acceptodds
Under review as a conference paper at ICLR 2027

Operator Sensitivity as an Inductive Prior for Uncertainty Quantification in DeepONets

Abstract

We investigate whether the input-output sensitivity of a learned operator can provide an inductive prior for the spatial structure of predictive uncertainty. We introduce Jacobian-DeepONet, which parameterizes asymmetric prediction intervals using the normalized pointwise Jacobian norm and two additional trainable scalars, jointly optimized with the operator under mean-squared-error and pinball losses. We also introduce Jacobian Conformal Bands (JCB), which calibrate sensitivity-shaped intervals on a frozen operator without retraining. Both methods exploit the Cartesian DeepONet factorization to compute Jacobian norms exactly through a latent Gram matrix using reverse-mode vector-Jacobian products, without materializing the full input–output Jacobian. On viscous Burgers, Darcy flow, and compressible Navier-Stokes, the methods achieve competitive interval widths after calibration and stronger alignment with reference physical sensitivities than constant-width and learned-quantile baselines. Monte Carlo perturbations validate the first-order variance approximation, while experiments with noisy inputs show improved coverage retention for JCB relative to conditional quantile bands under the tested perturbations. Our results suggest that operator sensitivity provides useful, problem-dependent structure for prediction intervals in DeepONets, complementing uncertainty models learned from residuals.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty quantification, neural operators, DeepONet, operator learning, scientific machine learning, partial differential equations, quantile regression, sensitivity analysis

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_z8FaOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

Here, the operator's own Jacobian tells prediction intervals where to widen. One version learns those intervals jointly with DeepONet. Another calibrates a frozen model afterward. A latent Gram matrix avoids constructing the full input–output Jacobian. I find that computational trick convincing, though the uncertainty interpretation needs more work.

Critical strengths

The latent-Gram calculation is the strongest contribution for me. It computes exact Jacobian row norms without storing the full output Jacobian. That's a useful saving on dense grids.

I appreciate the same-backbone tests and three-seed reporting. The frozen-operator comparisons also hold the mean prediction fixed. That makes the interval-shape comparison much cleaner.

Critical weaknesses

The Monte Carlo check validates the trained network against itself. It checks input-noise propagation, rather than true solver sensitivity or approximation-error uncertainty. I'd keep those claims clearly separated.

The gains depend heavily on the PDE. Darcy's trained bands are 6% wider than conditional quantile bands. Its frozen Jacobian bands also lose to the training-variance template on interval score.

Average coverage hides some ugly tail behavior. On noisy Burgers' hardest decile, JCB covers 55.1%, versus 81.8% for conditional quantiles. That limits its appeal for difficult cases.

Questions for authors

Do the results survive using calibration-only floor and compression statistics? I'd also want repeated calibration splits and actual post-calibration coverage beside every width comparison.

What exactly is the exchangeable calibration unit: whole fields or pooled spatial nodes? Please give a coverage argument that respects within-field dependence. That matters more to me than another aggregate PICP table.

Does the sensitivity advantage survive matching the reference perturbation model exactly? I'd want the same input coordinates, covariance, and observed sensors. Calling Darcy's mismatched reference “conservative” needs justification.

Does detaching the Jacobian during training preserve the gains? Please compare derivative accuracy and input curvature directly. That would tell me whether the claimed regularization actually happens.

1 comment

Thank you for the detailed review, the questions were very specific and helped us find where the claims went further than the evidence. We agree with the three weaknesses and changed the paper accordingly.

The Monte Carlo check does validate the trained network against itself. It compares two estimates of the same quantity, the output variance of the network under input noise, so it only verifies the first-order approximation. We now say this in Section 3.1 and in Appendix G, and keep it separate from the other two quantities: agreement with the solver sensitivity is measured by rho_Sens against solver derivatives, and agreement with the approximation error by rho_Err against the realized residuals.

The dependence on the PDE is real. The Limitations now state that on clean inputs the sensitivity prior is not the most efficient band, since the conditional quantile head and the ensemble reach a lower interval score on Burgers and Darcy. On Darcy the frozen Jacobian band is between 2% and 3% above the training standard deviation template in interval score, depending on the calibration split (Table 1, Appendix C.6), and below it with one score per field (Appendix C.6). Given this, we narrowed the claim to what holds across the three PDEs: the alignment of the width with the solver sensitivity, and the coverage kept when the input is perturbed.

The tail behavior is also true. We now report the hardest decile coverage under input noise for the trained bands and JCB (Table 12), and the text states that the conditional quantile head keeps the highest coverage there on all three PDEs. This is expected, as it is the only band fitted to the residuals, so it is told where the model error is large, while JCB reads its width from a frozen operator and only uses the residuals through one scalar. Training the prior jointly closes part of the gap (Jacobian-DeepONet covers 61.7% against 55.1% for JCB on Burgers, and 74.2% against 54.2% on Darcy). We state it as a limitation for difficult cases.

Questions for the Authors:

Q1. Yes for the floor. We added Appendix C.6, where the floor is computed on the calibration fields alone, the multiplier uses the finite-sample conformal rank, and the calibration/evaluation split is repeated 20 times for each of the three seeds. Coverage moves by under half a point and the interval score by under 1%, with a standard deviation of about one point over splits. The compression scale is only used by the log-compressed shape, a secondary variant in Table 11 and Appendix G, and it is still pooled over the test split; we now state this in Appendix A.5 and removed the sentence claiming that the multiplier absorbs it. Every calibrated width comparison in the paper now reports the post-calibration coverage beside it; the loss-weight sweep of Table 7 is the exception, reported as trained by design.

Q2. The exchangeable unit is the whole field. Nodes within a field are dependent, but fields are exchangeable. Pooling the N node scores of each of the n calibration fields and taking the rank ceil((1-alpha)(n+1)N) guarantees that the expected fraction of covered nodes in a new field is at least 1-alpha: adding the test block can only lower the value at that rank, and by symmetry between blocks each block has in expectation at least that fraction of its scores below it. This says nothing about a single field. For whole field coverage we also calibrate one score per field, its maximum normalized residual, which covers a new field with probability at least 1-alpha, at 2.7 to 4.2 times the interval score (Appendix C.6). The argument is now written in that appendix.

Q3. The word conservative was not justified and we removed it. rho_Sens compares the band, a row norm over all inputs under isotropic noise, with the solver derivative in a single direction, so it measures rank agreement with that direction and not with the same perturbation model. We now state it this way in Section 4 and Appendix B, and leave the exactly matched comparison as future work.

Q4. We did not run this ablation in the rebuttal period, and we agree that without it the regularization claim is not supported. We removed it from the main text and the contributions. Appendix A.2 now only states that backpropagating through the norm adds second-order terms to the gradient, and that whether they change derivative accuracy or curvature was not measured. We leave it as future work.

Reviewer_AUvhOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In my opinion, the best way to look at the contribution of this work is to compare it to prior approaches. UQ methods like Deep Ensembles or MC Dropout stick to the definition of epistemic uncertainty as measure the sensitivity of the prediction to the model's parameter space (the weights, the initial seed, etc). This method asks about the sensitivity to the input. Framing the UQ problem this way enables the method to give up on mathematical rigour for computational speed. There are very similar UQ approaches focusing on this Jacobian trick like Stable Distribution Propagation, that use the Jacobian trick for aleatoric uncertainty, but this approach focuses on epistemic uncertainty. The catch is that this Jacobian trick with conformal calibration will not be reliable for covariate shifts (or Out Of Distribution detection) that is the most essential feature of epistemic uncertainty (this is my opinion). Furthermore, this method is niche as it works only for the branch and trunk architecture of DeepONets. Addiitonally, DeepONets are used lesser and lesser now, as the SciML field has moved to transformer based models like Transolver, GeoTransolver, etc. With DeepONets, we can just use MC Dropout etc and get reliable OOD as well.

Critical strengths
  1. Where it is applicable, as in DeepONets, the approach does give some gains in speed.
  2. Where it is applicable, the method does use conformal calibration to ensure marginal coverage guarantees.
Critical weaknesses
  1. Niche: The proposed method is hardly universal even in SciML. It works ONLY for the branch and trunk architecture of DeepONets. With the SciML community transitioning over to transformer based models. In models like Transolver, GeoTransolver, DoMiNo, etc. this method would be very very computationally expensive.

  2. Covariate Shifts / Out Of Distribution Detection: The method implicitly assumes that under any covariate shift, the model will show high gradients lead to high sensitivity. This is an assumption and it is not proved for even a single case. I would encourage the authors to add tests showing that their method is robust to different types of covariate shifts encountered in SciML.

  3. Multi-Modal Training Data: A case where (in my opinion) this method would fail is multi-modal training data where the trainign dataset exists as 2 or more separate clusters. In such applications, I have found that DeepONets lead to smooth interpolation between the 2 clusters. This leads to smooth gradients and a constrained Jacobian. Here, the approach would output overconfident uncertainty predictions.

Questions for authors
  1. Covariate Shifts / Out Of Distribution Detection: The method implicitly assumes that under any covariate shift, the model will show high gradients lead to high sensitivity. This is an assumption and it is not proved for even a single case. I would encourage the authors to add tests showing that their method is robust to different types of covariate shifts encountered in SciML.

  2. Multi-Modal Training Data: A case where (in my opinion) this method would fail is multi-modal training data where the trainign dataset exists as 2 or more separate clusters. In such applications, I have found that DeepONets lead to smooth interpolation between the 2 clusters. This leads to smooth gradients and a constrained Jacobian. Here, the approach would output overconfident uncertainty predictions. I would urge the authors to add tests on such multi-modal training data and add the results in the revision.

1 comment

Thank you for the time to review this work, we appreciate the insights derived from it. As you mention on the weakness, the original paper was only tested and formulated for DeepONets, which was a clear limitation, specially on recent research that may not be inclined to them. Given that, we stretched the horizons of the paper and tested the JCB against Fourier Neural operators. For this we use the direct JVP instead of the gram matrix, achieving similar results but with higher cost. In future works we propose to use a Jacobian approximation instead of the full. This in theory should also work with transformer based architectures, but we chose to leave it for future directions given the time constraints.

The assumption holds only in part. The Jacobian-shaped width grows when the shift makes the operator more sensitive, as with a lower viscosity, but not when the shift moves the input to a regime where the error grows without a change in sensitivity, as with a larger input amplitude. There JCB covers less than a constant band. With the multiplier calibrated on clean data and frozen, JCB covers more than the constant band under the two viscosity shifts and less under the amplitude shift on Darcy, and the deep ensemble covers more than JCB under all three (Appendix H, Table 16). We therefore state in the Limitations that JCB is not an out-of-distribution detector. Input sensitivity and weight uncertainty expose different failure modes, input noise and covariate shift, which is why we suggest combining both (Section 6).

Considering multi-modal training, we leave it as a clear limitation and to test it in future directions. Sadly we prioritized adding more UQ methods to the benchmark because of the time constraints.

Thank you for the review.

Reviewer_wQgdOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work investigates first order sensitivity in DeepONets and investigates a possible link between the uncertainty and the sensitivity. The paper proposes two approaches. The first option is trained end-to-end, using a Jacobian norm and combination of several loss functions to learn asymmetric prediction intervals. This has the result of allocating network capacity to locations in space where the operator is sensitive -- a proxy for where the operator is poorly calibrated to the dynamics of the PDE -- and as a result, where the model may appear to be "uncertain". The second approach is a post-hoc fitting of a Jacobian conformal band to a trained DeepONet. This is essentially a first-order estimate of the sensitivity of the model to perturbations in the input space. The paper subsequently draws conclusions about the relationship between this sensitivity and the uncertainty.

Critical strengths
  1. The question of estimating uncertainty from pre-trained models is interesting and has practical implications for foundation models. This paper attempts to answer a reduced version of this question.
Critical weaknesses
  1. The number of PDEs used across the experiments is far too small to even determine if the results from this paper hold more generally. All three are likewise extremely simple toy problems which have very limited ties to modern machine learning.
  2. The paper only considers one model architecture -- the DeepONet. There are two significant issues with this. First, the DeepONet has extremely limited practical use. The larger research community has largely abandoned DeepONet style architectures precisely for this reason. Second, there is no clear extension of the uncertainty quantification to more general model architectures, as it relies specifically on components of the DeepONet.
  3. The work references both epistemic and aleatoric uncertainty, but largely neglects problems in the latter case. The PDEs from Section 4 are entirely deterministic. Section 5.3 is a step in the right direction to understand whether this approach can actually model the uncertainty that is inherent in turbulent 3D flows and chaotic systems, but it is severely undeveloped. Additionally, we see in this case that the uncertainty bands actually do a poor job of modeling this uncertainty.
  4. The paper is poorly written. The results section is particularly difficult to follow.
  5. It is unclear if the relationship between sensitivity and uncertainty holds more generally, for example in other models, trained with different loss functions or objectives.
  6. There are no external baselines in any experiments.
Questions for authors
  1. How does this approach compare to other approaches for modeling uncertainty in neural networks and neural operators? There are many extremely simple baselines when training from scratch, dropout and BNNs being two notable examples. Likewise, there are post-hoc uncertainty calibration approaches for pre-trained networks in the literature (arxiv 2502.20966v1).
  2. Many phenomena we observe in the real world come from noisy measurements or have incomplete data, see arxiv 2601.01829 for several examples. How does this approach fare in capturing uncertainty estimates of these systems?
  3. Can we go beyond empirical results, drawing a link more formally between the sensitivity analysis in this work and estimating both epistemic and aleatoric uncertainty?
1 comment

Reviewer wQgd

Thank you for your review, which gave clear direction on what was missing. We address the weaknesses first and then the questions.

Number and complexity of PDEs: We agree that three PDEs are a limited benchmark. We chose standard operator-learning problems (viscous Burgers, Darcy flow, compressible Navier–Stokes) so that the interval shapes could be compared on a common backbone against reference sensitivities computed by the solver. We do not claim that the results hold beyond these problems.

Architecture: DeepONets remain in active use in scientific and engineering applications, but we agree that restricting the method to them limited its relevance. The post-hoc method only requires Jacobian-vector products, and the revision repeats the full post-hoc comparison on a Fourier Neural Operator, with the deep ensemble and LUNO as baselines (Table 2). The same division found on the DeepONet holds on all three benchmarks: the weight-based bands are more efficient and more aligned with the model error, and the Jacobian band is the most aligned with the reference sensitivity. Transformer-based operators are left as future work (Section 6).

Aleatoric versus epistemic uncertainty: Our question is whether the sensitivity of the operator describes the uncertainty that enters through its input. JCB does not model the intrinsic stochasticity of the dynamics: to first order, its width is the standard deviation of the output under Gaussian noise on the input, and split conformal calibration absorbs the model error. This is why the solvers are deterministic: the uncertainty lies in the input, not in the dynamics. We agree that turbulent and chaotic systems are not covered; there the Jacobian norm grows with the horizon and the first-order approximation may break down, which we now state in the Limitations. Section 5.3 has been expanded to two noise structures (isotropic and spatially correlated), all DeepONet baselines, and both coverage and interval score (Table 3).

Relationship between sensitivity and uncertainty in other models: The FNO experiments test this relationship on a second architecture and find the same pattern (Section 5.2). Other training objectives were not evaluated.

External baselines: The revision adds Deep Ensembles and MC Dropout on the DeepONet, and Deep Ensembles and LUNO on the FNO (see Q1).

Presentation: We agree that the paper was hard to follow, largely because results were added until the end of the submission period. In the revision we simplified the abstract and introduction, define every term before it is used, and rewrote the results section to give short descriptions with references to the tables rather than repeating the numbers in the text.

Questions for the Authors:

Q1. We added Deep Ensembles and MC Dropout as training-time baselines on the DeepONet, and Deep Ensembles and LUNO on the FNO, the architecture LUNO was formulated for. LUNO is a linearized Laplace approximation and the post-hoc counterpart of JCB; since it approximates a Bayesian posterior over the weights, it also serves as our Bayesian baseline (we did not train full BNNs; VB-DeepONet is discussed in Related Work). The post-hoc approach you cite (Bergna et al., arXiv 2502.20966) is now discussed in Related Work. In short, the weight-based bands reach a lower interval score on clean inputs on most benchmarks, while the Jacobian band keeps its coverage better under input noise (Sections 5.2–5.3).

Q2. Section 5.3 addresses this setting with synthetic noise on the input. Each band is calibrated once on clean data, its multiplier is frozen, and coverage is then measured on perturbed inputs under both isotropic and spatially correlated noise. Under this protocol, JCB stays within a few points of its nominal coverage, while the conditional quantile head, whose width is fitted to the model error on clean inputs, loses up to 30 points. We did not evaluate the real datasets of arXiv 2601.01829 or stochastic dynamics, and we now say so in the Limitations; testing JCB on measured noise is the natural next step.

Q3. There is a formal link in the input-noise setting. To first order, the JCB width equals the standard deviation of the output under Gaussian input noise (Eq. 2), so it captures uncertainty propagated from the input, which plays the role of aleatoric uncertainty in measurement settings. It does not capture epistemic (weight) uncertainty, which requires a posterior over the weights, as in LUNO. Since LUNO and JCB linearize the same operator, in its weights and in its input respectively, a single linearized Gaussian process could carry both variances; we discuss this in Section 6.

Reviewer_b87UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present a method for uncertainty quantification in DeepONets. The first method requires a second-order optimization through the branch network, and the second method requires computing jvps for a frozen network. The method seems to improve performance compared to the presented baselines, however, standard methods like dropout or deep ensembles for UQ are not evaluated.

Critical strengths

The method seems to be a principled way to look at uncertainty quantification in DeepONets

Critical weaknesses

The paper is somewhat poorly written. Many concepts are introduced, but never defined (such as PICP, MPIW?). A figure describing the main method would also be beneficial for understanding. The primary weakness (even if the method works well) is that it is only applicable for DeepONet, which isn’t a very powerful backbone (compared to attention for example) The method degrades rollout RMSE, and also requires 2nd order optimization for the Jacobian DeepONet The method seems to depend on tuning to work and it is unclear that is uniformly beneficial.

Questions for authors

Why are other columns not bolded in Table 3? In the AI statement, could you clarify the role of LLMs in writing the paper? The choice of baselines seems to be a bit puzzling. Would it be possible to evaluate something simple and more standard like Monte-Carlo dropout or deep ensembles for UQ?

1 comment

Thank you for your review and suggestions. We address each point below.

Presentation: We agree that the manuscript was hard to follow: several metrics, such as PICP and MPIW, were used before being defined. In the revision, every metric is defined in Section 4, with formulas in Appendix B.3; the abstract and introduction have been simplified; and the results section has been restructured around a single line of argument. Both methods are summarized step by step in Algorithms 1 and 2 (Appendix A).

Table 3: The bold formatting was an error and has been corrected; all tables now bold the best value per block.

Use of LLMs: Gemini was used to turn the experimental designs proposed by the authors into runnable code. Claude Code was used for a reproducibility audit, re-deriving the reported tables from the stored checkpoints and metrics. During the revision, Claude Code was also used to replicate the formulations of the added UQ methods from their original papers, to check whether the new version addresses the critical points raised in the reviews, and to polish paragraphs so that the paper keeps a single line of argument. The authors verified all results and claims against the stored outputs and take full responsibility for the content.

Applicability beyond DeepONets: The reviewer’s concern that the method has only been tested on DeepONets is well taken. We focused on DeepONets because they allowed a more practical implementation through the Gram matrix optimization. That said, for this revision we also tested the method on Fourier Neural Operators, showing that it can be applied to other differentiable neural operators via the JVP, which is available for any differentiable operator. The revision repeats the full post-hoc comparison on the FNO (Table 2, Appendix C.5). There, JCB lowers the interval score of the better constant-shape band by 20% on Burgers and 16% on Navier-Stokes, matches it on Darcy, and has the highest alignment with the reference sensitivity, the same pattern as on the DeepONet. The caveat is the cost: 1.0 to 3.2 s per sample for the exact norm on the FNO, against 0.2 to 0.7 ms with the Gram matrix on the DeepONet. The randomized Jacobian-norm estimator of Appendix C.5 (Bekas et al., 2007), whose cost does not grow with the input dimension, is the route to larger and attention-based operators, which we list as a future direction in Section 6.

Accuracy and second-order optimization: Jacobian-DeepONet raises the MSE of the deterministic operator by 14% on Burgers and does not reduce accuracy on Darcy or Navier-Stokes; for comparison, the conditional quantile head raises it by 21% on Burgers (Section 5.1). The second-order terms make training 3.4x slower on Burgers (Table 10). JCB avoids both issues: it operates on a frozen operator, leaves the mean prediction unchanged, and adds no parameters.

Dependence on tuning: JCB has a single scale, set by split conformal calibration, so it requires no tuning. For Jacobian-DeepONet, the loss-weight sweep over lambda in {1, 4, 8, 20} (Table 7) shows that its interval score stays below that of the constant-width band at every weight, with nearly flat rank correlations. We agree that the method is not uniformly beneficial, and the Limitations now state that on clean inputs the conditional quantile head and the deep ensemble reach a lower interval score on Burgers and Darcy. Our claims are narrowed to what holds across benchmarks: alignment of the width with the solver sensitivity, and coverage retention under input noise.

Baselines: The question about other baselines is true, and we chose to tackle this directly for the review by adding MC dropout and Deep Ensembles to the previous benchmark. The previous baselines were derived from research that combined conformal prediction with DeepONets, thing that is not particularly clear for the other methods. The revision adds Deep Ensembles and MC Dropout on the DeepONet, and Deep Ensembles and LUNO (linearized Laplace) on the FNO, all calibrated with the same conformal procedure (Tables 1 to 3). On clean inputs, the ensemble has the lowest interval score on Burgers and Darcy, helped by a more accurate mean, while JCB has the lowest on Navier-Stokes; MC dropout's mean is 29 to 61% less accurate and its interval score is above that of JCB on all three benchmarks. Under smooth input noise, the ensemble loses about ten points of coverage on Burgers and JCB about three.

Thank you for the review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.