acceptodds
Under review as a conference paper at ICLR 2027

From Sparse Sensors to Full-Field Posteriors: Physics-Prior Bridges for Stochastic PDEs

Abstract

Reconstructing a stochastic physical field from fixed sparse sensors requires propagating observational constraints into unmeasured regions while preserving posterior uncertainty. We propose a physics-controlled two-stage diffusion-bridge inference framework in which the governing SPDE defines the path prior and sparse observations determine the posterior bridge. We propose a two-stage framework for physics-controlled diffusion-bridge inference. First, a local triangular transport calibrates an all-observation coarse Gaussian reference, yielding a tractable variational path density and unbiased local-window training under a physical reverse-KL objective. Stage 1 provides an important calibration for the coarse Gaussian reference and approximate posterior samples at discretized physical times without backpropagating through full controlled trajectories. Second, a stochastic endpoint network distills the underlying diffusion bridge process controlled by the stochastic PDE based on state pairs within the physics trajectory. At inference, terminal fields are sampled through a short reference prefix followed by a single stochastic endpoint prediction. We derive the conditional training identities and an endpoint error decomposition, establishing a theoretical foundation for connecting local path calibration to efficient terminal-state posterior sampling. Preliminary experiments on stochastic PDE benchmarks, including the one-dimensional Burgers equation, the two-dimensional FitzHugh–Nagumo system, and the Navier–Stokes equations, show promising full-field reconstruction from fixed sparse observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-informed machine learning, stochastic PDE, Diffusion Bridge

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

The paper presents a methodology for estimating the final full-field state of dynamical systems from collections of sparse measurements obtained over time. It uses a path-wise KL objective, Ensemble Kalman synthetic data, and distills a model as a last step to perform one-shot inference. The method is tested on a series of PDE benchmarks and compared to standard Kalman methodology and modern ML alternatives.

Critical strengths

None

Critical weaknesses

The paper presents a method with a training that contains a lot of redundancy; it uses both a KL term and Ensemble Kalman methods where the model is incorporated via the conditional dependence of the probabilistic model, and this is justified by saying that these two aspects offer different information without explaining why this is the case. Also, once training is done, the authors then get rid of the learned transitions and opt for a direct "one-shot" predictor. Overall this seems like a collection of ad-hoc pre-existing methods which in some cases can yield reasonable results, rather than a unified methodology for solving challenging and novel problems. It seems that one is better served by using standard EnKF with at most a learned surrogate.

Questions for authors

For a revised version, there are some typos to correct, such as "long-windoe" and a few more.

2 comments
AuthorsOpenReview ↗

Ensemble supervision on PRBridge training

Why use both the path KL and ensemble supervision

Thank you for your suggestions on the design of ensemble supervision loss. Both local path KL loss and the ensemble supervision loss use the same physical model and the same sparse observations. What differs is how those observations guide the model during the optimization. We further illustrated its effectiveness in the updated manuscript Appendix A with the case study.

In updated Appendix A, Fig. 4 illustrates why this distinction matters. With the most challenging setting of 64 sensors in Navier–Stokes, the positive-vorticity filament turns sharply inside the marked region at frame 100. The Gaussian reference flattens this turn, and KL-only training retains much of that shape, indicating that the path KL only is hard to find the correct the structure. EnKF captures the bend more closely via a large number of ensemble updates by capturing the spatial-censor covariate. Therefore, the model trained with ensemble supervision became a choice for capturing better structures, and this will bring significant improvement in both reconstruction error and its uncertainty. Table 2 tests the contribution of both losses across the NS64 test set: KL alone gives RMSE 0.5751, ensemble supervision alone gives 0.5155, and joint training gives 0.4764. CRPS follows the same ordering: 0.3868, 0.3086, and 0.2798, respectively. The figure identifies a concrete spatial error that ensemble supervision helps correct; the ablation shows that retaining the path KL also improves on ensemble supervision alone.

The path KL defines the target: to learn a path distribution that satisfies the physics dynamics given conditional observations. Its gradient can be estimated from a short window of jointly sampled reference states. The ensemble term gives a more direct field-level training signal: it propagates an ensemble over a longer physical window, uses the evolving state–sensor covariances to update unobserved locations, and trains the sampler towards more plausible structure recovery. These are complementary targets, and they act together for more reliable physics field reconstruction.

Why learn a single step for terminal inference?

The link between the two stages is the terminal state marginal of the learned path distribution. Stage 1 learns from physical transitions and sparse measurements throughout the trajectory, without requiring full-field training targets. It is already a usable sampler: once the Gaussian reference is prepared, a terminal field can be drawn from the final pair of reference states. For stage 2 distillation, it uses terminal samples from this teacher to train a stochastic map from the observation history, initial field, and fresh noise to a terminal field. The main advantage is efficiency: it reduces reference preparation for each new observation history. When intermediate states or complete trajectories are needed, Stage 1 remains available.

Table 2 provides a direct check of whether the teacher still matters after distillation. On NS 64 observation setting, both students produce terminal fields in one step, but the student distilled from the full stage 1 model teacher has lower RMSE and CRPS than the student distilled from the KL-only teacher:

NS64, settingRMSE ↓CRPS ↓Cov90 (%)Time for 32 samples (s/instance)
One-step student from KL-only teacher0.57130.390431.90.027
One-step student from full Stage 1 teacher0.48960.286762.80.028
Full Stage 1 teacher itself0.47640.279885.211.623

An EnKF with a learned surrogate is a useful alternative, and the comparison depends on what the surrogate replaces. A learned forecast solver would accelerate propagation while retaining EnKF’s sequential observation updates. A terminal generator trained on EnKF outputs would instead amortize the entire reconstruction and is the closest alternative to our Stage 2. Our experiments compare against standard EnKF; the two students above are both trained from PRBridge teachers. On the NS 64 observation setting, standard EnKF itself gives better CRPS than full Stage 1 (0.2560 vs. 0.2798), while Stage 1 gives lower RMSE (0.4764 vs. 0.4902).

Manuscript revision and presentation

We proofread and corrected typos, and revised the Introduction to clarify the roles of path-KL training, ensemble supervision, and one-step distillation. Appendix A (Fig. 4) now illustrates how ensemble supervision helps recover a vorticity filament in the 64-sensor Navier–Stokes example. Appendix B (Table 3) reports additional Cylinder results with 107 and 432 sensors. The corresponding changes are highlighted in blue. The updated manuscript has been uploaded and major revisions are highlighted in blue fonts.

I thank the authors for replying to my comments.

My concerns are not addressed in a manner I find convincing. In particular, finding a specific case in NS where their methods provide a benefit over standard EnKF methodology falls short. I will be keeping my score as it is and I thank the authors for their time.

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

This paper proposes PRBridge for reconstructing physical fields from sparse sensors under known stochastic PDE dynamics. The method combines three components: Gaussian-reference parameterization enabling path-space KL optimization without full-trajectory backpropagation, long-window ensemble supervision propagating observation information via known physics, and bridge distillation for single-step terminal sampling.

Critical strengths
  1. The paper attempts to tackle a practically relevant problem (sparse sensor reconstruction) with a structured approach which is intuitively makes sense.
  2. The mathematical formulation involving is technically sound.
Critical weaknesses

1.The main method appears to be a combination of known methods presented in an overly technical manner. 2. The problem requires knowing the complete stochastic PDE, all physical parameters, and the full initial field. This assumption is unrealistic in almost any practical scenario. 3. The three claimed insights are all existing techniques under different names. 4. The ablation (Table 2) shows that neither KL loss alone (CRPS 0.0308, Cov90 53.8%) nor ensemble supervision alone (CRPS 0.0426, Cov90 50.3%) work independently. Only their combination produces reasonable results (CRPS 0.0250, Cov90 77.8%). This suggests the design relies on hyperparameter tuning rather than a principled framework. 5. The paper mentions cylinder flow measurements three times (abstract, introduction, conclusion) but provides no results in the main paper.

Questions for authors
  1. Can you address your primary novelty or contribution? 2.Can you name a real-world application where the complete stochastic PDE, all parameters, and the full initial field are actually known?
  2. Table 2 shows EnKF outperforms PRBridge on CRPS and coverage without training. Why should practitioners use PRBridge?
  3. Where are the cylinder flow results mentioned in the abstract and conclusion?
4 comments
AuthorsOpenReview ↗

Contributions and practical scenario clarification

Primary methodological contribution

Thank you for suggesting that we address the main contribution more precisely. We have revised the Introduction and made the uploaded manuscript to make it explicit. Our central contribution on PRBridge is a physical-time path diffusion bridge-based sampler that uses only the sparse observation history for physical field reconstruction. The novelty specifically expands on three challenges:

Challenge 1: Global conditioning with local computation. How can a path sampler use the entire observation history while estimating the gradient of a global path objective without backpropagating through a full physical rollout?

Challenge 2: Correcting spatial structure beyond the sensors. When the Gaussian reference misrepresents an unobserved structure and KL-only training retains that error, how can the model use the physical evolution over a longer window to correct it? (examples can also be found in the updated Appendix A)

Challenge 3: Efficient sampling for a new observation history. Once the path sampler has been trained, how can we generate terminal fields for a new set of measurements without preparing a new Gaussian reference?

Thus our contributions can be listed three fold:

  • A globally conditioned path sampler with local optimization Our main methodological contribution is to enable local optimization on global path distributions via coupling a physics-informed Gaussian reference in a local invertible transport. The key distinction from the existing works is that the input state Gaussian reference is coupled by physics dynamics instead of being sampled independently across different states: its mean is obtained by smoothing a linearized physical model using the full sparse observation history, and its covariance couples fluctuations across time steps. After reference preparation, a randomly selected time index and at most three jointly sampled adjacent reference states suffice for an unbiased estimate of the discrete path-KL (Proposition 3.3) without backpropagating through a full physical rollout.

  • Long-window ensemble supervision. We propagate an ensemble through the stochastic dynamics and assimilate sparse measurements using the evolving state–sensor covariances. The resulting fields provide approximate training targets at unobserved locations. Appendix A illustrates the benefit: in the 64-sensor Navier–Stokes example, the Gaussian reference and KL-only reconstruction flatten a bend in a vorticity filament, whereas the jointly trained sampler recovers the bend more closely.

  • One-step terminal sampling. PRBridge Stage 1 can already generate a terminal field from the final pair of reference states. To remove the remaining cost of gaussian reference preparation, we distill the teacher’s terminal distribution into a stochastic student that maps the initial field, sparse observation history, and fresh noise to a terminal field in one forward pass.

Practical scenario and real-world use case relaxation

The assumption of known initial conditions can be relaxed when the initial measurement is incomplete or noisy: given the available measurements yy, one could draw plausible initial fields from p(x0∣y)p(x_0 \mid y) and run PRBridge conditionally on each draw. Pooling the terminal samples would propagate initial-state uncertainty according to

= \int p(x_T \mid x_0, y) p(x_0 \mid y)\,\mathrm{d}x_0$$ For PRBridge, estimating $p(x_0\mid y)$ would be extended with additional inference steps which are compatible with our current pipelines. The initial field already varies across trajectories in our simulations: for each PDE, we sample 1,000 initial fields from a prescribed distribution and generate ten independent realizations of the stochastic dynamics from each field. The real Cylinder flow experiment illustrates a setting in which the initial field can be measured directly. The first particle-image-velocimetry (PIV) frame provides the velocity field on the measurement plane. For all subsequent frames, the reconstruction uses only the velocity measurements at fixed sensor locations. This tests a concrete monitoring protocol: an initial full-field snapshot followed by sparse sensing under a prescribed flow model. With 28 sensors, Stage 1 achieves an RMSE of 0.2295 (main text, Table 1(c)); Appendix B reports the 107- and 432-sensor results. For more real-world application scenarios with initial condition relaxation, in heat-conduction experiments, a specimen can start at a known uniform temperature and then be monitored by a few thermocouples. In ensemble weather forecasting, observations inform an ensemble of plausible initial fields, each could serve as the input to a conditional reconstruction.
AuthorsOpenReview ↗

Ensemble filtering on PRBridge

When should practitioners use PRBridge rather than EnKF

A practical advantage of PRBridge is the inference efficiency. On FHN, EnKF must propagate its ensemble through the physical dynamics and update it at each of 100 observation times for every new trajectory. PRBridge uses long-window ensemble forecasts and analyses to supervise the sampler during training. At inference, once the Gaussian reference is prepared for a new observation history, Stage 1 draws terminal fields directly from the final pair of reference states. With 256 sensors, producing 32 terminal samples takes 2.297 seconds per instance with Stage 1 versus 14.384 seconds with EnKF, a 6.3-fold inference speedup. This makes PRBridge particularly useful when many reconstructions share the same physical system and sensor layout, so the offline training cost can be amortized. For the same 32 samples per instance, inference takes 2.297 seconds with Stage 1 versus 14.384 seconds with EnKF. Stage 1 requires 2.764 GPU-hours of offline training, so this trade-off is most attractive when that cost can be shared across many reconstructions.

For more metric evaluations, PRBridge always provides the lower reconstruction error and keeps comparable uncertainty results with EnKF. Table 1 provides a direct comparison with EnKF, while Table 2 isolates the effects of the training objectives and distillation. Across the six settings in Table 1, PRBridge Stage 1 has lower RMSE and relative L2L_2 error than EnKF in all six, lower CRPS in four, and coverage closer to the nominal 90% level in three. EnKF does perform better on both CRPS and coverage for Burgers with 16 sensors. The choice therefore depends on the reconstruction metric and how often the model will be used. A concrete use case for PRBridge is repeated terminal-field reconstruction under a fixed sensor layout. On FHN with 256 sensors, Stage 1 reduces RMSE from 0.0573 to 0.0483 and CRPS from 0.0276 to 0.0250, with similar coverage (78.4% for EnKF versus 77.8% for Stage 1).

Ensemble filtering and PRBridge also play complementary roles. In updated Appendix A, Fig.4 illustrates this in the 64-sensor Navier–Stokes experiment. At frame 100, the Gaussian reference flattens a sharp bend in the marked vorticity filament, and KL-only training leaves much of this error unchanged. EnKF captures the bend more closely, and PRBridge trained with the ensemble loss better reproduces its curvature as well. Across the full test set for this setting, adding ensemble supervision reduces RMSE from 0.5751 to 0.4764 (Table 2). Thus, PRBridge can benefit from ensemble filtering during training without rerunning the filter for each new reconstruction.

Supplement Cylinder flow results:

The results for most challenging setting of 28 sensors are already reported in Section 5 and Table 1(c) of the main paper. We have also updated Appendix B (Table 3) to report Cylinder flow results for 107 and 432 fixed sensors, corresponding to 1.49% and 6.03% of the 64 × 112 grid, respectively. Bold values denote the best result within each training-supervision group; the target for Cov90 is 90%.

Cylinder flow: 107 sensors (1.49%)

Training supervisionMethodRMSE ↓Relative L2L_2 ↓CRPS ↓Cov90 (%)
SparseGaussian smoother0.17010.23540.083772.5
SparseEnKF0.15210.21070.083677.3
SparsenMA0.21440.29440.109919.3
SparseSDE Matching0.17890.24740.080231.0
SparseNGDB0.15150.20970.074076.7
SparsePICSB0.16130.22330.090124.3
SparsePRBridge0.14840.20540.074579.5
DenseDiffusionPDE0.17000.23090.050577.8
DenseSDA0.16040.19100.060371.0
DensePalSB0.13700.18970.071637.3
DensePRBridge0.08550.11830.036682.4

Cylinder flow: 432 sensors (6.03%)

Training supervisionMethodRMSE ↓Relative L2L_2 ↓CRPS ↓Cov90 (%)
SparseGaussian smoother0.08990.12750.042676.3
SparseEnKF0.08200.12180.041480.6
SparsenMA0.22520.30920.103214.1
SparseSDE Matching0.09790.13550.049148.5
SparseNGDB0.08090.11460.038377.9
SparsePICSB0.10740.15240.058135.6
SparsePRBridge0.07640.10820.036381.1
DenseDiffusionPDE0.11760.15610.034681.5
DenseSDA0.06430.09820.032780.6
DensePalSB0.08300.11780.035264.4
DensePRBridge0.05430.07690.024883.8

Among the sparse methods, PRBridge achieved the lowest RMSE at both sensor budgets: 0.1484 versus 0.1521 for EnKF at 107 sensors, and 0.0764 versus 0.0820 at 432 sensors. Its CRPS was close to the best at 107 sensors (0.0745 versus 0.0740 for NGDB) and lowest at 432 sensors (0.0363).

Thank you for the responses. I accept the correction regarding Table 1(c) and Appendix B. However, my core concerns remain unaddressed. In particular, the initial-field relaxation is given as a formula and explicitly left for future work, so the practical scenario is still untested. I will follow the discussion before finalizing my assessment.

AuthorsOpenReview ↗

Empirical evaluation with an unknown initial field

Thank you for discussion. For more practical scenario without knowing the exact initial condition, we have added a small-scale experiment on a Navier–Stokes subset with 16x16 fixed sensors.

The initial field is randomly sampled from the sparse observations of the initial field, and the gaussian reference and ensemble filtered training signals are calculated based on the randomly sampled initial field, then updated using the same sparse observation history towards the final state. Then we trained our PRBridge based on the input of the same gaussian reference and supervision ensemble filtering

Method / evaluated quantityRelative L2L_2 error ↓
Unknown initial field: prior ensemble mean0.9001
EnKF initialized from the unknow initial field0.2192
Unknown-initial field PRBridge — test set0.2156

The first row measures error at the sampled initial frame; all remaining rows measure terminal reconstruction error. The prior initialization differs substantially from the true initial field. Nevertheless, after assimilating the sparse observation history, the prior-initialized EnKF achieves lower terminal reconstruction accuracy, and these ensembles therefore provide substantially corrected terminal field estimates for supervision. On the test subset, the learned model achieves a slightly lower relative �1��1� error than the initialized EnKF. These results provide empirical support for the unknown initial field extension: ensemble-based supervision enables a model to reconstruct the terminal field from sparse observations without access to the true initial field.

Reviewer_uk6WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper combines a Gaussian reference with learned transport and ensemble supervision to reconstruct full physical fields and their uncertainty from sparse observations under known stochastic dynamics and uses distillation for faster sampling.

Critical strengths
  1. The local training objective is well motivated. The derivation explains how the model uses the full observation history without differentiating through complete physical trajectories.

  2. The experiments include simulated systems and real measurements. The ablations help explain the contribution of the main components. The evaluation considers both reconstruction accuracy and uncertainty.

Critical weaknesses
  1. I am concerned about how the auxiliary targets are constructed. The starting samples already depend on the full observation history. The ensemble update then uses observations from the same history again. This can shift the targets away from the intended posterior and affect the learned uncertainty. The paper acknowledges that these targets are approximate but does not directly check this source of bias.

  2. The real-data experiment selects the Gaussian reference using full-field validation data. The paper states this clearly but does not show whether similar results can be obtained using only sparse validation observations. Since the reference affects both reconstruction and uncertainty, this leaves an important gap in the evidence for the sparse-data setting.

Questions for authors
  1. How does the auxiliary update avoid counting the same observation information again? Please check the posterior mean and uncertainty in a simple setting where the true posterior is available. This would help show whether the auxiliary targets introduce systematic bias.

  2. Can the reference be selected using only sparse observations? Please compare reconstruction accuracy and uncertainty under this choice and explain how the reference settings are selected.

If these concerns are answered well with clear explanations and supporting results, I would be willing to change my recommendation to acceptance.

4 comments
AuthorsOpenReview ↗

Auxiliary supervision targets and gaussian reference selection

Auxiliary update and its potential shift on reconstruction

Thank you for your suggestions. The ensemble analysis reuses window observations, and we evaluated such effect in a linear–Gaussian system with a tractable posterior, and also provided the visualization of complicated PDE cases in the updated Appendix A.

We generated 40-step trajectories from xn=Axn−1+ηnx_n=A x_{n-1}+\eta_n with fixed x0,j=0.5sin⁡(2πj/16)x_{0,j}=0.5\sin(2\pi j/16) and A=0.98(0.75I+0.25S)A=0.98(0.75I+0.25S), where SS is a cyclic shift (with train:val:test = 8000:1000:1000). At every step we observed coordinates 0 and 8 with additive noise. Process and observation noises were independent Gaussian draws with covariances 0.01I160.01I_{16} and 0.04I20.04I_2, respectively. This gives an exact posterior for the 14 unobserved coordinates. For the first 100 test trajectory comparison, A was initialized at step 35 using only past observations, whereas B was initialized by smoothing with all 40 observations. Both then used the same physical forecast and Kalman updates from steps 36 to 40. A recovered the exact terminal posterior. Even with an exact starting marginal, B shifted the mean average Eμ=0.0613E_\mu=0.0613 and reduced total variance by 1.40%. We further compared full dataset among Gaussian reference, EnKF, and Stage1 PRBridge trained with KL alone or with auxiliary supervision at different weights; the results can be found below:

MethodMean EμE_\muePe_PVVCov90 (%)W90W_{90}
Analytic auxiliary targets: same final five observations
A: filtering initialization001——
B: full-history smoothing initialization0.06130.027030.98605——
Terminal distributions: baselines and learned Stage 1 samplers
Gaussian reference, analytic0.00002.55563.4948100.000.9806
Gaussian reference, 256 draws0.08702.58523.4892100.000.9782
EnKF, M=64M=640.22610.38380.945185.240.5937
B1: KL only0.04420.60320.574975.880.4864
B2: KL + auxiliary, λ=0.1\lambda=0.10.04820.57960.628578.970.4879
B2: KL + auxiliary, λ=0.25\lambda=0.250.05580.53140.682979.780.4980
B2: KL + auxiliary, λ=1\lambda=10.06480.50980.718282.370.5281

On the unobserved coordinates, EμE_\mu measures standardized mean error, ePe_P relative covariance error, and VV the total variance ratio, all relative to the exact posterior. On the linear–Gaussian benchmark, adding auxiliary supervision improved the trained sampler’s empirical coverage of nominal 90% intervals: from 75.88% with KL-only training to 82.37%, but at the cost of a higher posterior-mean error. The auxiliary target shows a trade-off: it recovers the posterior with better coverage 90, but at a slight cost in mean accuracy.

However, the auxiliary targets could provide a useful training signal in complex stochastic dynamics. In the updated Appendix A, a Navier–Stokes example with the most challenging setting of 64 fixed sensors, illustrates how ensemble supervision can help correct a spatial feature that KL-only training leaves unresolved. At frame 100, the positive-vorticity filament turns sharply in the circled region of the ground-truth field. The Gaussian reference smooths out this bend, and the KL-only reconstruction retains much of that error. Both ensemble filtering and the model trained with the auxiliary loss recover a bend closer to the one in the ground truth. The two losses use the same observations but provide different training signals: the forecast–analysis ensemble gives an explicit target for the spatial correction, while the path-space KL evaluates how well the learned trajectories agree with the stochastic dynamics and observation likelihoods.

Sparse gaussian reference selection in Cylinder flow

We select the Gaussian reference by minimizing the average predictive negative log-likelihood of sparse validation with the settings of the 28-sensor layout. After re-building the reference using all sparse sensors, we re-train the PRBridge Stage1. From the following table, sparse selection yielded nearly identical Stage 1 RMSE, slightly better CRPS coverage 90% from 78.1% to 80.6% with wider intervals.

MethodReference selectionRMSERelative L2L_2CRPSCov90 (%)W90W_{90}
Gaussian referenceDense val(Table 1)0.25240.34750.106169.60.2508
Gaussian referenceSparse val0.26900.35660.112569.10.2716
PRBridgeDense val (Table 1)0.22950.30760.106378.10.2391
PRBridgeSparse val0.23060.31600.097780.60.2874

More result comparisons on Cylinder flow for different sparse observation settings: 107 sensors (1.49%) and 432 sensors (6.03%) across baselines can be found in the updated Appendix B.

Thanks for your reply. and I have two more concerns for your reply: using the same observations twice introduces bias, and I would like to check how to deal with this bias in the final model. For the experiment results Adding more supervision increases the mean error from 0.0442 to 0.0648, and I could like get any explanation for this. For the second parts of the experiment results, the MSE is about 0.3X. Which is not reasonable for the physics problem. Please check the model setting or the training setting. Current results just show me the model is not trained very well.

AuthorsOpenReview ↗

Error analysis

Thank you for the follow up. We added experiments to examine the effect of auxiliary supervision on the final model regarding bias and further inspected the reconstruction errors.

Auxiliary supervision and final-model bias

We compared two schedules under the same lightweight linear gaussian experiments discussed above. The refinement shares the first 16,000 updates with joint training, then sets the auxiliary weight to zero while retaining the model and optimizer states. Besides adding a regular fixed weight for the auxiliary ensemble supervision, we add another weight scheduling from \lambda =1 -> 0 in the optimizing steps:

Training scheduleMean EμE_\muMean ePe_PCov90 (%)
KL + auxiliary throughout0.06480.509882.37
KL + auxiliary → KL-only0.05070.527981.62

The final KL-only phase reduces posterior-mean error relative to continued joint training.

Errors of Gaussian reference input

Appendix A illustrates an NS case where the reference mean contains an incorrect local structure. To simulate and examine such imperfect initialization, we deliberately introduced bias into the Gaussian reference input mean to examine whether ensemble-supervised training can help improve an imperfect reference. On the same linar gaussian system above, starting from this biased input, training with joint KL and ensemble supervision reduces the empirical posterior-mean error and improves covariance agreement and CRPS. These results show that theauxiliary trained sampler can reduce errors when there is biased input reference distribution. This complements Appendix A, where ensemble supervision helps correct a local spatial structure that remains unresolved under KL-only training.

DistributionMean EμE_\muMean ePe_PCov90 (%)CRPS
Biased Gaussian reference input0.100302.589098.070.1713
Output after KL + auxiliary training0.07250.587779.560.1119

Another explanation for error is that because the predictive mean is estimated from a finite ensemble, a broader predictive distribution can increase this contribution, making the measured error appear larger than KL-only trained setting and systematic mean bias.

Error analysis on experiments:

The Cylinder values around 0.3 are relative L2L_2 errors, rather than MSE. As an optimization sanity check, we inspected the KL-derived validation objective under fixed reference, normalization, and evaluation settings. The log error decreased from 1.742 to −4.216 for NS and from 1.978 to approximately −1.901 for Cylinder. Negative values are consistent with Eq. (12).

To explain why the larger RMSE appears to be large in the experiments. To understand why recognizable structures coexist with appreciable relative error, we additionally examined the same Appendix A NS predictions on the same 128 validation histories: For the original Gaussian-reference model, spatial mismatch accounts for 96.41% of the full-field normalized squared error. Additional spectral diagnostics on the EnKF-reference variant locate approximately 69% of the error at intermediate spatial frequencies, while approximately 88% of the ground-truth energy lies at larger scales.

Thus, the main structures can appear plausible while the phase and high-frequency components remain inaccurate. Figure 3 additionally illustrates conditional sample variability in FHN. Such positional or phase variability can smooth local features in the ensemble mean. More generally, residual conditional uncertainty contributes to expected reconstruction error against an individual realized field, even for an accurate posterior mean. This provides context for interpreting the qualitative figures alongside the remaining pointwise errors.

Thanks for your time and effort. But for me, the relative error for the physics is around 30 percent, which means the model is not trained very well, or there are some issues about the training setting. I keep my score.

Comments on the paper

1 comment
Area_Chair_1DrbOpenReview ↗

Please engage in the discussion

Reviewers: if you have not replied to the authors' response, please read it and engage in the discussion. Thanks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.