- 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.
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.