acceptodds
Under review as a conference paper at ICLR 2027

Geometry–Condition Factorized Neural Operators for Electromagnetic Modeling

Abstract

Electromagnetic forward modeling is fundamental to applications ranging from mineral exploration to cerebral electrical impedance tomography (EIT), yet repeated high-fidelity simulations remain computationally expensive. Starting from the Maxwell-governed forward operator, we analyze its local Fr\'echet sensitivity with respect to the three-dimensional electrical-property field. The resulting sensitivity exhibits a pronounced geometry–condition organization: the corresponding unfolding shows substantially stronger spectral concentration than alternative variable partitions, indicating that observation geometry and excitation condition admit distinct yet coupled representations. Guided by this structure, we propose the Geometry–Condition Factorized Neural Operator (GCFNO). A shared volumetric encoder extracts the latent electromagnetic state, a geometry factor models receiver interactions through coordinate-conditioned message passing, while a condition factor performs frequency-guided aggregation of depth-resolved features. A nonlinear readout then couples the two factors to predict sensor responses. On 3-D magnetotelluric modeling, GCFNO achieves the lowest three-seed mean total RMSE, while substantially improving held-out-frequency and receiver-geometry generalization. On head EIT, GCFNO achieves the lowest three-seed mean patient-macro error among evaluated surrogates. More broadly, these results suggest a path toward problem-specific neural operators whose architectures are derived from the governing physics, enabling more accurate and generalizable surrogates for complex physical systems. The methods and related analysis code of this manuscript are available at https://anonymous.4open.science/r/iclr2027-code-artifact-788E.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-guided neural operators, Electromagnetic forward modeling, Geometry–condition factorization, Magnetotellurics, Electrical impedance tomography

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_efAdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose the GCFNO for electromagnetic modeling. Motivated by local sensitivity analysis, it factorizes receiver–horizontal geometry from frequency–depth condition and couples them through a nonlinear readout. The authors provide local approximation and estimation analyses and evaluate GCFNO on synthetic 3D magnetotellurics (MT) and head EIT benchmarks.

Critical strengths
  1. The geometry–condition factorization is physically motivated and clearly connected to electromagnetic sensitivity.
  2. The paper provides useful theoretical support, including a spectral-tail local approximation bound and reference-model separation/estimation analysis.
  3. The experimental results were remarkable, and there were also abundant visual verifications.
Critical weaknesses
  1. I'm not very familiar with this field. Are the two benchmark datasets sufficient to validate the proposed method?
  2. The theory relies on local, small-perturbation, smoothness, and translation-invariance assumptions, so its connection to the trained global nonlinear operator is indirect.
  3. Ablations are insufficient for the main components, such as geometry message passing, depth aggregation, FiLM/readout, and residual design; evaluation is mostly synthetic with limited real-data/noise validation.
Questions for authors

Please see Weaknesses!

2 comments
AuthorsOpenReview ↗

Response to Reviewer efAd

We thank the reviewer for the constructive feedback and respond to each point below.

Weaknesses R1: MT and head EIT provide two complementary electromagnetic benchmarks with substantially different physical and computational characteristics.: MT evaluates 900 3-D volumes (630/135/135 split) on a structured 50×5050\times 50 surface grid across 15 frequencies, whereas head EIT evaluates 193 subject-specific anatomies (135/29/29 split, ~76k finite-element cells) across 17 frequencies, 31 current-injection patterns, and 33 scalp electrodes. They span open-domain geophysical modeling with skin-depth attenuation and bounded, anatomically irregular biomedical modeling with boundary-electrode measurements. Together, these evaluations provide complementary evidence for the effectiveness of sensitivity-guided geometry--condition factorization across representative electromagnetic forward-modeling scenarios.

Weaknesses R2: Thanks for sharing the same concern with Reviewer 1. Proposition 1 and Proposition 4 analyze the local Fr'echet derivative under reference symmetry; they do not assume that the full nonlinear forward operator is globally linear or translation invariant. GCFNO uses this structure as an architectural inductive bias: in Eq. (13), latent feature access satisfies ∂di,f,k∂Dd,k(η)=wi(η)ad(f)\frac{\partial d_{i,f,k}}{\partial D_{d,k}(\eta)} = w_i(\eta)a_d(f), separating receiver--horizontal geometry (i,η)(i,\eta) from frequency--depth dependence (f,d)(f,d), while the shared encoder EθE_\theta and nonlinear readout RθR_\theta learn the nonlinear material dependence end-to-end from solver responses. Appendix Fig. 10 shows that learned EIT depth weights shift from interior regions at low frequencies to boundary-adjacent regions at high frequencies, consistent with this organization. Appendix C.4 (Eqs. 47--51) further links local factor estimation to nonlinear prediction error by accounting for rank-truncation bias, estimation variance, and the second-order Taylor remainder.

Weaknesses R3: We added matched component ablations on 3-D MT using identical splits and training protocols, along with the relative degradation compared with GCFNO full:

VariantGeometryDepthResidualRMSE ↓\downarrowReduction vs. GNO ↑\uparrowDegradation vs. GCFNO Full
GNO matchedGNNNoneNo0.018577—+7.21%
+ Residual decoderGNNNoneYes0.0176445.02%+1.82%
+ Depth attention, w/o FiLMGNNAttentionYes0.0173716.49%+0.25%
GCFNO fullGNNAttentionYes0.0173286.72%0.00%
w/o geometry message passingNode-wiseAttentionYes0.0182341.85%+5.23%
w/o depth attentionGNNMeanYes0.0173726.49%+0.25%

On the fixed MT grid, the residual decoder gives the largest sequential gain (5.02% over matched GNO), and geometry message passing has the largest leave-one-out impact (+5.23% degradation when removed). Depth attention and FiLM each contribute ~0.25% on fixed-grid RMSE, while structured depth conditioning becomes more important under held-out-frequency completion, lowering macro-averaged RMSE from 1.264 (GNO freq-only) to 0.546. On input degradation (Appendix G.2, Table 8), GCFNO maintains lower error than GNO across all log-resistivity noise levels (0.010.01--0.100.10, e.g., 0.017775 vs. 0.019430 at 0.100.10), all coarse-grid factors (2, 4, 5), and boundary smoothing at level 0.25 (0.033771 vs. 0.034397).

The author has addressed my concerns, and I no longer have any further comments. Furthermore, I continue to give positive ratings.

Reviewer_DLk1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GCFNO predicts electromagnetic responses by separating geometry processing from frequency and excitation conditioning while preserving depth information. Experiments on 900 synthetic magnetotelluric volumes and 193 head anatomies with simulated EIT voltages report improved prediction accuracy and MT generalization.

Critical strengths

The method addresses costly repeated electromagnetic solvers in the inverse problem. Its geometry–condition design is physically motivated by local sensitivity analysis, supporting its potential usefulness in inversion. MT mean RMSE improves by 7.6% over GNO across three seeds. The frequency-input control and evaluation on unseen EIT subjects strengthen the empirical evidence.

Critical weaknesses

1.This paper has a plausible motivation. EIT Calibration setup is not introduced, matching of readout components, model sizes, and training budgets is unclear. 2. The evaluation relies on simulated electromagnetic responses, without validation using experimental measurements. The reported gains therefore do not yet establish practical usefulness under measurement noise, calibration errors, and model mismatch. 3.A substantial practical advantage remains unproven. EIT improves mean RMSE by about 3.8% on simulated data, with appreciable seed variation. Calibration robustness is untested, and the illustrative reconstructions lack systematic comparisons of imaging quality and total computational cost against existing methods.

Questions for authors

1.EIT Calibration robustness is untested. How are electrode contact impedances and instrument calibration handled? Is validation on experimental EIT measurements available to demonstrate that the reported gains translate to practical imaging? 2. What evidence supports total computational savings and robustness to EIT calibration or acquisition changes? 3. What physical evidence supports frequency-dependent boundary-distance aggregation in EIT beyond the learned weights? 4. Given the extensive work on electromagnetic imaging, including EIT and ECT, the authors should clarify which limitations of existing physics-based and learned methods it addresses. Comparisons of reconstruction quality and total computational cost against appropriate established baselines are needed to demonstrate a meaningful practical advantage.

1 comment
AuthorsOpenReview ↗

Response to Reviewer DLk1

We thank the reviewer for recognizing GCFNO's physical motivation and potential. We clarify the EIT protocol and calibration assumptions and add systematic inversion comparisons, including measured clinical EIT data.

w1 The head model uses 33 electrodes (32 plus COM), 31 fixed directed injection patterns, and 17 frequencies (5–2000 Hz). Device currents range from 45 to 280 μA; FEM generates current-normalized complex potentials. Each electrode comprises 32 surface triangles; its area-weighted centroid on the 75k mesh is the neural geometry input. FEM responses use an approximately one-million-tetrahedron mesh. Contact impedance is modeled by the Complete Electrode Model (CEM), with boundary condition Uℓ−u=zℓjnU_\ell-u=z_\ell j_n and nominal zcontact=0.001z_{\mathrm{contact}}=0.001, shared across electrodes, frequencies, and subjects and fixed during inversion.

All eight learned surrogates use the same training data, split, and maximum epoch budget, with backbone-specific readouts. The table reports parameter storage, best-validation epoch, and reconstruction performance.

ModelParameters (MB)RMSE ↓Macro Dice ↑Forward relative error ↓Reconstruction time (s) ↓
GCFNO0.5260.2542491.03%0.30%1.23
Voxel-FNO7.4930.2550290.98%0.46%1.55
Perceiver IO1.3560.4494284.89%11.48%5.37
DeepSets0.4500.4628285.61%23.83%3.57
3D U-Net1.4230.5075279.13%16.43%3.88
DeepONet0.4781.165838.28%5.24%1.75
EFKAN10.0561.191868.39%47.97%3.84
GNO0.5040.2557490.90%1.03%1.19
FEM–GN—0.4075530.94%2.66%2262.176

GCFNO achieves the best RMSE, Macro Dice, and forward error, with parameter storage close to GNO and 14.2× smaller than Voxel-FNO. It reduces online inversion time relative to FEM–GN by approximately 1839×. Reconstruction visualizations are included in the revised manuscript.

w2 We appreciate the reviewer’s concern regarding the practical relevance of simulated EIT. Our simulations use anatomically realistic head models reconstructed from real-subject MRI scans and follow the target device's acquisition protocol (33 electrodes, 31 injection patterns, and 17 frequencies from 5–2000 Hz). The FEM solver incorporates electrode–tissue contact effects through the Complete Electrode Model (CEM), with a nominal contact impedance of 0.001. This provides a physically grounded evaluation under device-aligned acquisition conditions. Furthermore, we evaluate GCFNO on experimentally measured voltages from the UCLH Stroke EIT Dataset [1]. Compared with FEM–GN, GCFNO reduces centroid localization error from 66.42 mm to 17.21 mm. These results provide initial evidence of GCFNO's potential for practical stroke imaging. We will present the visualization results of imaging from real data in the revised appendix.

Q1 Please see our responses to W1 and W2 for the EIT acquisition protocol, CEM-based contact impedance modeling, and validation using experimentally measured UCLH data.

Q2 GCFNO reduces online reconstruction time from 2262.176 s for FEM–GN to 1.23 s, achieving an approximately 1839× speedup under our controlled inversion benchmark. Its compact parameter storage (0.526 MB) further supports efficient deployment. The additional UCLH experiment demonstrates applicability to independently acquired clinical measurements. These results support computational efficiency and initial cross-setting applicability.

Q3 Appendix Fig. 10 shows learned profiles shifting from interior features at lower frequencies toward boundary-adjacent features at higher frequencies. This empirical diagnostic is consistent with frequency-dependent spatial sensitivity in multifrequency EIT. We present it as learned aggregation behavior.

Q4 FEM–GN provides physics-based inversion but requires repeated forward and sensitivity solves; learned surrogates reduce this cost but depend on how geometry and excitation are represented. GCFNO uses sensitivity-guided geometry–condition factorization for differentiable forward modeling and iterative inversion. Our comparisons include FEM–GN, seven learned surrogates, and measured UCLH data. We present the computation time of different surrogate models and that of FEM-GN in the table within our response to w1. Importantly, on measured UCLH data, GCFNO reduces lesion localization error from 66.42 mm to 17.21 mm (74.1%). These results demonstrate its potential for efficient and accurate EIT inversion.

[1] Goren, N., Avery, J., Dowrick, T., et al. (2018). Multi-frequency electrical impedance tomography and neuroimaging data in stroke patients. Scientific Data, 5, 180112. https://doi.org/10.1038/sdata.2018.112

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

The paper aims to predict electromagnetic measurements from physical inputs using a neural network. To motivate its design, the authors study how small changes to those inputs affect the measurements. They observe that rearranging a sensitivity matrix to group spatial relationships separately from frequency/depth relationships gives an approximately low-rank representation in their tests. They use this observation to motivate the network's structure. The reported results show improved prediction accuracy on their chosen applications.

I understand the main contribution to be better prediction accuracy and generalization. The broader motivation is to reduce the cost of repeated physical simulations. From my applied perspective, I would like to understand how the mathematical motivation relates to what the network learns and what tradeoff it offers when used in place of the original solver.

Critical strengths
  • I appreciate the attempt to connect the network design to physics, rather than treating its structure only as an empirical choice. Studying low-rank structure in the sensitivity matrix offers a principled starting point for explaining what information the network might reuse.
  • The authors explore the same organizing idea in more than one application. I find the attempt to provide a common physical and mathematical explanation valuable, even though my familiarity with these particular applications is limited.
Critical weaknesses

The following concerns come from my experience with inverse problems in general, rather than expertise in these particular applications. They reflect what would help me understand and assess the method's practical value.

  • I would like a clearer connection between the local analysis and what the network learns. The mathematical approximation describes small changes around a particular input, while the network is trained to predict full measurements for many different inputs. That is not inherently inconsistent: a nonlinear network can learn different local behavior in different places. However, the paper does not show whether the trained network follows the physical sensitivity structure used to motivate it, or how large a change to the input the local approximation can handle accurately. Good response predictions alone do not answer those questions.

  • The practical cost and the effect on reconstruction remain unclear. The paper focuses on prediction quality, but does not report the time needed relative to the physical solver or other networks. The head-EIT results in Table 1 appear close to FNO, and I do not know whether this difference is meaningful for the intended application. Even a modest quality gain could be useful; knowing its computational cost would help put it in context. Figures 11--12 show reconstructed interiors, but without a corresponding reconstruction using the physical solver, I cannot tell which differences from ground truth come from the inverse problem and which are added by the network approximation. For example, a smooth reconstruction could result from limited measurements or regularization even with the original solver. If GCFNO produces a similarly useful answer much faster, that would be a meaningful result. If it loses additional detail, that tradeoff should be measured. The inversion algorithm, starting point, prior information, and stopping criteria are also not clear from the manuscript. If these details are already in the code, a short description in the paper or supplement would help readers assess the examples.

Questions for authors
  • How far does the local explanation hold in practice? For a few held-out inputs, could you vary the size of input changes and compare the physical response change, the local factorized approximation, and the network's prediction? This would help show where the approximation is accurate and whether the network learns the behavior motivating its design.

  • How were Figures 11--12 reconstructed? Were the network weights frozen and gradients taken with respect to the input, or was a different procedure used? Please specify the objective, regularization, algorithm, starting point, and stopping criteria, as well as which anatomical or geometric information was already known. Since the result can depend on initialization, it would help to know how the starting point was chosen and whether other starts gave similar results. Were any checks made that the network remained accurate on the inputs visited during inversion?

  • What is the speed--quality tradeoff? Please report forward-evaluation times for the physical solver, GCFNO, and relevant neural baselines under a clearly specified workload and hardware. For the head-EIT results in Table 1, could you explain whether the difference from FNO is meaningful in practice and whether it comes with a difference in computational cost? For an example in Figures 11--12, could you compare two reconstructions that solve the same inverse problem: one using GCFNO for the forward calculations and one using the original physical solver? Please use the same input parameterization, initial guess, prior information, and comparable regularization and stopping criteria, and explain any changes required to the algorithm. I am asking for a reference reconstruction with the original solver, rather than a new inverse method. Reporting time, reconstruction quality, and agreement with the observed measurements as evaluated by the physical solver would help distinguish existing reconstruction limitations from errors introduced by the approximation.

1 comment
AuthorsOpenReview ↗

Response to Reviewer wmgr

We thank the reviewer for the constructive questions. We clarify the sensitivity–network connection and provide additional inversion and perturbation analyses.

W1 The proposed grouping is (p,ξ)∣(f,z)(p,\xi)\mid(f,z); Group A/B are (p,f)∣(ξ,z)(p,f)\mid(\xi,z) and (p,z)∣(ξ,f)(p,z)\mid(\xi,f), we have added experiments under 9×99\times9 and 16×1616\times16 grids as below:

PartitionMatrixE1E_1E4E_4E8E_8R95R_{95}A E4E_4B E4E_4
4×4256×2080.07%98.44%99.77%382.25%81.30%
9×9576×4576.06%96.05%99.24%452.78%53.62%
16×161024×8075.21%94.94%98.41%545.99%47.61%

The proposed grouping retains high rank-4 energy as resolution increases, while its advantage over both alternatives grows. GCFNO implements this separation in Eq. (13) through the latent access kernel wi(η)ad(f)w_i(\eta)a_d(f).

We also compute GCFNO input Jacobians at the same 12 held-out MT backgrounds, using identical variable unfoldings:

GroupingPhysical E4E_4GCFNO E4E_4Physical R95R_{95}GCFNO R95R_{95}
Proposed98.44±0.26%98.44\pm0.26\%90.46±1.62%90.46\pm1.62\%2.83±0.382.83\pm0.387.28±0.867.28\pm0.86
A82.55±1.07%82.55\pm1.07\%53.05±5.15%53.05\pm5.15\%6.83±0.296.83\pm0.2922.70±1.1022.70\pm1.10
B81.30±1.95%81.30\pm1.95\%51.65±6.12%51.65\pm6.12\%7.00±0.257.00\pm0.2525.52±2.1125.52\pm2.11

The learned Jacobians preserve the physical sensitivity's preferred grouping across backgrounds. Appendix Fig. 10 further shows frequency-dependent depth aggregation in EIT: lower frequencies emphasize interior features, while higher frequencies emphasize boundary-adjacent features.

W2. Head-EIT inversion comparison. We compare eight learned surrogates with FEM–GN:

ModelParameters (MB)RMSE ↓Macro Dice ↑Forward error ↓Time (s) ↓
GCFNO0.5260.2542491.03%0.30%1.23
FNO7.4930.2550290.98%0.46%1.55
Perceiver IO1.3560.4494284.89%11.48%5.37
DeepSets0.4500.4628285.61%23.83%3.57
3D U-Net1.4230.5075279.13%16.43%3.88
DeepONet0.4781.165838.28%5.24%1.75
EFKAN10.0561.191868.39%47.97%3.84
GNO0.5040.2557490.90%1.03%1.19
FEM–GN—0.4075530.94%2.66%2262.176

All learned models use the same data, split, and epoch budget, with backbone-specific readouts. GCFNO achieves the best RMSE, Macro Dice, and forward error, with 14.2× smaller parameter storage than FNO. Its 1.23-s reconstruction time is comparable to GNO (1.19 s) and much shorter than FEM–GN (Finite Element Method, Gauss-Newton). GCFNO weights remain frozen; Autograd computes gradients with respect to input admittivity from the voltage-misfit objective. Starting from the same homogeneous model, Adam runs 160 steps. FEM–GN allows up to eight outer iterations per frequency, stopping on line-search failure or relative objective decrease below 2×10−42\times10^{-4}. More details are provided in the revised appendix.

In addition, we also compare the inversion results of GCFNO and traditional inversion methods on the measured head EIT dataset UCLH Stroke EIT Dataset; see the revised appendix. Experimental results show that Compared with FEM–GN, GCFNO reduces lesion localization error from 66.42 mm to 17.21 mm. These results provide evidence of GCFNO's potential for practical stroke imaging

*Q1. Proposition 1 bounds the local approximation error for ΔX=αV\Delta X=\alpha V by αϵR∥V∥F+α2L2∥V∥F2\alpha\epsilon_R\|V\|_F+\frac{\alpha^2L}{2}\|V\|_F^2. We perturb held-out MT models from a reference X0X_0 along a fixed spatial pattern VV, and compare the Maxwell response, the rank-4 local prediction using the Jacobian at X0X_0, and GCFNO's full prediction:

α\alphaPhysical response changeRank-4 local errorGCFNO full-response error
00.00%—1.78%
0.100.30%9.18%1.78%
0.250.75%9.83%1.79%
0.501.50%11.40%1.84%
0.752.27%13.36%1.93%
1.003.04%15.50%2.08%

As α\alpha increases, the rank-4 error rises from 9.18% to 15.50%, consistent with the nonlinear remainder, while GCFNO's full-response error remains 1.78%–2.08%. Together with the Jacobian results, this shows that sensitivity guides GCFNO's input-dependent structure, while nonlinear learning captures responses beyond the fixed rank-4 local approximation.

Q2. Please refer to the response to Weekness 2.

Q3. As we mentioned in our response to your Weekness 1, the learning-based agent method represented by GCFNO can achieve an average speedup of nearly 800 times compared with traditional inversion methods. GCFNO's structural inductive bias provides more stable gradients for inversion, yielding the lowest reconstruction RMSE. Although its forward accuracy gain over FNO is modest, small prediction errors can propagate through iterative optimization and affect final reconstruction quality.

Reviewer_AN4ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript Geometry-Condition Factorized Neural Operators for Electromagnetic Modeling describes GCFNO, a learning framework designed for Electromagnetic forward modeling. Motivated by electromagnetic sensitivity analysis, GCFNO split information into geometry and frequency. The network is designed to handle different geometries and a residual correction is used to improve prediction accuracy. The framework is tested on multiple problems and shown an advantage over other several learning-based baselines.

Critical strengths
  1. The manuscript shows direct applications of the proposed method, and the method is supported by reasonable physical motivation.

  2. The authors provide some of the source code for the paper, and have enough experiments to support the main claims of the paper.

Critical weaknesses
  1. The presentation of the paper could be improved… Figures have inconsistent font sizes and color schemes/colormaps. Some text is too small, and the overall alignment could be better. The texts/markers could also be better designed. For example the “50x50” lable on figures 3 seems strange. The source code also contains emojis and Chinese characters.

  2. The low-rank analysis is not on the final nonlinear model and thus there seems to be a gap between the theory and the final prediction error. Also, the squared energy can be a little bit misleading and exaggerate the low-rank property of the matrices. And there are no general guarantee of the low-rank property.

Questions for authors
  1. It would be good if the authors could provide some guidance on how to design the network and choose parameters. The model is very complicated and requires a lot of design choices and training parameters. For different applications how should one choose an encoder and decoder? Also since the network is a little bit complicated, for a fixed budget (memory and computational), how should one assign the parameters to different components?

  2. Also more details on experiments would be great. Are the baselines selected to have similar numbers of parameters? Is the training cost roughly the same? Any LR sweep? Training time? etc.

2 comments
AuthorsOpenReview ↗

Response to Reviewer AN4Z

We thank the reviewer for the constructive feedback and provide point-by-point responses below.

Weaknesses R.1: Thank you for these concrete suggestions. We will standardize figure typography, panel alignment, line widths, and colormaps; enlarge labels and legends; and revise the receiver-grid annotation in Fig. 3. We will also clean the public code artifact by removing nonessential emojis and non-English comments/output strings, while preserving technical documentation.

Weaknesses R.2: We appreciate these insightful theoretical questions and address each of the three points below:

  1. Why we use squared energy: We use normalized squared singular values er=sr2/∑jsj2e_r=s_r^2/\sum_j s_j^2 in Fig. 2 because, by the Eckart–Young theorem, the uncaptured energy 1−ER1-E_R directly corresponds to the relative squared Frobenius approximation error used in Proposition 1 (Eq. 8) and Proposition 7 (Eqs. 40–41). The spectral concentration remains evident without squaring: under the proposed (p,ξ)∣(f,z)(p,\xi)\mid(f,z) grouping, s4/s1≈0.095s_4/s_1\approx0.095, with the top four singular values capturing approximately 86.4% of the nuclear norm, compared with 61.2% and 59.8% for the alternative groupings. We have added the unsquared singular-value spectrum in Appendix B.1 for clarity.

  2. Theoretical guarantees of the low-rank property: Appendix B.2 establishes sufficient conditions for the low-rank structure. Under horizontal translation invariance and transverse Sobolev regularity s>0s>0, Lemma 3 and Proposition 4 guarantee a rank-N(Q)=O((1+Q)2)N(Q)=\mathcal{O}((1+Q)^2) approximation with error bounded by Ms(1+Q2)−s/2M_s(1+Q^2)^{-s/2}. Corollary 5 extends this result to heterogeneous sensitivity kernels within an L2L^2 distance η\eta of the reference kernel, adding η\eta to the error bound. Because diffusive electromagnetic fields act as spatial low-pass filters (skin-depth smoothing), localized 3-D conductivity anomalies induce smooth, bounded kernel perturbations η\eta, explaining why high spectral concentration persists across all 12 heterogeneous 3-D models in Fig.~2.

  3. Bridging local theory to the final nonlinear model: We establish the connection between local sensitivity analysis and nonlinear operator learning from both architectural and statistical perspectives. Architecturally, GCFNO incorporates the geometry–condition factorization into its latent feature-access mechanism, where the kernel wi(η)ad(f)w_i(\eta)a_d(f) explicitly separates receiver–horizontal geometry from frequency–depth dependence. Statistically, Appendix C.4 extends the local estimation analysis to nonlinear prediction by accounting for rank-truncation bias, estimation variance, and the second-order Taylor remainder. Under the sufficient condition in Eq. 51, the structured estimator achieves strictly lower local prediction risk than the unrestricted estimator, benefiting from the reduced effective degrees of freedom R(a+b−R)R(a+b-R) compared with abab. Together, these results provide a theoretical motivation for sensitivity-guided architectural design and offer an explanation for the improved predictive accuracy and data efficiency observed in our experiments.

Questions R3: GCFNO separates spatial encoding, geometry interaction, and frequency–depth conditioning into distinct components, but this modular design does not lead to an unusually large model. A practical design rule is to match the encoder to the input discretization and the decoder to the observation geometry. Regular volumetric inputs are efficiently handled by the convolutional encoder used here, whereas irregular meshes can use a graph- or point-based encoder. A fixed-grid decoder is sufficient when receiver locations and frequencies remain unchanged, while a coordinate- and frequency-conditioned decoder is preferable when these queries vary across applications.

Questions R4: We thank the reviewer for highlighting the importance of reproducibility. The trainable parameter counts of the implementations used in our 3-D MT benchmark are summarized below.

ModelParameters (M) ↓Rel. to GCFNOMT Total RMSE ↓Head EIT RMSE (mV) ↓
3D U-Net1.3342.65×0.047420.3899
FNO14.16928.11×0.031780.3882
DeepONet0.0930.18×0.068600.6601
Perceiver IO0.3420.68×0.060430.4798
DeepSets0.1340.27×0.057320.4071
EFKAN15.62331.00×0.048099.3627
GNO0.5471.09×0.019070.3922
GCFNO (Ours)0.5041.00×0.017620.3735

GCFNO achieves a favorable accuracy–parameter trade-off: compared with GNO, it uses 7.9% fewer parameters and reduces MT RMSE by 7.6%, while outperforming both smaller and substantially larger baselines. We will further report training epochs, learning-rate selection, hardware configurations, and wall-clock training times in the revised Appendix to clarify computational fairness.

Thank you for the clarifications and for partially addressing my comments. However, the comparison against the baselines does not seem to be apples-to-apples, as the settings vary a lot across different methods. It is somewhat hard to tell how fair the comparison against the baselines is.

Nonetheless, the proposed revisions would improve the paper. I therefore maintain my score of 3.

Reviewer_j8z8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Geometry–Condition Factorized Neural Operator (GCFNO) for forward models in EM problems. The architecture is motivated by a low-rank approximation of the local sensitivity of the Maxwell's equations (forward problem). The neural architecture combines several existing tricks and modules, with numerical experiments on magnetotellurics (MT) and head electrical impedance tomography (EIT).

I think the paper would be considerably stronger if geometry–condition factorization were formulated as a meta-method that can be incorporated into existing neural operators such as FNO, GNO, or DeepONet. The authors could then evaluate whether the same factorization mechanism consistently improves different backbones on both MT and EIT. As of now, this paper fits more for an engineering journal than ICLR.

Critical strengths
  1. The authors assemble two electromagnetic forward modeling benchmarks, MT and EIT, involving different physical settings and spatial discretizations. This provides a useful benchmark for neural operators in this field.
  2. The proposed neural operators work on unstructured meshes.
  3. The paper makes an effort to motivate the architecture mathematically through the properties of the forward operator.
Critical weaknesses
  1. The theoretical analysis reads rather like a placeholder. The low-rank factorization in (7) is not explicitly implemented or enforced in the actual neural operator. (13) is only for latent feature aggregation, how it matches (6)?

  2. Proposition 4 reads like another placeholder, which is essentially a Fourier approximation result for translation invariant kernel, and there is no particular property of Maxwell's equations (as forward problem) used, nor does it motivate the neural architecture.

  3. In Figure 2, based on the Appendix B, the sensitivity tensor is estimated by partitioning the field variable into 4x4 blocks. How the authors are sure that the decaying spectrum (rank-4) is not majorly caused by this coarsening? At least test on finer grids.

  4. As for the neural operator itself, GCFNO essentially is a mixed and match with existing techniques (graph message passing, attention-like depth aggregation, FiLM, etc).

  5. The experiment improvements are too marginal (esp for EIT), and lacks a systematic ablations. For example, for all those combined modules, which one attributes to the improved RMSE most? In the meantime, the claimed geometry generalization is also rather limited, since the "unseen" receiver layouts are evaluated on the same underlying MT solver grid, and the improvements over GNO are generally small.

  6. EIT is an inverse problem, thus I find posing GCFNO as a forward problem surrogate in EIT weird. The reason is that the paper did not demonstrate whether replacing the forward solver (like finite element) with GCFNO improves computational efficiency or preserves reconstruction accuracy in EIT.

Questions for authors
  • Page 4, line 190: what is ρ\rho? it is not defined before.
  • What discretization is used to generate MT reference data? H(curl)-conforming FEM?
3 comments
AuthorsOpenReview ↗

Response to Reviewer j8z8

We thank the reviewer for the constructive assessment. Our findings extend beyond a specific engineering application by showing how physical structure can guide neural operator design. Similar structures may exist in heat transfer, fluid mechanics, and aerodynamics, potentially inspiring physics-informed architectures. We respond below; detailed experiments and analyses are provided in the revised appendix.

Weaknesses R1: Eq. (7) describes the local Jacobian, while GCFNO predicts the full nonlinear forward map. GCFNO uses sensitivity-derived variable grouping as an architectural inductive bias, avoiding direct prediction of nonunique, high-dimensional SVD factors. In Eq. (13), receiver sampling provides the geometry pathway and frequency-dependent depth aggregation the condition pathway, reflecting Eq. (6). Appendix Fig. 10 shows learned depth weights shifting from interior to boundary-adjacent regions as frequency increases, supporting this physical interpretation.

Weaknesses R2: Appendix A.1 derives Maxwell sensitivity from forward/adjoint fields (Lemma 2, Eq. (17)) and the frequency–skin-depth relation (Eq. (21)). Under a horizontally invariant reference medium and excitation, Lemma 3 establishes translation symmetry in receiver and perturbation coordinates. Proposition 4 then establishes a finite Fourier separation into geometry modes and frequency–depth coefficients. Appendix B.2 specifies the sufficient assumptions and reference-setting scope.

Weaknesses R3: We refined the centered finite-difference sensitivity diagnostic (h=0.03h=0.03) from 4×4 to 9×9 and 16×16 partitions. The table shows captured squared energy. The proposed grouping is (p,ξ)∣(f,z)(p,\xi)\mid(f,z); Group A is (p,f)∣(ξ,z)(p,f)\mid(\xi,z) and Group B is (p,z)∣(ξ,f)(p,z)\mid(\xi,f). The last two columns report alternative rank-4 energies.

PartitionMatrixE1E_1E4E_4E8E_8R95R_{95}Group A E4E_4Group B E4E_4
4×4256×2080.07%98.44%99.77%382.25%81.30%
9×9576×4576.06%96.05%99.24%452.78%53.62%
16×161024×8075.21%94.94%98.41%545.99%47.61%

The proposed grouping remains spectrally concentrated at finer resolutions (Appendix B.1); at 16×16, its rank-4 energy exceeds both alternatives by over 47 percentage points.

Weaknesses R4: Our contribution is sensitivity-guided operator-level factorization; message passing, attention, and FiLM implement geometry, condition, and fusion pathways. Incorporating the geometry–condition structure into other backbones also improves MT prediction:

BackboneBaseline RMSE+ Geometry–Condition RMSE
FNO0.031780.02180
DeepONet0.068600.019091

Full log-resistivity and phase metrics are in the revised appendix. These results support transfer of the grouping principle across architectures.

Weaknesses R5: We added matched 3-D MT component ablations using identical splits and training protocols:

VariantGeometryDepthResidualRMSEReduction vs. GNO
GNO matchedGNNNoneNo0.018577—
+ Residual decoderGNNNoneYes0.0176445.02%
+ Depth attention, w/o FiLMGNNAttentionYes0.0173716.49%
GCFNO fullGNNAttentionYes0.0173286.72%
w/o geometry message passingNode-wiseAttentionYes0.0182341.85%
w/o depth attentionGNNMeanYes0.0173726.49%

The residual decoder yields the largest sequential gain (5.02%); removing geometry message passing raises full-model RMSE by 5.23%. The isolated fixed-grid effects of depth attention and FiLM are ~0.25% each. The EIT improvement over GNO is modest (0.3922 → 0.3735 mV; 4.75%).

Weaknesses R6: Appendix I presents new Autograd-based Head-EIT inversion experiments. Across 29 held-out subjects, GCFNO achieves the best mean values on all six reconstruction metrics (RMSE 0.30503, Macro Dice 75.944%), with modest gains over seven surrogates. On one matched held-out subject, GCFNO reduces measured reconstruction time from 2262.176 s (FEM–GN) to 1.705 s (1326.8× speedup), while improving RMSE (0.40755 → 0.28583) and Macro Dice (30.938% → 80.257%). FEM–GN attains a lower independent-FEM voltage error (0.951% vs. 11.874%); Appendix I discusses this measurement-fit/reconstruction-fidelity trade-off.

Questions For Authors R7: The input is log⁡10ρ\log_{10}\rho, where ρ\rho is isotropic resistivity (Ω·m) and solver conductivity is σ=1/ρ\sigma=1/\rho. The frequency-domain 3-D MT solver uses lowest-order first-family Nédélec H(curl)H(\mathrm{curl}) elements on structured hexahedra (one degree of freedom per oriented edge). Complex curl–curl systems for two polarizations are solved with GPU-accelerated BiCGStab. Stretched padding and background continuation reduce boundary truncation; surface fields yield apparent resistivity and phase.

Thank you for the clarifications and additional experiments. The finer sensitivity partitions, cross-backbone comparisons and component ablations address several concerns. However, the connection between the Jacobian factorization and the implemented architecture remains qualitative, and the ablations attribute most of the sequential improvement to the residual decoder. Generalization to genuinely new geometries remains unestablished. The EIT inversion results are encouraging, but the single-subject FEM comparison and substantially larger independent-FEM voltage error limit the conclusions about practical benefit.

These additions strengthen the empirical evidence but do not sufficiently resolve my concerns about methodological novelty and generality. I therefore retain my score of 2.

Thank you for your thoughtful feedback. We fully agree that both MT and EIT are fundamentally inverse problems, and the practical value of a forward surrogate should ultimately be evaluated through inversion accuracy and efficiency. We provide additional inversion experiments below to address this concern for both you and future readers. In the revised manuscript and appendix, we have added inversion visualizations and quantitative comparisons across 3D MT, synthetic EIT, and measured EIT (see the newly added figures and tables). Overall, GCFNO achieves strong reconstruction accuracy and computational efficiency across these inversion tasks, and accurately recovers and localizes CT-consistent low-impedance lesions in measured clinical cases.

Specifically, our new measured-data inversion experiment on the public UCLH multi-frequency head EIT dataset covers 34 recordings from 28 subjects (18 patients and 10 healthy volunteers). Using a 24/7 excitation split for inversion and held-out validation, we apply the official 280 μA280\,\mu\text{A} calibration, bad-channel masking, fitting-set gain, and fixed contact impedance, with no CT/MRI lesion labels used during inversion:

Measured ExperimentMethodHeld-out Rel. L2L_2: Initial →\rightarrow Post-InversionImproved RecordingsMedian Time
17-freq. absolute voltageGCFNO72.07% →\rightarrow 62.13%34/342.88 s
17-freq. absolute voltageFEM–GN69.76% →\rightarrow 58.75%34/3427.95 s
Rel. 5 Hz freq.-differenceGCFNO99.24% →\rightarrow 81.23%34/342.90 s
Rel. 5 Hz freq.-differenceConventional FEM100.00% →\rightarrow 79.12%34/3416.02 s

GCFNO reduces the held-out measurement error on all 34 recordings, while achieving a 9.74×9.74\times speedup over FEM–GN for absolute-voltage inversion and a 5.55×5.55\times speedup over conventional FEM for frequency-difference inversion.

We further evaluate spatial localization using CT/MRI lesion annotations introduced only after inversion. On Patient 17 (100–5 Hz frequency-difference), the results are:

MethodLesion AUC ↑\uparrowHotspot Dice ↑\uparrowCentroid Error ↓\downarrow
GCFNO0.7390.15917.2 mm
FEM0.3610.00066.4 mm

These results demonstrate that a frozen GCFNO forward surrogate can be directly used for iterative inversion of real multi-frequency EIT measurements, improving measurement consistency on held-out excitation patterns while substantially reducing online reconstruction time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.