acceptodds
Under review as a conference paper at ICLR 2027

Library-based Corrector Network

Abstract

PDEs with boundary, corner, or interior layers exhibit localized multiscale behavior that is hard to capture with standard neural architectures. Classical asymptotic analysis decomposes such solutions into a smooth component and a problem-specific analytical corrector, but designing correctors requires case-by-case analysis and does not transfer across problems. We propose the Library-based Corrector Network (LibCorrNet), which replaces the problem-specific corrector with a shared, family-level structural prior. LibCorrNet combines a Fourier neural operator for the smooth component with a corrector branch that evaluates a fixed candidate-function library on a learned stretched coordinate, combined via adaptive, input-conditioned weights. The same library and architecture are reused across different PDE classes without redesigning a corrector module for each problem. Across several representative classes of singularly perturbed and multiscale PDEs, LibCorrNet consistently outperforms strong neural operator baselines, with the largest gains observed under limited training data. Ablation studies show complementary contributions from coordinate learning, candidate transformation, and adaptive weighting, and the method remains effective even when the exact analytical corrector is excluded from the library, supporting its use as a reusable structural prior rather than a problem-specific design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Layer Problem, PDE, Elliptic PDEs, Corrector function

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NAJkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LibCorrNet, a neural operator for singularly perturbed and multiscale PDEs. It combines an FNO-based global branch with a corrector branch that evaluates a fixed library of functions on learned coordinates and combines the resulting features through spatially varying neural weights. The same library is reused across separately trained PDE configurations, avoiding the need to prescribe a different analytical corrector for each problem.

Experiments cover boundary, corner, and interior layers, nonlinear reaction--diffusion equations, and Helmholtz problems. The authors report the lowest mean relative L2L^2 error in 19 of 20 main settings using approximately 82K parameters. Additional experiments investigate component ablations, candidate-library selection, perturbation-scale adaptation, a circular domain, and oscillatory forcing.

Critical strengths
  1. The architecture offers a practical compromise between generic neural operators and problem-specific analytical enrichment. A fixed candidate library, learned coordinate representation, and adaptive weighting can accommodate different localized structures without redesigning the corrector module for every benchmark. The authors also appropriately acknowledge that intermediate representations are not uniquely identifiable and that the method retains, rather than eliminates, structural prior knowledge.

  2. The empirical study covers several PDE families, multiple training budgets, and ten independent runs, with 1,024 held-out test instances per configuration. LibCorrNet achieves the lowest reported mean error in 19 of 20 main settings using fewer parameters than the compared models. The component ablations, circular-domain experiment, and oscillatory-source tests provide useful additional evidence beyond a single favorable benchmark.

Critical weaknesses
  1. The central layer-resolution claim is not adequately validated on the 52×5252\times52 evaluation grid. Its spacing is approximately 0.020.02, whereas convection-dominated layers at ε=10−5\varepsilon=10^{-5} have width O(10−5)O(10^{-5}). An enriched reference solver does not remove this evaluation bottleneck. Without layer-adapted sampling, convergence checks, or localized and derivative-based errors, the reported global L2L^2 accuracy may largely reflect agreement away from the thin layers.

  2. The benchmark definitions contain material asymptotic inconsistencies. For the stated interior operator −εu′′+(x−x0)u′-\varepsilon u''+(x-x_0)u', homogeneous solutions satisfy u′∝exp⁡((x−x0)2/(2ε))u'\propto\exp((x-x_0)^2/(2\varepsilon)), rather than the Gaussian derivative underlying an error-function transition. Appendix A.0.4 also describes an O(ε)O(\varepsilon) outflow layer for a reaction--diffusion equation without advection, whereas Table 5 uses O(ε)O(\sqrt{\varepsilon}) profiles. These discrepancies leave the implemented equations and reference correctors unclear.

  3. The claimed robustness to excluding the exact analytical corrector is insufficiently demonstrated. Appendix C.4 samples candidates from a pool that still contains exponential and error-function families, but does not identify a controlled exclusion configuration for each target PDE. Comparing random libraries with an Exact ϕ\phi baseline does not establish that performance remains robust when the corresponding analytical profile or function family is deliberately unavailable.

  4. The ablations do not fully isolate the value of the proposed structural prior. Removing candidate transformations compares against an identity transformation, rather than a matched branch using generic nonlinear features. Moreover, the boundary-layer results without ϕ\phi are worse than FNO at every reported training budget, contradicting the broad claim that coordinate learning alone improves performance. More controlled comparisons are needed to attribute the gains to the candidate library.

  5. Input-adaptive coordinate learning is only weakly tested by the main benchmarks. The coordinate branch excludes ff, while ε\varepsilon and bb are fixed within the described configurations; its input therefore remains unchanged as forcing samples vary. Several source constructions also depend on only one spatial coordinate. These settings support learning reusable templates within restricted configurations, but provide limited evidence of adaptation to sample-dependent layer locations, scales, or genuinely two-dimensional structures.

  6. Essential implementation details are missing despite the method's sensitivity to its internal representation. The manuscript does not clearly specify the dimensionality of ξΘ\xi_\Theta, how scalar and two-variable candidates are evaluated on it, the complete weighting-network architecture, or the precise non-positive-domain constraint used for exponentials. The enriched solver's construction and convergence are also insufficiently documented. With code deferred until after review, these omissions materially limit reproducibility.

  7. The statistical evidence does not establish the reliability of several reported rankings. Table 9 provides variability only for LibCorrNet on four families, without corresponding baseline statistics or Helmholtz uncertainty. Some differences are extremely small, such as 0.09790.0979 versus 0.09800.0980 for Helmholtz at NT=128N_T=128. Using official default hyperparameters also does not demonstrate comparable tuning effort, leaving the robustness of the 19-of-20 ranking uncertain.

Questions for authors
  1. At which spatial locations and with what quadrature are the reported L2L^2 errors evaluated? Can the main conclusions be verified using layer-adapted evaluation points, localized errors, and derivative-sensitive metrics? Please also provide convergence evidence for the enriched reference solver and clarify how predictions are evaluated between the original 52×5252\times52 grid points, where most thin-layer variation occurs.

  2. Does the interior-layer implementation use b1=x−x0b_1=x-x_0 or b1=−(x−x0)b_1=-(x-x_0)? Please reconcile the implemented equation with the error-function profile and Figure 8. For nonlinear reaction--diffusion, what boundary conditions, solution-selection procedure, and layer scaling were actually used? Do the inconsistencies in Appendix A affect only the exposition, or also reference generation and baseline construction?

  3. Which precise experiments remove the analytical corrector associated with each target PDE? Please list the retained candidates, exclusion rules, and number of library-selection seeds. Were entire matching function families excluded, or only individual expressions? Also, does the Exact ϕ\phi row represent a controlled LibCorrNet variant or the ComFNO baseline, whose reported values appear identical?

  4. Are models trained separately for each fixed (ε,b)(\varepsilon,b) configuration, making the learned coordinate map identical across forcing samples? What evidence demonstrates adaptation when layer locations or scales vary within one trained model? Can the library-specific benefit be isolated using a matched residual architecture with generic nonlinear features while preserving the coordinate and weighting modules?

  5. Please provide baseline run-to-run variability, paired comparisons where available, and the hyperparameter-selection protocol used for all methods. Could you also specify the coordinate output dimensions, candidate-evaluation rules, weighting-network layers, and exponential constraint? These details would help determine whether the small performance margins and claimed parameter efficiency are reproducible under comparable experimental conditions.

Reviewer_PQ7aOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Library-based Corrector Network (LibCorrNet) for solving PDEs with localized multiscale structures. The method combines an FNO-based global branch with a candidate-function corrector branch, using learned coordinate transformations and adaptive weights to reduce reliance on problem-specific analytical correctors. Experiments cover five PDE families, and LibCorrNet achieves the lowest relative L2L_2 error in 19 out of 20 main settings.

Critical strengths
  1. The paper addresses a meaningful problem and offers an interesting way to incorporate classical asymptotic correction ideas into neural operators.
  2. LibCorrNet achieves strong performance across several PDE benchmarks, particularly in data-limited settings.
  3. The paper includes component ablations, candidate-library analysis, and cross-regime experiments.
Critical weaknesses

Several expressions in Appendix Table 5 appear inconsistent with the governing equations. For example, when b1=−1b_1 = -1, the stated exponential corrector grows with xx and does not satisfy the corresponding homogeneous convection–diffusion equation. The interior-layer corrector lacks a coordinate shift associated with the layer location x0x_0. Furthermore, the nonlinear reaction–diffusion boundary-layer thickness is described as ε\varepsilon, while the corresponding analytical expressions use a ε\sqrt{\varepsilon} scale. Equation (24) also does not fully specify the boundary conditions on the two-dimensional domain. These issues should be corrected, and their potential impact on the numerical experiments and analytical baselines should be clarified.

Questions for authors

Some references contain author information inconsistent with their publication records, and a few citations are only indirectly related to the claims they support. The appendix also contains “LaTeX” artifacts, incorrect figure references, and duplicated computational-cost discussions.

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

LibCorrNet combines an FNO global branch with a coordinate-learning branch. A fixed library of candidate functions is evaluated on the learned coordinates, and spatially adaptive neural weights combine these features into a correction added to the global prediction. The same library and architectural form are reused across separately trained PDE benchmarks.

Experiments cover boundary, corner, and interior layers, nonlinear reaction–diffusion, and Helmholtz problems. The paper reports the lowest mean error in nineteen of twenty main settings, with additional component and library-size ablations, repeated-run statistics, a circular-domain example, and fine-tuning across perturbation scales.

Critical strengths
  1. The motivation describes a meaningful trade-off between strong but problem-specific analytical structure and generic learned representations. Reusing a family-level library while learning its realization offers a practical intermediate design. The paper also correctly defines reuse as applying the architecture with retraining, rather than claiming zero-shot transfer between unrelated PDEs.

  2. Learning coordinates is a substantive part of the idea. A correct functional profile evaluated at an inappropriate location or scale need not represent a layer well; separating the candidate function from its spatial realization follows naturally from stretched-coordinate asymptotics. This gives the coordinate branch a clearer rationale than simply adding another feature pathway.

  3. The reported study supports practical potential across the chosen families. Multiple PDE classes, training-set sizes, ten independent runs, component ablations, and qualitative internal-feature analyses provide useful evidence. The authors appropriately state that a learned global/corrector decomposition is not uniquely identifiable, which prevents the visualizations from being treated as exact analytical recovery.

Critical weaknesses
  1. The central insight is not sufficiently distinguished from a flexible nonlinear correction branch. The classical motivation involves a smooth component plus a structured, localized corrector, but the network imposes no separate smoothness or localization requirement and allows spatially varying weights together with learned coordinates. Such a branch can represent general prediction error without implementing the intended asymptotic separation. No theorem is necessary to make the architecture useful, but the paper needs evidence or a precise argument that the chosen candidate families contribute something beyond generic learned features. The current component ablations establish utility of the parts, not that particular explanation.

  2. The scope of the reusable prior is not fully characterized. Replacing an exact profile with a fixed function pool reduces one form of manual design, but the pool is still chosen around the considered families, and Appendix B uses problem-dependent forcing constraints. Appendix C.4's random subsets do not clearly establish a controlled exclusion of the relevant profile family. The useful insight would be to show what structural knowledge must be supplied, what the learned coordinates can absorb, and where the shared library ceases to be appropriate. Retraining across related benchmarks alone does not answer that question.

  3. The evaluation is not yet aligned with the physical difficulty used to motivate the method. The main singular-perturbation setting combines epsilon=10^-5 with a uniform 52×52 learning grid. High-fidelity enriched reference solutions are valuable, but coarse-grid global L2 error can remain small while missing the internal shape or flux of a layer much narrower than the grid spacing. This does not show that LibCorrNet fails to represent layers; it means the reported advantage has not yet verified the specific layer-resolution capability that would make the corrector insight compelling.

Questions for authors
  1. What distinguishes the proposed corrector-family prior from a capacity-matched generic correction branch? Please explain what structural property is retained by the network and which existing evidence links the gains to that property rather than general feature flexibility.

  2. Which prior knowledge and problem-dependent preparation are required to reuse the model, and what exactly is excluded in the claimed profile-removal study? Can the existing library analyses delimit the generality of the insight rather than only show that several candidate subsets work?

  3. Can the enriched reference solutions support evaluation within the thin layers, using dense profiles or a layer-/flux-sensitive measure? Would the observed rankings then support the physical capability that motivates the architecture, as well as accurate samples on the 52×52 grid?

1 comment

W2 & Q2. Reusability and Required Prior Knowledge

The forcing functions in the nonlinear reaction–diffusion and boundary-layer problems are chosen to satisfy the prescribed periodic boundary conditions. For the interior-layer problem, the compatibility condition (f(0)=0) ensures that the limiting equation remains consistent at the turning point, where (b(0)=0) and (b'(0)>0), and admits the asymptotic corrector profile considered in our paper. In contrast, when (f(0)≠0)(f(0)\neq0), the limiting problem becomes singular at the turning point and requires a different treatment, such as the lifting-based correction discussed in the existing numerical analysis literature. These conditions characterize the asymptotic structure of the benchmark problems rather than restrictions imposed by LibCorrNet. We clarify this distinction and the periodic boundary conditions in Appendix B of the revised manuscript and will add the corresponding references.

As shown in Table 6, all benchmarks use the same candidate library, architecture, and hyperparameters, following the common input specification ((x,f,b(x),ε))((\mathbf{x},f,\mathbf{b}(\mathbf{x}),\varepsilon)). Models are trained separately for each PDE, but no PDE-specific analytical correctors or stretched coordinates are prescribed. Learned Fourier coordinates and spatial weights adapt the shared candidates to different layer locations and scales.

Table 2 evaluates candidate transformations, while Appendix C.4 examines candidate subsets. We agree that random subsets alone do not establish robustness to removing an entire relevant profile family. Therefore, we conducted controlled experiments excluding the analytical corrector corresponding to each PDE, as well as its entire associated function family. The results have been added to Table 3 of the revised manuscript.

Our claim is reduced dependence on PDE-specific analytical correctors, not elimination of asymptotic prior knowledge or zero-shot transfer. The demonstrated reusability is limited to the elliptic-type PDE benchmarks considered.

W3 & Q3. Layer-Adapted Evaluation and Reference Validation

We agree that the original (52\times52) grid, although consistent with previous studies [1, 2] and ComFNO, cannot adequately resolve (O(\varepsilon)) boundary and corner layers for (\varepsilon=10^{-5}). Uniform resolution with (h=O(\varepsilon)) would require quadratically increasing grid points in 2D, resulting in substantial computational costs. However, the interior-layer problem exhibits a distinguishable transition even for small (\varepsilon) on the original (52\times52) grid, as illustrated by the reference solution in Figure 1 ((https://imgur.com/a/Mq0kxyR)).

We therefore performed additional evaluations for boundary and corner layers using nonuniform grids, adding evaluation points at the (10^{-3}), (10^{-4}), and (10^{-5}) spatial scales near the relevant boundaries. We focused on these problems because the interior-layer and nonlinear reaction–diffusion structures are already distinguishable on the original grid.

Relative errors and maximum absolute errors ((\times10^{-2})); 800 training and 120 test samples.

ProblemModelFull Rel (L^2)Layer Rel (L^2)Full Grad Rel (L^2)Layer Grad Rel (L^2)Max Abs Error
BoundaryLibCorrNet0.460.441.901.9015.45
CNO0.550.501.050.9219.67
FNO0.550.522.042.0023.03
CornerLibCorrNet4.585.1210.4111.5430.61
CNO5.477.158.198.6231.42
FNO4.925.118.238.6734.58

LibCorrNet achieves the lowest full-domain relative (L^2) errors in both problems and the lowest localized error for boundary layers. For corner layers, its localized error is lower than CNO's and comparable to FNO's. It also achieves the lowest maximum absolute errors in both problems. Figures 8 and 9 ((https://imgur.com/a/Mq0kxyR)) visualize the maximum absolute error at each evaluation point across 120 test samples for the boundary-layer and corner-layer problems, respectively, including magnified views of the thin-layer regions. The experimental results and corresponding visualizations have been added to Section 4.2 of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.