acceptodds
Under review as a conference paper at ICLR 2027

Response Geometry for Decisions with Neural PDE Surrogates

Abstract

Neural PDE surrogates can fit familiar solution fields while responding incorrectly to changes in physical parameters. We study whether this missing response information improves three decisions: choosing a forward predictor, choosing a surrogate for parameter recovery, and repairing an existing predictor. Response geometry compares parameter derivatives of model and reference fields. An exact pathwise error identity connects these derivatives to extrapolation; sensor projection explains local inverse sensitivity, and the rollout chain rule explains why one-step derivative fit can miss accumulated error. We evaluate these connected uses on the one-dimensional viscous Burgers equation, the two-dimensional steady Darcy flow equation with heterogeneous permeability, and the two-dimensional forced incompressible Navier-Stokes equations in vorticity form, separating ID-only probes, shift-calibrated scores, and target-assisted comparisons. A prespecified response selector reduces early-time Navier-Stokes OOD relative field error from 0.9934 to 0.3377, while its Burgers counterpart ranks population risk better but selects worse. Completed repair tests reduce Navier-Stokes field error from 0.6284 to 0.4899 using rollout-plus-response supervision, with a corrected paired-test gain and additional compute disclosed. The same repair worsens inverse parameter recovery. Burgers shows smaller path-repair gains; Darcy's reliable small gain is a direct-field control. These findings make response geometry useful as a task-specific decision variable: its information can improve selection and prediction, but field accuracy, inverse recovery, and long-time physical fidelity require separate tests.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, response geometry, out-of-distribution generalization, model selection, inverse problems, sensitivity analysis, surrogate model repair, scientific machine learning

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

This paper investigates the "parameter response" of neural operators. The authors define a parameter response Jacobian and a relative response geometry error and use those to organize three decisions: selecting a frozen forward model under a protected parameter shift, selecting which frozen model to invert from sparse sensors, and "repairing" a pretrained predictor with physics, tangent, and rollout-response losses. The supporting "theory" is a pathwise telescoping identity (the fundamental theorem of calculus), a Taylor curvature bound, a Gauss--Newton/Fisher linearization, and the tangent-linear chain rule. Experiments are carried out on Burgers, Darcy and Navier-Stokes equations. The main finding is that a shift-calibrated response score reduces Navier--Stokes OOD field error from 0.99340.9934 to 0.33770.3377.

Critical strengths
  1. The evaluation protocol is frozen before the protected outcomes, and the negative results are reported rather than hidden.
  2. The rank-vs-top-one dissociation on Burgers is a genuinely useful observation: better population ranking does not imply better model selection.
Critical weaknesses
  1. The headline forward result is confounded, and the missing control is decisive. The response selector is tuned on labeled shifted-development outcomes, while the baseline is a strict in-distribution field-L2 rule. No shift-calibrated field-only control is reported. The winning candidate is also a different family (SC-FNO vs. DeepONet). Thus the abstract's 0.9934→0.33770.9934 \to 0.3377 cannot be attributed to response geometry; it conflates using shifted labels, using response features, and switching architectures. This is the paper's only large positive selection result, so the framing does not stand.
  2. The inverse evidence is null or exploratory. The prespecified sensor-geometry score improves inverse log-parameter RMSE from 0.86070.8607 to 0.69110.6911 but does not survive multiplicity correction. The compact score that looks strong reaches 0.06240.0624---numerically identical to the oracle---with p≃0.0622p \simeq 0.0622 after being selected post hoc, and the authors themselves call it exploratory. There is no confirmatory inverse result, and the repair experiment makes inverse recovery worse.
  3. The statistical and reproducibility record is inadequate for the claims. Tables 2 and 3 report no intervals and no seed replication; the 3636 candidates per PDE are individual (family,capacity,seed)(family, capacity, seed) checkpoints (108/3=4×3×3108/3 = 4 \times 3 \times 3), not seed-averaged ensembles; and the forward headline rests on a single parent-level sign-flip test over 1212 parents (p≃0.0146p \simeq 0.0146). No architectures, loss weights, optimizer, seed integers, or compute table are given, the relative-OOD denominator is never defined, and the inverse failure fractions appear only in prose. Tables 2--4 are also mislabeled or incomplete (the "Response score" column holds OOD error, Table 3's failure column is ambiguous, Table 4 omits the field-control inverse RMSE), and the custom, unreleased generator makes the numbers externally incomparable.
  4. Novelty: This is standard sensitivity analysis repackaged as a decision framework, and the experiments do not support the repackaging. Eq. 3 is the fundamental theorem of calculus, Eq. 4 is Taylor's theorem, Eq. 5 is Gauss--Newton/Fisher linearization, and Eq. 10 is the chain rule; the authors concede there is no new theorem. There are also two important uncited works: sTCL (Yang, Lu & Mou, arXiv:2609.32797), which enforces derivative consistency on the fly without offline tangent labels, and PhysGuard (arXiv:2606.16602), which uses a Fisher information matrix to protect physics-critical directions during fine-tuning of a neural operator. Both narrow the repair contribution. What remains is a decision framing, and its central selection claim is confounded.
Questions for authors
  1. What happens if you run the same selector with the response features removed, tuned on the same labeled shifted cases? If that field-only selector also approaches 0.33770.3377, the headline gain is not about response geometry.

  2. The response selector picks an SC-FNO while the baseline picks a DeepONet. What is the top-one result within each family? Without that comparison, the gain may be an architecture switch, not a response-geometry effect.

  3. Four non-isolating repair arms, including a random-reference arm, pass the same corrected field test, the winner is not equal-compute, and the compact inverse score is exploratory, equals the oracle at 0.06240.0624, and has p≃0.0622p \simeq 0.0622. What claim remains that is both confirmatory and specific to response geometry?

Reviewer_Au9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces response geometry, which measures the discrepancy between the parameter derivatives of a neural PDE surrogate and those of the reference solution operator. The paper provides a pathwise error identity, an analysis of sensor-conditioned inverse sensitivity, and a rollout chain-rule argument, followed by experiments on the Burgers, Darcy, and Navier–Stokes equations. The results suggest that response information can improve model selection and forward prediction in certain settings.

Critical strengths
  1. This paper reveals that, in neural opertors, low in-distribution field error does not necessarily imply reliable behavior under changes in physical parameters and propose that considering parameter responses in addition to field accuracy is meaningful for scientific computing, which a novel perspective.

  2. The discussions and analyses of pathwise error accumulation, sensor-conditioned sensitivity, and rollout dynamics are conviencing and valuable.

  3. The evaluation covers three PDE systems and multiple neural operator families, providing comprehensive results for their pathwise error identity.

Critical weaknesses
  1. This paper is hard to follow, need to improve the readiblaty, for example what is 'reference solves' and what are L~ID\tilde{L}_{ID} and C~m\tilde{C}_{m}? I even do not find LIDL_{ID}.

  2. Its main contribution is just to organize existing response diagnostics and related training objectives into a framework for model selection and repair. And the experiments show that the proposed identity does not yeild consistent improvements.

  3. Currently, neural operators are increasingly moving toward more general-purpose, pretrained models that aim to handle diverse PDEs, physical parameters, and problem settings within a unified framework. In this context, the paper's focus on selecting among a fixed set of candidate surrogates or repairing existing checkpoints may limit its broader impact.

Questions for authors

See weaknesses.

Reviewer_7DznOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether parameter-response information helps select and improve neural PDE surrogates. It compares surrogate and reference derivatives, then uses these measurements for forward-model selection, sensor-based parameter recovery, and predictor repair. Experiments on Burgers, Darcy, and Navier-Stokes equations show that the benefit depends on the task. In particular, better forward prediction does not necessarily lead to better inverse recovery.

Critical strengths
  1. Evaluating forward selection, inverse selection, and repair together is useful. The paper makes clear why field error, derivative error, and sensor-conditioned inverse accuracy should not be treated as interchangeable measures.

  2. The negative results are informative. On Burgers, a better overall ranking still leads to a worse selected model. On Navier-Stokes, repair improves field accuracy while worsening parameter recovery (Tables 2-4). Reporting these outcomes makes the empirical conclusions more credible.

Critical weaknesses
  1. The forward-selection experiment does not sufficiently isolate the contribution of response features. Equation (6) uses labeled shifted-development outcomes to fit a score containing ID field error, model capacity, and response features. The main baseline in Table 2 selects by ID field error without this calibration. The Navier-Stokes gain is large, but this comparison cannot tell us how much comes from response information rather than access to shifted-development labels and score fitting.

  2. The repair results leave a related question. Table 4 includes a useful rollout-only control, with error 0.5263 versus 0.4899 for rollout plus response. However, the main reported gain is against field-only training, and the paper does not establish the separate derivative contribution under a matched compute budget. Section 5 explicitly acknowledges the extra training cost. This matters because a simpler field- or rollout-based procedure could be preferable when given the same resources.

Questions for authors
  1. How does the selector perform with response features removed, while keeping ID field error, capacity, development labels, and the fitting procedure unchanged? Please also compare with direct selection by development-set field error. The important outcomes are the selected model's error and selection regret, not only ranking correlation.

  2. For the Navier-Stokes repair, what is the paired improvement from adding response supervision to the rollout-only control across the same starting checkpoints and evaluation cases? How does this comparison change when additional training time and reference-solver work are accounted for?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.