acceptodds
Withdrawn from ICLR 2027

Boundary-Action Equivariant Neural Operators for Sparse-Sensor PDE Inversion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physical transformations in sparse-sensor PDE inversion act jointly on the domain, boundary data, sensor coordinates, observations, and solution field. Boundary concatenation and soft equivariance penalties do not enforce one coherent transformation law across this inverse-problem interface. We introduce boundary-action equivariant neural operators whose lifting, latent blocks, and readout are constructed as intertwiners for a physically admissible group, and retain the smallest singular value of the physical sensor-map Jacobian as a separate observability diagnostic rather than conflating symmetry with identifiability. On PDEBench Darcy flow with held-out D4 orbit elements, BC-EQUIVOP obtains 4.7% relative L2 error, 3.9% boundary error, and 2.1% commutation error, compared with 8.7%, 10.6%, and 9.9% for an FNO with boundary channels. Its 16.2-ms latency versus 7.4 ms for that FNO quantifies the computational price of structural coherence. The results support block-level boundary–sensor structure as the operative inductive bias; the measured 2.1% end-to-end residual is not a claim of exact implementation-level equivariance, and observability remains an independent limitation.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural Operators, PDE Inversion, Equivariant Learning, Group Equivariance, Scientific Machine Learning, Sparse Sensors, Physics-Informed Learning, Inverse Problems, Observability

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

The paper proposes a neural operator for sparse-sensor PDE inversion that is designed to respond consistently when boundary conditions, sensor locations, observations, and outputs are transformed together by a symmetry (e.g., rotations and reflections of a square domain). It also proposes a separate diagnostic for whether the sensors carry enough information. Experiments on one Darcy-flow benchmark report lower error than several neural-operator baselines, at a higher computational cost.

Critical strengths

The motivation is reasonable: treating boundary data, sensors, and outputs under one consistent symmetry is a sensible goal for inverse problems.

Critical weaknesses
  1. The method is not described clearly enough to understand or evaluate. The paper explains the architecture only at a conceptual level and does not specify how the model is actually built. As a result, I could not understand how the method works, why it should work, or what is new compared to existing equivariant neural operators. This is a fundamental issue rather than a minor clarity problem: without a concrete description, the contribution cannot be assessed or reproduced.

  2. The paper is not mature enough for the main track. With about five pages of main text, it reads more like a workshop paper or research proposal: the evaluation plan is described in detail, but results are reported for a single benchmark, and most of the experiments the paper itself identifies as necessary to validate its claims are missing. Addressing this would require substantial new work, not a revision.

Questions for authors
  1. Could you describe the architecture concretely (e.g., how sensor data and boundary conditions are fed into the model, and what the main layers are)?
  2. What exactly is new compared to existing group-equivariant neural operators such as G-FNO, or EqGINO, which uses orbit-based kernels? Since EqGINO is listed as a baseline in your experimental protocol, please also report its results.
  3. Can you report the experiments described in your evaluation plan that are missing from the results?
Reviewer_sPqvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a neural operator for sparse-sensor PDE inversion that jointly transforms boundary data, sensor locations, observations, and outputs under compatible symmetries. It also proposes a physical-Jacobian diagnostic for observability. Darcy-flow experiments report improved prediction accuracy and reduced equivariance errors, with increased inference cost.

Critical strengths

Treating boundary data, sensors, and observations under a consistent joint action is a useful design principle for symmetry-aware inverse problems.
The paper clearly distinguishes equivariance from identifiability and acknowledges the limitations of its reported evidence.

Critical weaknesses

The method is insufficiently specified: concrete group actions and layer constructions are missing. Proposition 1 assumes equivariant components without establishing how the proposed implementation achieves them, leaving the technical contribution unclear.

Experimental support is limited to one aggregate Darcy table and brief ablations. Matched-augmentation controls, observability diagnostics, and variability across the stated five seeds are not reported.

Questions for authors

What are the explicit group actions and constructions of the sensor lifting, operator blocks, and readout? Which implementation steps account for the reported 2.1% equivariance error?

Can you provide the matched-augmentation comparison and observability evaluations described in the protocol, together with variability across the stated five seeds? These results are necessary to assess the claimed benefits.

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

The paper studies sparse-sensor PDE inversion, where a solution field or an unknown coefficient is reconstructed from limited observations and known boundary data. Its central argument is that physical transformations must be applied consistently across the inverse problem. For example, rotating the domain requires corresponding transformations of the boundary data, sensor locations, observations, and predicted field.

BC-EquivOp consists of a sparse-sensor lifting, equivariant latent blocks, and a readout. Each component is required to intertwine the corresponding group representations; if these conditions hold exactly, their composition is equivariant. Training combines an inverse-task loss with an optional physics residual, while equivariance is intended to follow from the architecture. The paper also proposes a separate diagnostic based on the physical sensor-map Jacobian to assess weak local observability, while acknowledging that its effectiveness has yet to be validated empirically. The results are promising, but the current evidence does not isolate the proposed boundary-sensor mechanism from other differences between models.

Critical strengths

I find the motivation convincing. When a sparse inverse problem is transformed, the measurements and boundary information must remain consistent with the field being reconstructed. Treating these inputs together is a useful way to formulate the problem. The paper also makes a sound distinction between equivariance and observability: a model can respect the correct symmetry and still lack enough measurements to identify the unknown. The reported results are encouraging, and including boundary error, commutation error, and latency alongside reconstruction error helps assess the practical trade-off. The no-physics variant and component ablations are relevant to the proposed explanation, although the evidence is incomplete.

Critical weaknesses
  1. I could not reconstruct the proposed architecture from Sections 4.1-4.2. These sections state the properties that the lifting, latent blocks, and readout should satisfy, but do not define the maps used in the experiments. The alternatives offered, such as group pooling or regular-representation channels, and convolution or attention, leave the implementation unclear. Proposition 1 is correct as a conditional statement, but it assumes the equivariance of the components that the method needs to establish. Explicit definitions of the boundary encoding, sparse lifting, nonlinear operations, and readout would help assess both the construction and its contribution beyond an appropriately adapted G-FNO.

  2. The main experiment needs a more precise description of the physical problem. Section 3 allows the output to be either a solution field or an unknown coefficient, while Table 2 does not say which is reconstructed. The Darcy equation, boundary and source distributions, sensor configuration, and corresponding D4 actions are also unspecified. The boundary conditions are particularly relevant here. If they are homogeneous Dirichlet throughout, for example, transforming them leaves them unchanged. Such a setting would provide limited evidence for the benefit of handling nontrivial boundary transformations. I would like the authors to clarify what boundary data are used and how the experiment tests this part of the method.

  3. The reported comparisons do not yet isolate the source of the improvement. Section 6 identifies matched augmentation as a decisive control, but its results are absent. It is also unclear how sparse observations are supplied to G-FNO and whether that baseline uses the same physics loss. The paper states that capacity and training budgets are matched, without providing the corresponding values. Although at least five seeds are mentioned, Table 2 contains no variability estimates. Finally, the split should specify both the underlying PDE instances and the held-out group elements, together with the transformations allowed during training. These details are needed to judge whether the gain comes from the proposed joint action and whether it is consistent across runs.

  4. The connection between the structural claim and the implemented model needs further evidence. The authors appropriately report the nonzero 2.1% commutation error, but do not define the metric or identify where the residual arises. Tests of the lifting, latent blocks, and readout before and after training would make this connection clearer. There is also a conceptual issue with H4. Breaking a symmetry of the physical problem does not necessarily remove the network's algebraic equivariance. A network built from exact intertwiners can still commute under its declared action even when its predictions are physically inaccurate. The proposed control should distinguish architectural commutation, physical symmetry, and prediction accuracy; it should not require commutation or every predictive benefit to disappear when the physical symmetry is broken.

Questions for authors
  1. What is the reconstruction target in Table 2, and what are the explicit actions on the boundary data, coefficients, sensor coordinates, and observations? Does the experiment include boundary data that change nontrivially under these actions?
  2. Can the authors provide the implemented lifting and readout, along with component tests that account for the 2.1% commutation residual?
  3. How does the method compare with matched augmentation and with G-FNO using matched input processing and physics regularization? Please include variability across seeds and the exact instance and group-element split.

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

Related papers

Discussion (0)

Sign in to comment.