 windows, with spread/skill on the 314 windows without the outlier recording in brackets. (supplementary Table S4)

MethodCylinderControlled cylinderFSIPublic foilCombustion
FROST, 4 members0.1571 / 0.12 [0.40]0.1490 / 0.460.2134 / 0.340.0298 / 0.510.7750 / 0.35
FROST, 4 members + calibration0.1571 / 0.28 [0.96]0.1490 / 0.970.2134 / 0.750.0298 / 0.980.7750 / 0.84
FROST, 16 members0.1549 / 0.11 [0.37]0.1459 / 0.430.2105 / 0.320.0289 / 0.480.7664 / 0.32
FROST, 16 members + calibration0.1549 / 0.28 [0.97]0.1459 / 0.960.2105 / 0.750.0289 / 0.990.7664 / 0.83

We also evaluated the order-0.5 variogram score over spatial, temporal and space–time pairs of grid points. It is a proper multivariate score: it compares how the differences between pairs of values vary across members with how they vary in the measurement, so an ensemble with the right marginals but the wrong correlations is penalised, which fair CRPS cannot detect.

Table R2. Order-0.5 variogram score in units of 10⁻³ (lower is better), four members unless stated, mean over three seeds; cylinder on all 320 windows. Bold marks the lowest four-member value; the score improves with the member count, so the 16-member row is shown for reference. (supplementary Table S5)

MethodCylinderControlled cylinderFSIPublic foilCombustion
Fine-tuned FNO alone6.282.784.282.107.63
Gaussian noise around the FNO forecast8.403.764.955.1815.99
MC dropout on the FNO18.4214.125.8615.268.03
PhysGuard8.482.824.222.147.88
FROST - EDM residual diffusion, 11 calls5.443.323.841.236.88
FROST, 6 steps5.002.063.471.037.07
FROST, 30 steps4.871.953.390.976.89
FROST, 30 steps, 16 members4.681.723.150.876.55

Thirty-step FROST has the lowest four-member variogram score on cylinder, controlled cylinder, FSI and public foil, and is level with EDM diffusion on combustion (6.89 against 6.88). It lowers the score of the forecaster it corrects on all five systems, from 2.78 to 1.95 on controlled cylinder and from 2.10 to 0.97 on public foil, and with 16 members it is lower again on every system. As a direct test of temporal dependence, shuffling member identities between successive rollout blocks, which keeps every block's marginals, worsens FROST's cross-block variogram score on four of five systems. FROST's members therefore carry dependence across space and time beyond what fair CRPS measures.

W2. Weakly supported sim-to-real framing. A forecaster trained only on real data and a weak simulation-only forecaster both reach, once corrected, almost the same accuracy as the simulation-pretrained and fine-tuned one. The corrected error barely depends on the base forecaster, and no generator of equal capacity trained on the history alone is reported on the main systems. We therefore cannot assess whether the gains come from a large corrector that compensates for weak forecasts, rather than from correcting a sim-to-real mismatch.

FROST is motivated by the analysis in Section 2: a forecaster trained by regression predicts the conditional mean, so it cannot represent the stochastic part of real measurements, and its errors on real data are systematic. Table 2 shows that this gap appears for every forecaster architecture we tested and comes from the regression objective rather than from the training data mixture. The corrected errors are close across forecasters because the corrector is trained on an ensemble of forecasters and models what all of them miss.

To isolate the role of the forecast, we trained a history-only generator with the same network, data and training budget as FROST. It generates the future field directly from the history, while FROST uses the time-resolved forecast as a condition and generates a residual around it. Both use four members and six corrector calls per member per block. On all five systems, FROST has 1.5–4.2× lower rollout error and 1.3–6.5× lower fair CRPS than the history-only generator.

Table R3. History-only generator against FROST: the same network, data and training budget, four members and six corrector calls per member per block; FROST uses the time-resolved forecast as a condition, the history-only generator does not. Cylinder on all 320 windows. (supplementary Table S6)

SystemHistory only: rollout L2FROST: rollout L2RatioHistory only: fair CRPSFROST: fair CRPSRatio
Cylinder0.38750.15712.47×0.047080.017722.66×
Controlled cylinder0.62030.14904.16×0.048930.008265.92×
FSI0.51200.21342.40×0.038250.012643.03×
Public foil0.07740.02982.60×0.022130.003396.53×
Combustion1.13950.77501.47×0.009290.006941.34×

To test whether a large corrector compensates for weak forecasts, we reduced the corrector from 41 M to 10 M parameters. Rollout error changes by at most 1.6% and fair CRPS by at most 2.4% on all five systems, at six and at thirty steps (Table R4). On cylinder we went further, and every corrector from 0.6 M to 163 M keeps the six-step rollout error below the FNO's.

Table R4. The 41 M corrector of the paper against a 10 M corrector of the same architecture; rollout relative L2 / fair CRPS, four members (supplementary Table S7)

SystemSteps41 M10 MChange in L2Change in CRPS
Cylinder60.1571 / 0.017720.1579 / 0.01810+0.5%+2.1%
300.1535 / 0.016580.1549 / 0.01698+0.9%+2.4%
Controlled cylinder60.1490 / 0.008260.1500 / 0.00827+0.7%+0.1%
300.1478 / 0.007290.1502 / 0.00740+1.6%+1.4%
FSI60.2134 / 0.012640.2149 / 0.01278+0.7%+1.1%
300.2147 / 0.012080.2162 / 0.01221+0.7%+1.1%
Public foil60.0298 / 0.003390.0296 / 0.00331−0.6%−2.4%
300.0298 / 0.003160.0300 / 0.00315+0.6%−0.4%
Combustion60.7750 / 0.006940.7740 / 0.00688−0.1%−0.8%
300.7722 / 0.006490.7733 / 0.00645+0.1%−0.5%

Q1. Can the authors report results on the main systems for a generator of the same capacity, conditioned only on the history (without the forecast) and trained with the same budget as FROST?

Yes. Please see our response to W2: with the same capacity and training budget, FROST has 1.5–4.2× lower rollout error and 1.3–6.5× lower fair CRPS than the history-only generator on all five systems.

Q2. Can the authors compare FROST with other probabilistic correctors, such as score-based diffusion or a CRPS-trained corrector?

We trained two additional probabilistic correctors with the same data. First, an EDM residual-diffusion corrector with FROST's backbone, residual target, and history-and-forecast conditioning. Six-call FROST has lower rollout error and CRPS than seven-call EDM on all five systems, and at eleven calls each FROST has the lower rollout error on all five.

Table R5. EDM residual corrector against FROST, with the same network, residual target, measured-history and time-resolved-forecast conditioning, training data and random seeds. Mean rollout relative L2 and fair CRPS (lower is better), four members, three seeds; each call is a corrector evaluation per member per prediction block; cylinder on all 320 windows. (supplementary Table S8)

SystemFROST, 6 callsFROST-EDM, 7 callsFROST, 11 callsFROST-EDM, 11 callsFROST-EDM, 19 calls
(a) Rollout relative L2 ↓
Cylinder0.15710.21310.15480.16580.1591
Controlled cylinder0.14900.33540.14720.20440.1641
FSI0.21340.26420.21360.22510.2196
Public foil0.02980.05960.02990.03280.0310
Combustion0.77501.18630.77110.82010.7828
(b) Fair CRPS ↓
Cylinder0.017720.020600.017170.017080.01700
Controlled cylinder0.008260.014760.007740.008860.00764
FSI0.012640.013930.012300.011940.01199
Public foil0.003390.006430.003270.003590.00358
Combustion0.006940.012580.006740.006190.00614

Second, we trained a one-call corrector directly under fair CRPS on cylinder, with the same network, history-and-forecast conditioning, training split and update count as FROST (three seeds). It generates each member in a single forward pass from noise and is trained on the fair CRPS of four samples per training example.

Table R6. Cylinder, four members, mean over three seeds, as trained (no calibration). All 320 test windows. (supplementary Table S9)

MethodRollout L2 ↓Fair CRPS ↓
Fine-tuned FNO0.16320.02201
CRPS-trained corrector0.16160.01791
FROST, 6 steps0.15710.01772
FROST, 30 steps0.15350.01658

Training directly on fair CRPS does not close the gap. FROST is ahead of it on both.

Q3. Can the authors report evaluations with more ensemble members and with dependence-sensitive scores such as the Variogram Score?

Yes. Please see our response to W1. From 4 to 16 members, fair CRPS changes by at most 0.2%, spread/skill by at most 0.04, and ensemble-mean rollout error improves by 1.1–4.5% on the five systems at six and thirty steps; Table R1 adds the calibrated 16-member results. On the order-0.5 variogram score, thirty-step FROST is lowest on cylinder, controlled cylinder, FSI and public foil and level with EDM diffusion on combustion (Table R2), and shuffling members between blocks worsens it on four systems.

Q4. Can the authors clarify whether FROST is a general correction method, or what specifically distinguishes it for sim-to-real adaptation?

FROST is a general forecast-conditioned correction interface that we evaluate primarily for sim-to-real PDE forecasting, where a simulation-trained operator must be adapted using scarce, partial, and noisy measurements. It learns a distribution over discrepancies between observed future states and the frozen forecaster's time-resolved predictions.

Thank you again for taking the time to review our paper and for your valuable feedback! We're happy to provide any further clarification.

References

  • Johnson, C., and Bowler, N. (2009). On the reliability and calibration of ensemble forecasts. Monthly Weather Review, 137(5), 1717–1720.
  • Van Schaeybroeck, B., and Vannitsem, S. (2015). Ensemble post-processing using member-by-member approaches: theoretical aspects. Quarterly Journal of the Royal Meteorological Society, 141(688), 807–818.
Reviewer_pGMxOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper sets out to improve sim-to-real transfer performance in PDE surrogates. They introduce a frozen-operator stochastic transport based methods where on top of the base predictor, the model to correct the residuals. They use realpdebench as the dataset and show that frost improves upon the base model to capture residual fluctuations in residuals.

Critical strengths
  1. I think this is a very timely question to ask how to improve sim2real generalization in PDE based surrogates. The field needs to broadly move towards these types of studies.

  2. The baseline and dataset coverage is broad and appropriate.

  3. I also really appreciate the author's efforts on documenting a lot of the details of their experiments.

Critical weaknesses
  1. Unclear advantage over probabilistic approaches: The comparison with probabilistic retrofitting in Appendix D.5 is interesting, but actually default FROST method gives worse CRPS on 3/5 systems. Also I dont understand how the comparison between Diaconu et al. and the proposed method is actually fair. In page 4, the authors say "In the main experiments, Gθ is pretrained on simulation and fine-tuned on real measurements before being frozen". Then in the crps comparison in Appendix D.5, page 33 table 19 caption says "The retrofit adapts the same simulation pretrained FNO on the same real split, under the fair CRPS". So the CRPS retrofitting is not being done on a simulator that had seen both sim + real finetuning? I dont see how this makes the comparison fair then because Diaconu et al. explicitly states that CRPS method really gives gains when one starts CRPS tuning on top of an already deterministic checkpoint, which in this case would need to have a deterministic checkpoint that some sim + real finetuning. I might be mis-understanding what exactly has been done but I would really appreciate better clarity on this comparison?

  2. While FROST seems to improve CRPS in some cases, the predicted ensembles are very underdispersed. Example, table 29 has values like 0.116 for Cylinder compared to ideal value of 1. I feel like this is a significant limitation which, surprisingly is not really described well in the main text at all, and I had to dig deep into the appendix to find spread-to-skill values.

Questions for authors
  1. Could you please explain or correct by criticisms 1 and 2? I am happy to change my mind and go from reject to accept if you can clarify my questions. Can the authors also evaluate FROST on some of the benchmarks presented in Diaconu et al. (2026), starting from the same deterministic checkpoints and comparing CRPS, spread–skill ratios, and computational cost? I get that won't be sim-2-real but I would still like to see as a direct comparison against what was published in the CRPS paper. This is basically my main request.

Minor point:

  1. I feel like the related work can be expanded slightly, e.g. Mukhopadhyay et al. "Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence". Although the setting is different from FROST, I think it provides some useful context for the broader question of sim2real generalization in scientific ML, and could be briefly mentioned in the introduction or related work.
5 comments
AuthorsOpenReview ↗

Response to Reviewer pGMx

Thank you for recognizing the importance of sim-to-real generalization, the breadth and depth of our baselines and datasets, and our experimental documentation!

W1. Unclear advantage over probabilistic approaches: The comparison with probabilistic retrofitting in Appendix D.5 is interesting, but actually default FROST method gives worse CRPS on 3/5 systems. Also I dont understand how the comparison between Diaconu et al. and the proposed method is actually fair...because Diaconu et al. explicitly states that CRPS method really gives gains when one starts CRPS tuning on top of an already deterministic checkpoint, which in this case would need to have a deterministic checkpoint that some sim + real finetuning.

Thank you for pointing this out! Appendix D.5 reports FROST's uncalibrated CRPS, while probabilistic retrofitting (PR) is trained on fair CRPS and is therefore calibrated during training. We can apply a common post-hoc calibration technique (Johnson and Bowler, 2009; Van Schaeybroeck and Vannitsem, 2015), which rescales each ensemble member's deviation from the ensemble mean using a validation-set-derived factor.

Notably, calibration significantly improves both spread/skill and CRPS while maintaining the L2 advantage for FROST but has lesser effect for PR, because it is already trained on fair CRPS.

We also show probabilistic retrofitting on top of both real finetuned and simulation only FNO and observe that the choice of the starting deterministic checkpoint does not make a consistent difference in the sim-to-real setting.

Table P1. Fair CRPS (lower is better), four members, mean over three seeds. (supplementary Table S1)

MethodCylinderControlled cylinderFSIPublic foilCombustion
FROST, 6 steps0.017720.008260.012640.003390.00694
FROST, 6 steps + calibration0.015990.006930.010710.002920.00611
PR, simulation-pretrained start (Appendix D.5)0.016360.009100.012080.013040.00568
PR, simulation-pretrained start + calibration0.016440.009200.011440.012340.00570
PR, fine-tuned start0.025940.008730.012370.011770.00564
PR, fine-tuned start + calibration0.025850.008570.011250.010950.00565

With calibration, six-step FROST has the lowest CRPS on cylinder, controlled cylinder, FSI and public foil; PR has the lower CRPS on combustion. Calibration lowers FROST's CRPS by 10–16% and changes PR's by at most 9%, and by at most 2% on combustion, controlled cylinder and cylinder, because PR's training objective already calibrates its spread. The choice of the starting deterministic checkpoint changes PR's calibrated CRPS by at most 11% on controlled cylinder, FSI, public foil and combustion, but the effect is inconsistent.

The methods can also be combined to achieve better results! PR updates the forecaster's internal parameters and introduces stochasticity, whereas FROST leaves the selected forecaster frozen and learns a separate forecast-conditioned residual distribution from histories and predictions.

W2. While FROST seems to improve CRPS in some cases, the predicted ensembles are very underdispersed. Example, table 29 has values like 0.116 for Cylinder compared to ideal value of 1. I feel like this is a significant limitation which, surprisingly is not really described well in the main text at all, and I had to dig deep into the appendix to find spread-to-skill values.

After the post processing calibration stated above, FROST achieves a better spread-to-skill. We will include the comparison in the main text.

Table P2. Spread/skill (ideal 1), four members, mean over three seeds. FROST uses six steps. Cylinder: all 320 windows, and in brackets the 314 windows without the outlier recording. (supplementary Table S2)

MethodCylinderControlled cylinderFSIPublic foilCombustion
FROST, 6 steps0.12 [0.40]0.460.340.510.35
FROST, 6 steps + calibration0.28 [0.96]0.970.750.980.84

Additionally, we observe that the especially low cylinder value is driven in part by one recording among the 92 cylinder recordings, at Reynolds number 3656. Its measured mean velocity is 7.7 times and its fluctuation amplitude 16 times those of neighboring recordings (supplementary Figures S1–S2). Removing its six windows changes FROST on calibrated cylinder spread/skill from 0.28 to 0.96 and uncalibrated cylinder spread/skill from 0.12 to 0.40.

Q1. Could you please explain or correct by criticisms 1 and 2? I am happy to change my mind and go from reject to accept if you can clarify my questions. Can the authors also evaluate FROST on some of the benchmarks presented in Diaconu et al. (2026), starting from the same deterministic checkpoints and comparing CRPS, spread–skill ratios, and computational cost? I get that won't be sim-2-real but I would still like to see as a direct comparison against what was published in the CRPS paper. This is basically my main request.

We compare directly to probabilistic retrofitting on Rayleigh–Bénard, from the same public deterministic Lola checkpoint used by Diaconu et al. and with the benchmark's evaluator and 16 members. FROST corrects the released autoencoder's latent representation while the deterministic forecaster and the autoencoder stay frozen.

FROST at six steps reduces fair CRPS from 0.0560 to 0.0528 and VRMSE from 0.6427 to 0.6272 against PR trained with the published recipe, and its spread/skill is 0.986 without any calibration. It takes 12% more time per trajectory at inference, and the corrector trains in 32 minutes, against 9 hours for our reproduction of PR. Paired over the same 175 trajectories, FROST's CRPS and VRMSE are lower on every seed, with 95% intervals excluding zero. Thirty steps lower CRPS further, to 0.0520.

Table P3. Rayleigh–Bénard, public Lola checkpoint, benchmark evaluator, 16 members, 175 test trajectories. Medians over trajectories; parentheses give the benchmark's bootstrap interval and ± the standard deviation over three corrector seeds. Inference times are on one H200. (supplementary Table S3)

MethodFair CRPS ↓VRMSE ↓Spread/skillTraining timeInference, s / trajectory
Deterministic checkpoint0.1014 (0.0939, 0.1093)0.8026 (0.7896, 0.8210)0.122–3.1 (1 member)
PR, published0.055 (0.051, 0.060)0.632 (0.621, 0.652)0.957about 12 h on one H100, as reported–
PR, reproduced with the published recipe0.0560 (0.0505, 0.0589)0.6427 (0.6283, 0.6550)0.9349 h14.7
FROST, 6 steps0.0528 ± 0.00040.6272 ± 0.00120.98632 min16.4
FROST, 30 steps0.0520 ± 0.00050.6252 ± 0.00580.991same corrector24.2
FROST on top of PR, 6 steps0.0530 ± 0.00040.6335 ± 0.00281.067PR + 32 min16.5

Minor point. I feel like the related work can be expanded slightly, e.g. Mukhopadhyay et al. "Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence". Although the setting is different from FROST, I think it provides some useful context for the broader question of sim2real generalization in scientific ML, and could be briefly mentioned in the introduction or related work.

Thank you so much for suggesting Mukhopadhyay et al. (2026), Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence. Their simulation-adapted model transfers to laboratory turbulence without training on experimental trajectories. FROST studies the complementary setting in which measured trajectories are available for adapting a frozen forecaster through a learned correction. We will include this connection in the related work of the revised manuscript!

References

  • Diaconu, C., Cranmer, M., Turner, R. E., Marwah, T., and Mukhopadhyay, P. (2026). Probabilistic retrofitting of learned simulators. International Conference on Machine Learning.
  • Johnson, C., and Bowler, N. (2009). On the reliability and calibration of ensemble forecasts. Monthly Weather Review, 137(5), 1717–1720.
  • Mukhopadhyay, P., et al. (2026). Emergent transfer of a physics foundation model from simulation to laboratory turbulence. arXiv:2606.01470.
  • Van Schaeybroeck, B., and Vannitsem, S. (2015). Ensemble post-processing using member-by-member approaches: theoretical aspects. Quarterly Journal of the Royal Meteorological Society, 141(688), 807–818.

Thank you again for your thoughtful review and valuable feedback! We just wanted to kindly follow up on our response and let you know that we're happy to provide any further clarification!

Reviewer_LZRBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FROST (Frozen-Operator Stochastic Transport), a framework for sim-to-real adaptation of learned PDE forecasters through stochastic residual correction. Instead of relying exclusively on parameter-space fine-tuning, FROST freezes a pretrained or fine-tuned neural operator and learns a conditional distribution over the discrepancy between its predictions and real experimental measurements.

The residual distribution is modeled using stochastic interpolants, conditioned on both the observed history and the deterministic forecast. At inference, FROST generates corrected trajectories through ODE or SDE sampling, with each ensemble member evolving autoregressively using its own generated history.

The authors evaluate FROST on five RealPDEBench systems, demonstrating improvements in long-horizon prediction accuracy and fair CRPS relative to conventional fine-tuning. The paper also investigates residual structure, probabilistic calibration, adaptation across different forecaster architectures, and generalization to held-out physical parameters.

Critical strengths
  1. Strong empirical validation on real experimental PDE forecasting tasks.

A major strength of this work is its extensive evaluation on five real-world physical systems, rather than relying exclusively on synthetic PDE benchmarks.

FROST consistently improves the ensemble-mean rollout accuracy and fair CRPS of the corresponding fine-tuned FNO across all five systems. The improvement is particularly substantial on the public foil dataset, where the rollout relative L2 error decreases from approximately 0.075 to 0.030.

The evaluation also includes different forecaster architectures, held-out physical parameters, data-efficiency experiments, and multiple correction baselines. The use of multiple training seeds and trajectory-level bootstrap confidence intervals further strengthens the credibility of the reported improvements.

  1. A flexible correction framework supported by informative ablations.

FROST separates deterministic prediction from stochastic correction, allowing the same general procedure to be applied to different pretrained forecasters without modifying their internal architectures.

The experiments show that the method improves six different backbone architectures on the cylinder dataset. Furthermore, a corrector trained on a mixture of forecasters successfully transfers to a held-out CNO architecture.

The finding that FROST can improve a simulation-only forecaster without fine-tuning its parameters on real data is particularly interesting, suggesting that prediction-space adaptation can complement or partially replace conventional parameter-space adaptation.

The matched comparisons against deterministic residual regression, Gaussian perturbations, and PDE-Refiner-style correction also provide useful evidence that learning structured residual distributions is more effective than simply adding noise or applying deterministic corrections.

Critical weaknesses
  1. Limited methodological novelty relative to existing generative residual correction approaches.

Although the application to real experimental PDE forecasting is valuable, the central methodological components are closely related to existing work.

The general strategy of combining a deterministic predictor with a generative residual model has already been explored in residual corrective diffusion models (Mardani et al., 2025). Related work has also combined neural operators with diffusion-based correction to improve spectral fidelity and autoregressive forecasting (Oommen et al., 2025), while residual-augmented flow-matching approaches have been investigated for probabilistic PDE modeling (Bhola and Duraisamy, 2025).

FROST adopts an established stochastic-interpolant formulation and applies it to conditional residual generation around a frozen operator. Its principal novelty therefore appears to lie in the sim-to-real formulation, conditioning strategy, and experimental validation rather than in a fundamentally new generative algorithm.

  1. The practical trade-off between probabilistic calibration, physical fidelity, and computational cost remains unclear.

FROST achieves consistent improvements in fair CRPS, but its predictive ensembles remain substantially underdispersed.

For example, on cylinder, the main six-step ODE configuration obtains a spread-skill ratio of approximately 0.116, far below the ideal value of one. Increasing the sampling budget to 30 SDE steps improves this ratio to approximately 0.205, but substantial underdispersion remains. Similar limitations appear on several other systems.

Furthermore, improvements in fluctuation statistics are not uniform. Although FROST significantly improves fluctuation-energy error on cylinder, controlled cylinder, and FSI, the corresponding error increases on public foil and combustion. This suggests that improved point accuracy and marginal probabilistic scores do not necessarily imply uniformly better physical trajectory fidelity.

These limitations are particularly relevant given the additional computational cost. The default configuration uses four ensemble members and six corrector evaluations per member per prediction block, requiring 240 corrector evaluations over a ten-block rollout, in addition to operator evaluations. The stochastic configuration increases this cost further.

Questions for authors

Q1. Novelty and comparison with closely related residual generative approaches.

Could the authors clarify the methodological differences between FROST and existing conditional residual diffusion approaches, particularly Mardani et al. (2025) and Oommen et al. (2025)?

Would a conditional residual diffusion baseline, trained with the same data, backbone, conditioning information, and comparable sampling budget, achieve similar results? This comparison would help establish whether the contribution is primarily the sim-to-real application or whether the stochastic-interpolant formulation offers a distinct practical advantage.

Q2. Probabilistic calibration and joint trajectory fidelity.

The reported fair CRPS improves consistently, but spread-skill ratios remain far below one, even with the more expensive SDE sampler.

What are the primary causes of this persistent underdispersion? Could the authors provide additional calibration diagnostics across rollout horizons and held-out physical parameters?

Since fair CRPS evaluates coordinate-wise marginal distributions, an assessment of temporal dependencies or a multivariate proper scoring rule would also help determine whether the generated trajectories represent the desired joint predictive distribution.

3 comments
AuthorsOpenReview ↗

Response to reviewer LZRB

Thank you for highlighting FROST's strong real-data validation, informative ablations, and flexibility across forecasters!

W1. Limited methodological novelty relative to existing generative residual correction approaches.

FROST's contribution is forecast-conditioned adaptation of frozen PDE forecasters to real measurements, not the invention of a new generative transport algorithm. The submitted related work situates this design beside Mardani et al. (2025)'s atmospheric residual diffusion, Oommen et al. (2025)'s neural-operator diffusion correction, and Bhola and Duraisamy (2025)'s residual transport between fidelity levels. Our empirical setting is scarce, partial, noisy experimental measurements, with a simulation-trained operator that remains frozen during correction, even when it has been fine-tuned previously on the same measured system. The residual model uses the observation history and the time-resolved forecast as inputs.

Our evaluation shows that FROST improves six distinct cylinder forecasters; a corrector trained on five forecasters improves an unseen CNO from 0.2239 to 0.1709 without retraining; and a simulation-only FNO's error drops from 0.4600 to 0.1565 through correction alone (paper Tables 2–3 and 5).

W2. FROST achieves consistent improvements in fair CRPS, but...underdispersed...Furthermore, improvements in fluctuation statistics are not uniform...additional computational cost.

Dispersion. The cylinder value of 0.116 comes mostly from one recording among the 92 cylinder recordings, at Reynolds number 3656, whose measured mean velocity is 7.7 times and fluctuation amplitude 16 times those of neighboring recordings. Without its six test windows, six-step spread/skill is 0.40 instead of 0.116, and thirty-step spread/skill is 0.66 instead of 0.184 (supplementary Figures S1–S2).

Predicting the spread of a real system from a few hundred measured trajectories is hard: a deep ensemble of three FNOs reaches only 0.08–0.34. With limited data and few sampling steps, the flow sampler places its members close to the conditional mean, and more steps raise spread/skill on all five systems without retraining. Because the remaining mismatch is mostly one of scale, a common post-hoc calibration, which multiplies each member's deviation from the ensemble mean by one factor fitted on validation windows of the training recordings, brings six-step FROST's spread/skill to 0.75–0.98 and lowers its CRPS by 10–16%. On the Rayleigh–Bénard benchmark, where the training set is larger, FROST reaches a spread/skill of 0.986 at six steps with no calibration.

Table L1. Spread/skill , four members, mean over three seeds. Cylinder: in brackets is with the outlier recording removed. (supplementary Table S2)

MethodCylinderControlled cylinderFSIPublic foilCombustion
FROST, 6 steps0.12 [0.40]0.460.340.510.35
FROST, 6 steps + calibration0.28 [0.96]0.970.750.980.84

Physical fidelity. We added equal-weight-per-window measurements of fluctuation-energy error and fluctuation amplitude. On this aggregation, FROST's ensemble members improve both metrics against the fine-tuned FNO on all five systems, with paired trajectory-bootstrap intervals excluding zero. The originally submitted fluctuation-energy statistic instead pools windows, giving high-energy windows greater influence. We will label both aggregations explicitly in the revision.

Table L2. FROST (six steps) minus the fine-tuned FNO. Negative favours FROST. (supplementary Table S10)

SystemΔ rollout rel. L2Δ fair CRPSΔ fluctuation-energy errorΔ |RMS ratio − 1|
Cylinder−0.0061 [−0.0098, −0.0031]−0.0043 [−0.0070, −0.0030]−0.041 [−0.072, −0.017]−0.020 [−0.033, −0.010]
Controlled cylinder−0.0239 [−0.0298, −0.0183]−0.0045 [−0.0050, −0.0041]−0.086 [−0.106, −0.067]−0.049 [−0.060, −0.039]
FSI−0.0140 [−0.0169, −0.0111]−0.0051 [−0.0052, −0.0050]−0.023 [−0.029, −0.016]−0.013 [−0.016, −0.009]
Public foil−0.0456 [−0.0476, −0.0435]−0.0102 [−0.0118, −0.0090]−1.846 [−2.486, −1.293]−0.500 [−0.636, −0.379]
Combustion−0.0168 [−0.0319, −0.0013]−0.0022 [−0.0026, −0.0018]−0.070 [−0.114, −0.029]−0.087 [−0.142, −0.038]

Inference cost. We measured how accuracy scales with the number of corrector calls and the corrector size and the training objective. FROST offers a tunable accuracy–cost frontier, with gains largely saturating near 10 M corrector parameters. On cylinder's 314 test windows without the outlier recording, a 10 M MeanFlow corrector uses one call per member per block and takes 0.49 s per window on one RTX PRO 6000, versus 0.28 s for the fine-tuned FNO, while reducing rollout error from 0.1513 to 0.1463 and calibrated CRPS by 38% (0.01185 to 0.00738). Please see supplementary Figure S3 for more details.

Q1. Novelty and comparison with closely related residual generative approaches...Would a conditional residual diffusion baseline, trained with the same data, backbone... achieve similar results?

We do not propose a new diffusion formulation. The methodological distinction is the use of an existing, frozen simulation-trained forecaster as the time-resolved conditioning signal for a residual distribution learned from limited experimental observations (please see our response to W1). Because FROST uses only the forecaster's outputs, it never touches the forecaster's weights or hidden layers and the forecaster can be a black box.

To test whether the stochastic-interpolant transport is the better choice for this task, we trained an EDM residual corrector with the same network, residual target, measured-history and time-resolved-forecast conditioning, training data, and random seeds as FROST. The table reports mean rollout relative L2 and fair CRPS (lower is better) with four ensemble members over three seeds. Six-call FROST has lower rollout error and CRPS than seven-call EDM on every system.

Table L3. EDM residual corrector against FROST (supplementary Table S8)

SystemFROST, 6 callsEDM, 7 callsFROST, 11 callsEDM, 11 callsEDM, 19 calls
(a) Rollout relative L2 ↓
Cylinder0.15710.21310.15480.16580.1591
Controlled cylinder0.14900.33540.14720.20440.1641
FSI0.21340.26420.21360.22510.2196
Public foil0.02980.05960.02990.03280.0310
Combustion0.77501.18630.77110.82010.7828
(b) Fair CRPS ↓
Cylinder0.017720.020600.017170.017080.01700
Controlled cylinder0.008260.014760.007740.008860.00764
FSI0.012640.013930.012300.011940.01199
Public foil0.003390.006430.003270.003590.00358
Combustion0.006940.012580.006740.006190.00614

Q2. Probabilistic calibration and joint trajectory fidelity.

Please see our analysis above for the under-dispersion.

For joint trajectory fidelity, we evaluated the order-0.5 variogram score over spatial, temporal and space–time pairs, a proper multivariate score that penalises wrong correlations between grid points, which fair CRPS cannot detect (in units of 10⁻³, lower is better; four members unless stated, mean over three seeds; cylinder on all 320 windows; bold marks the lowest four-member value).

Table L4. Order-0.5 variogram score in units of 10⁻³ (lower is better), four members unless stated, mean over three seeds; cylinder on all 320 windows. Bold marks the lowest four-member value; the score improves with the member count, so the 16-member row is shown for reference. (supplementary Table S5)

MethodCylinderControlled cylinderFSIPublic foilCombustion
Fine-tuned FNO alone (1 member)6.282.784.282.107.63
Gaussian noise around the FNO forecast8.403.764.955.1815.99
MC dropout on the FNO18.4214.125.8615.268.03
PhysGuard (1 member)8.482.824.222.147.88
FROST - EDM residual diffusion, 11 calls5.443.323.841.236.88
FROST, 6 steps5.002.063.471.037.07
FROST, 30 steps4.871.953.390.976.89
FROST, 30 steps, 16 members4.681.723.150.876.55

Thirty-step FROST has the lowest four-member score on cylinder, controlled cylinder, FSI and public foil, is level with EDM diffusion on combustion, and lowers the score of the forecaster it corrects on all five systems. Shuffling member identities between rollout blocks, which keeps each block's marginals, worsens FROST's cross-block variogram score on four of five systems, so the generated rollouts carry temporal dependence beyond what fair CRPS measures.

Thank you again for taking the time to review our paper and for your valuable feedback! We're happy to provide any further clarification.

References

  • Bhola, S., and Duraisamy, K. (2025). Residual-augmented flow matching operators for probabilistic partial differential equations. arXiv:2512.12749.
  • Mardani, M., et al. (2025). Residual corrective diffusion modeling for km-scale atmospheric downscaling. Communications Earth & Environment, 6, 124.
  • Oommen, V., Bora, A., Zhang, Z., and Karniadakis, G. E. (2025). Integrating neural operators with diffusion models improves spectral representation in turbulence modeling. Proceedings of the Royal Society A, 481(2309), 20240819.

I thank the authors for their detailed response, which has addressed most of my concerns. Although I still have reservations about the novelty of the Sim2Real procedure, I maintain my recommendation of weak accept (3).

Comments on the paper

2 comments

Dear reviewers,

Thank you for your reviews. The author-reviewer discussion closes on November 18, so please respond to the authors' rebuttal before then. Please say whether it addressed your concerns, and update your score if your view has changed.

AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

We sincerely thank the Area Chair and all three reviewers for their thoughtful evaluations and for recognizing the timeliness, importance, and broad empirical value of this work. As Reviewer pGMx emphasizes, improving sim-to-real generalization is a timely and important direction for the field, while Reviewers ro2v and LZRB highlight our extensive real-world validation, generality across pretrained forecasters and generative residual formulations, and informative ablations. Sim-to-real generalization remains a fundamental bottleneck in translating advances in scientific machine learning into practical applications, as most learned PDE surrogates are developed and evaluated on numerical simulations rather than real physical systems. Our work addresses this gap by identifying the structured, forecast-dependent, and stochastic discrepancies that conventional deterministic fine-tuning leaves unresolved, and introducing a forecast-conditioned prediction-space adaptation framework that preserves pretrained models while learning their discrepancies from real observations.

Our central methodological contribution is the decomposition of a simulation-pretrained neural operator and a forecast-conditioned stochastic residual corrector for sim-to-real adaptation. Across five real-world systems, our experiments demonstrate the value of this decomposition through improved long-horizon forecasting, probabilistic benefits beyond mean correction (identical-mean CRPS), and more faithful physical trajectory dynamics (temporal spectra, fluctuation energy, and autocorrelation), with substantial improvements over the evaluated baselines.

We have conducted extensive additional experiments directly addressing the specific comparisons and analyses requested by the reviewers. These provide further evidence for the paper's central contributions:

  1. Additional external validation (pGMx): Under matched starting checkpoints, FROST achieves lower rollout error and fair CRPS than Probabilistic Retrofitting on 4/5 RealPDEBench systems. On the additional public Rayleigh–Bénard benchmark, FROST also achieves lower CRPS and VRMSE under the original evaluation protocol.
  2. The value of forecast-conditioned adaptation (ro2v): Matched-capacity history-only controls have 1.5–4.2× higher rollout error across all five systems, demonstrating that the improvements arise from leveraging pretrained forecasts rather than simply additional generative capacity.
  3. Alternative generative methods (LZRB, ro2v): Under matched training and sampling budgets, FROST achieves lower rollout error than residual diffusion across all five systems at eleven correction calls, further supporting our methodological choices.

We further address the reviewers' requests for comprehensive probabilistic evaluation through 16-member ensemble experiments, dependence-sensitive variogram scores, and detailed calibration analyses across rollout horizons and physical parameters. We also provide additional physical-fidelity and computational-cost comparisons, with complete results in the rebuttal supplementary material, updated manuscript, and individual responses.

Taken together, we believe the original contributions, coupled with the substantial additional evidence addressing the reviewers' requests, make a strong case for the significance and real-world applicability of this work. We greatly appreciate the opportunity to strengthen the paper!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.