acceptodds
Under review as a conference paper at ICLR 2027

Residual-Conditioned Subspace Correction for Helmholtz Neural Operators

Abstract

Helmholtz discretizations can share Poisson’s sparsity pattern while exhibiting indefiniteness and near-resonant sensitivity, motivating neural-operator designs informed by matrix properties. We propose residual-conditioned subspace correction (ReSCo), which combines a pretrained multigrid neural operator with a fixed POD error basis and a learned coefficient-recovery module. The module uses physical residuals and projected features to predict global wavefield corrections. To improve generalization, we train it on samples excluded from backbone training. On the Caltech Helmholtz validation benchmark across three random seeds, the method reduces mean relative error from 0.00998 to 0.00619, achieving a mean per-seed improvement of 37.27%, compared with 6.2% when correction training uses backbone-seen samples. These results demonstrate the value of combining matrix-informed correction with training on residuals.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, Helmholtz equation, Residual-conditioned subspace correction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YumQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Residual-Conditioned Subspace Correction (RESCO), a supervised method for improving Helmholtz predictions from a frozen neural operator. It constructs a fixed POD basis from normalized prediction errors on the backbone's training data. A lightweight network predicts correction coefficients using the medium, frequency, backbone prediction, physical residual, and POD projections. The resulting correction is added to the backbone output. The correction head is trained on a separate labeled dataset excluded from backbone training. Its targets are the coefficients of the solution error projected onto the POD basis, while the physical residual is used as an input.

The analysis separates the remaining solution error into an error outside the POD space and an error in the predicted coefficients, and examines why residual-based reduced solves may not minimize field error. Experiments on the Caltech Helmholtz benchmark use a frozen MgNO backbone and report a 31.20% mean per-seed reduction in relative L2 error on the original test split. The study measures inference costs and examines the correction basis, residual inputs, coefficient learning, operator features, and the data used for correction training. The paper also tests the corrected predictions as initializations for GMRES. In the tested solver configuration, this improves final field accuracy relative to MgNO-initialized GMRES, without reducing the iteration count or total online solution time.

Critical strengths
  1. Clear motivation and useful error analysis. The paper explains why a smaller PDE residual does not necessarily imply a more accurate Helmholtz solution. Its error decomposition separates errors outside the POD space from errors in the predicted coefficients. Together with the analysis of reduced solves, this helps explain why a suitable correction space does not guarantee that residual-based methods recover the coefficients that minimize field error.

  2. Promising results and useful ablations. Using additional labeled correction data, RESCO improves the frozen MgNO on the Caltech Helmholtz benchmark, with a reported 31.20% mean per-seed reduction in relative L2 error on the original test split. The experiments compare POD and random bases, inputs with and without physical residuals, learned and constant coefficients, and correction training on backbone-seen and backbone-unseen examples. These comparisons, including the matched generic POD/residual head, help show how the main design choices affect performance in the tested setting.

  3. Clear reporting of inference costs and limitations. The paper reports trainable head parameter counts and batch-amortized inference times that include the backbone and the required physics, feature, and correction computations. The comparison with the operator-feature extension shows the extra inference cost associated with its accuracy gains. The GMRES experiments, spectral analysis, and transfer tests also provide useful evidence about where the correction helps and where its benefits are limited.

Critical weaknesses
  1. A comparison with a full-field correction network is missing. RESCO improves a frozen MgNO using extra labeled data, but there is no comparison with a network that directly predicts a full-field correction. The generic correction head still uses the POD space, so it does not show whether restricting the correction to this space is beneficial compared with predicting a full-field correction. A small full-field correction network using the same frozen backbone, extra labels, data splits, and physical residuals would be a useful baseline. Comparing accuracy and training and inference costs would help assess the benefit of the POD design.

  2. The contribution beyond existing correction methods needs clearer explanation. The method combines a fixed POD error basis, physical-residual inputs, and correction training on examples not seen by the backbone. The ablations show that these choices are useful in the tested setting. However, the paper could explain more clearly what this combination adds to existing supervised correction methods. The core error decomposition is useful, but follows from standard orthogonal projection arguments. Relevant methods such as MF-WNO (Thakur et al., 2022) and IRNO (Liu et al., 2026) are not discussed. Although their original settings differ, comparing what each method predicts, what inputs it uses, and how it is trained would make RESCO's contribution easier to assess.

  3. Two claims in the appendix need further explanation. In Appendix B.6, Eq. (18) does not specify the function class defining the integral probability metric, explain the assumptions needed to control the squared coefficient loss, or define the complexity term. The equation also does not by itself show that training on backbone-unseen errors makes the training and evaluation distributions closer. It is unclear whether the authors intend this statement as a theoretical result or as a motivation for the training protocol. In Appendix B.4, near resonance does not necessarily make all three reduced quantities deteriorate. Their behavior depends on subspace alignment, and the reduced problem can remain well-conditioned even when the full operator is ill-conditioned. The core error decomposition remains valid, but these two claims need clearer assumptions or narrower wording.

Questions for authors
  1. Comparison with a full-field correction baseline. Would it be possible to compare RESCO with a compact full-field correction network using the same frozen MgNO, additional labeled examples, and validation/test splits? Giving this baseline access to the same physical-residual information would help clarify the role of the fixed POD correction space. Reporting field accuracy, additional training cost, and inference cost would help assess the practical benefits of this design.

  2. Relationship to related supervised correction methods. Could the authors explain how RESCO differs from MF-WNO (Thakur et al., 2022) and IRNO (Liu et al., 2026)? A brief discussion of what each method predicts, what inputs it uses, and how it is trained would help clarify RESCO's contribution. This discussion could also note differences in the original tasks and training data.

  3. Clarification of the distribution-shift bound. Could the authors clarify the definitions and assumptions behind Eq. (18) in Appendix B.6? In particular, which function class defines the integral probability metric, under what assumptions does it control the squared coefficient loss, and what does the complexity term represent? Does the analysis show that using backbone-unseen examples makes the correction-training distribution closer to the evaluation distribution, or is this intended as an empirical motivation for the training protocol?

  4. Clarification of the near-resonance discussion. Appendix B.4 states that all three reduced quantities deteriorate near resonance. Could the authors clarify how this statement depends on the alignment of the correction subspace with sensitive modes? The full operator can be ill-conditioned while the reduced problem remains well-conditioned. Explaining the relevant conditions and the role of regularization in the reduced solves would help relate this statement to the more cautious discussion in Section 3.

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

This paper studies how to improve the field-prediction accuracy of a pretrained Helmholtz neural operator. The authors propose Residual-Conditioned Subspace Correction (RESCO): the MgNO backbone is frozen, a fixed rank-16 POD basis is constructed from the backbone's prediction errors on its training data, and a small coefficient-prediction head is trained to output a low-dimensional correction. The head uses the medium, frequency, initial prediction, physical residual, and POD projection of the residual as inputs. It is trained on additional labeled data that were not seen by the backbone during its training, with the goal of learning the backbone's generalization error. The final prediction is the sum of the backbone output and the correction in the POD subspace.

Critical strengths

A measurable accuracy gain in the explicit frozen-backbone setting. The main method trains a correction head and obtains an approximately 31% mean per-seed reduction in field error on the original test set. The authors report online cost including the physical-feature computation, and they disclose the negative results that the physical residual does not improve, GMRES is not accelerated, and transfer to changed settings is limited. These experiments have value as an evaluation of post-processing for a frozen pretrained model in this specific setting.

Critical weaknesses

1. No equal-total-label-budget comparison, making the relative value of the correction strategy difficult to assess. The main method is trained on additional labeled data that were not seen by the backbone, while the principal comparison remains the originally frozen MgNO. RESCO-SEEN matches the correction head's sample count and training budget and is useful for studying the source of the training error, but it does not remove the additional-label confound when RESCO is compared with the unmodified backbone. The paper does not compare against MgNO fine-tuned on the same additional labels, or against other lightweight adaptation methods using the same label budget.

2. The gains are mainly confined to field prediction under a fixed setting, and transfer experiments show a limited range of applicability. The main experiments focus on one data source, one backbone, and a real-valued Helmholtz problem at a fixed frequency. The paper does report grid-, frequency-, and source-count-transfer experiments, but the results show that, when the grid changes, the learned coefficients may fail to transfer reliably even when the oracle still indicates useful subspace capacity; when the frequency changes, the representational capacity of the fixed POD basis itself declines substantially; and when the number of sources increases, relative gains may remain while the absolute field error becomes very large.

Questions for authors

Q1 Since RESCO introduces additional labeled data, could the authors fine-tune MgNO, or train another lightweight adaptation module, using the same additional-label set and training budget, and report accuracy, training cost, and online cost? Please also disclose the number and selection rule of the additional correction-training samples.

Q2 Was the rank of the POD basis set to 16 empirically, or is it supported by theory or development-set experiments? Please report how different ranks affect oracle representation capacity, actual coefficient-recovery accuracy, the amount of training data required, and cost.

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

The paper adds a small correction module on top of a frozen, pretrained MgNO for the Helmholtz equation. The backbone’s error is corrected only within a fixed 1616-dimensional POD subspace, built from the backbone’s errors on its own training data. A small network (∼144\sim 144k parameters) predicts the coefficients in this subspace from the medium, the backbone prediction, and the physical residual b−Ahu0b - A_h u_0. The authors also give an error decomposition to explain why classical residual-based ways of computing these coefficients (ridge, Galerkin) fail, and why learning them directly works better. The main empirical point is that the correction head should be trained on samples the backbone has not seen.

Critical strengths
  1. Training the corrector on backbone-unseen errors gives a clear and well-controlled gain. When the correction head is trained on samples the backbone has not seen, the error reduction grows from about 66% to 31–3731–37%. The comparison uses the same head, the same amount of data, and the same training budget, so the effect can be attributed to the training data itself.
  2. The representation/coefficient error decomposition explains clearly why residual-based coefficient recovery (ridge, Galerkin) fails, and the experiments confirm it. The ablations include a random basis, constant coefficients, and an oracle upper bound. The authors also openly report the negative results: no residual reduction, no solver speedup, and poor transfer.
Critical weaknesses
  1. ReSCo uses an extra labeled set that MgNO never sees. There is no baseline with the same amount of data, for example MgNO fine-tuned on it. So it is unclear how much of the 31–3731–37% gain comes from the method and how much from the additional labels.
  2. The physical residual does not decrease. In the GMRES study, starting from zero is both faster and more accurate than starting from the corrected prediction. The correction also does not transfer to other grids or frequencies.
  3. The paper does not seem to cite or discuss FCG-NO [1]. It trains a neural operator on residuals to approximate the error and uses it inside flexible CG. It also shows that training residuals must match the deployment distribution, which is very close to the seen/unseen finding here. Following FCG-NO, the head could be used as a nonlinear preconditioner inside flexible GMRES, not only as an initial guess. This might address the lack of solver speedup.

[1] Rudikov, A., Fanaskov, V., Muravleva, E., Laevsky, Y. M., & Oseledets, I. (2024). Neural operators meet conjugate gradients: The FCG-NO method for efficient PDE solving. arXiv preprint arXiv:2402.05598.

Questions for authors

See the weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.