acceptodds
Under review as a conference paper at ICLR 2027

EFSA: Physically Interpretable Facet Modeling for High-Frequency RCS Prediction

Abstract

Learning physically interpretable and computationally efficient representations for high-frequency radar cross section (RCS) prediction is a critical yet unresolved challenge. While mesh-based Graph Neural Networks (GNNs) provide an excellent paradigm for understanding geometric and electromagnetic interactions, they are inherently bottlenecked by rigid sub-wavelength constraints, where shrinking wavelengths lead to catastrophic graph size expansion. To fundamentally decouple neural representations from these strict physical size requirements, we introduce the Equivalent Facet Scattering Amplitude (EFSA), which models the RCS contribution of facets at arbitrary scales. EFSA combines a Physical Optics (PO) prior with a learnable complex correction factor. Specifically, the PO prior absorbs the orders-of-magnitude variations in RCS contributions across differently scaled facets within the same target, while the correction factor captures the intra-facet coherent interference of facets significantly larger than the incident wavelength. Experiments on real-world targets demonstrate that our framework enables physically interpretable RCS prediction in high-frequency regimes while outperforming existing methods in predictive accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Learning physics, Radar cross section

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

The paper introduces EFSA for radar cross-section prediction on coarse meshes. It combines a Physical Optics estimate with learned amplitude and phase corrections for each facet. Tests on basic shapes and ModelNet airplanes show lower average prediction errors than the selected baselines, with some loss of accuracy as the mesh is coarsened.

Critical strengths

The paper addresses a real computational limitation, using larger surface elements could reduce the cost of high-frequency scattering prediction.

a clear physical intuition: each facet contributes an amplitude and phase, and these contributions are combined coherently. This is a useful starting point for building a physically grounded model.

Critical weaknesses

whether learning provides a useful advantage over those methods? The central contribution is not yet convincing to me. The learned correction describes interference within a flat triangle, which is already expressed as a known integral in the paper. Established analytical methods can evaluate this integral. Comparing only with a facet-center approximation does not answer this question.

the benefits remain insufficiently demonstrated. All reference data come from a Physical Optics solver, so the results show agreement with that approximation rather than validation against the actual scattering problem. There is no full-wave or experimental comparison, and no runtime evidence that the network is preferable to accurate PO integration. The paper therefore establishes a promising surrogate, but not yet a clear advance in physical modeling or computational capability.

Questions for authors

Why predict the facet integral with a neural network rather than evaluate it analytically? What are the accuracy and runtime differences on the same meshes?

Is the network learning any physics beyond the PO approximation, or mainly reproducing its integration over each facet?

Can you validate the method against full-wave calculations or measurements on a small set of representative objects?

What is the measured speedup over an accurate PO implementation, including preprocessing?

1 comment
AuthorsOpenReview ↗

# Response to Reviewer PqxR

Response to Reviewer PqxR

Thank you for asking us to distinguish the role of learning from accurate PO integration and to clarify the scope of the numerical validation.

P01. Analytical integration and same-mesh accuracy and runtime

EFSA offers an end-to-end efficiency advantage over analytical PO on the same airplane OBJ meshes under the tested hardware configurations. Both approaches produce a complete monostatic RCS map, with preprocessing, postprocessing, and output writing included in the timing. See “Inference Efficiency” for the benchmark.

Analytical integration evaluates the phase integral exactly for a prescribed planar triangle and PO current. EFSA learns a complex correction to the facet-center estimate so that the supplied coarse mesh can be used to approximate the refined-mesh PO response. The refined-mesh PO solutions provide the supervision and evaluation reference for this surrogate task.1


Method3 GHz4.5 GHz6 GHz
Analytic PO30.5230.6329.49
EFSA10.7710.7010.52

EFSA was 2.80–2.86× as fast as analytical PO across the three tested frequencies, including preprocessing and output writing. The table reports mean seconds per target. See “Inference Efficiency” in the main text and “Additional Inference Timing Results” in the appendix.

P02. What the network learns from PO supervision

The learned amplitude and phase corrections reduce RCS error relative to PO-init on both datasets, as shown in “Ablation Study.” Both facet-level supervision and target RCS references come from PO. We clarify in “Dataset Preparation” that the reported accuracy measures how well EFSA approximates the refined-mesh PO response from coarse input meshes.


DatasetPO-init MAE (dB)EFSA MAE (dB)
Basic Geometry13.096.13
ModelNet-Airplane8.694.63

The learned corrections reduce MAE to roughly half that of PO-init on both datasets. See “Ablation Study” for the complete comparison and “Dataset Preparation” for the PO references.

P03. Full-wave or measured validation

The present study evaluates Basic Geometry and airplane geometries from ModelNet against refined-mesh PO references. We agree that comparing with an independent full-wave calculation or measurements would address a separate validation question about agreement with physical scattering. The manuscript's accuracy conclusions are limited to the PO reference used in the evaluation.

P04. Measured speedup including preprocessing

EFSA is faster end to end than analytical PO at all three frequencies and faster than multilevel PO at 4.5 and 6 GHz on the same airplane meshes under the tested CPU and GPU configurations. Timing includes OBJ preprocessing, RCS postprocessing and numerical output writing, from process launch to exit.12


Method3 GHz4.5 GHz6 GHz
Analytic PO30.5230.6329.49
Multilevel PO9.1612.4014.37
EFSA10.7710.7010.52

Speedup (baseline time / EFSA time)3 GHz4.5 GHz6 GHz
Analytic PO2.83×2.86×2.80×
Multilevel PO0.85×1.16×1.37×

With preprocessing included, EFSA is faster in five of the six PO baseline–frequency comparisons. Its speedup over analytical PO stays between 2.80× and 2.86×. The first table reports mean seconds per target, and the second gives the ratio of baseline time to EFSA time. See “Inference Efficiency” in the main text and “Additional Inference Timing Results” in the appendix.

Footnotes

  1. A. Altintas and A. Çelik. “Large Flat Plate Models in the Physical Optics Method for RCS Calculations.” 10th International Conference on Mathematical Methods in Electromagnetic Theory, 586–588, 2004. ↩ ↩2

  2. Z. Y. Xue, Y. M. Wu, W. C. Chew, Y. Q. Jin, and A. Boag. “The Multilevel Fast Physical Optics Method for Calculating High Frequency Scattered Fields.” Progress In Electromagnetics Research, 169, 1–15, 2020. ↩

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

The paper proposes EFSA, a framework for predicting radar cross sections from coarse triangular meshes. Each facet’s complex scattering contribution is represented as a physical-optics prior multiplied by learned amplitude and phase corrections. A graph neural network predicts these corrections using geometry, incident direction, wavelength, and facet electrical size. The method is evaluated on basic geometric shapes and ModelNet airplane models using PO-generated reference data, including tests with different mesh resolutions.

Critical strengths
  1. The paper addresses a relevant computational problem. Avoiding excessive mesh refinement at higher frequencies is a useful objective for learning-based scattering prediction.
  2. The representation has an explicit physical interpretation. Separating the baseline scattering response from amplitude and phase corrections makes the predicted intermediate quantities easier to understand.
Critical weaknesses
  1. The need for the learned correction is insufficiently justified. In Eq. (4), the correction factor is an integral determined by the known facet geometry, incident direction, and wavelength. Analytical evaluation of physical-optics integrals over electrically large planar triangles has already been studied. However, the paper compares against PO-init, which uses a facet-center approximation, without comparing against accurate analytical facet integration. The paper should discuss established analytical and accelerated PO methods, including Altintas and Celik (2004), and Boag (2004). These works address direct facet integration and efficient scattering evaluation across observation conditions. Without comparison against an appropriate analytical or accelerated PO baseline, the practical benefit of learning the correction factor remains unclear.
  2. Physical validation is limited to PO-generated data. All reference responses are generated by a PO solver. The experiments on “real-world targets” use object meshes rather than measured radar responses.
  3. The computational efficiency claims lack direct evidence. Table 1 reports memory consumption and prediction errors for different mesh resolutions within EFSA. However, the paper does not report inference-time comparisons against accurate PO integration or the competing methods.
  4. The ablation study is insufficient to establish the contributions of the proposed design. The comparison with PO-init demonstrates that adding a learned correction improves accuracy, and the mesh-resolution experiment provides useful evidence of robustness. However, the paper does not isolate the benefits of the physical factorization, amplitude and phase corrections, electrical-size encoding, or additional complex-valued supervision.

Altintas, A., & Celik, A. (2004, September). Large flat plate models in the physical optics method for RCS calculations. In 10th International Conference on Mathematical Methods in Electromagnetic Theory, 2004. (pp. 586-588). IEEE. Boag, A. (2004). A fast physical optics (FPO) algorithm for high frequency scattering. IEEE Transactions on Antennas and Propagation, 52(1), 197-204.

Questions for authors

When changing the mesh discretization, how much of the prediction difference comes from changes in the underlying geometry, and how much comes from sensitivity to the triangulation itself? Does EFSA produce consistent total scattering predictions under different triangulations of exactly the same surface?

4 comments
AuthorsOpenReview ↗

# Response to Reviewer QxJS part1

Thank you for pointing us to the analytical PO literature and for asking which parts of the model and evaluation support our conclusions.

Q01. Analytical integration and the purpose of the learned correction

We agree that Eq. (4) admits exact analytical phase integration for a prescribed planar triangle and PO current. EFSA takes the supplied coarse mesh and learns a complex correction to a facet-center estimate to approximate the refined-mesh PO response. The full model lowers RCS error relative to both direct prediction of complex facet contributions and PO-init. The same-OBJ benchmark also shows an end-to-end efficiency advantage over analytical PO under the tested hardware configurations. See “Ablation Study” and “Inference Efficiency” for the results.[^1]


Method3 GHz4.5 GHz6 GHz
Analytic PO30.5230.6329.49
EFSA10.7710.7010.52

EFSA was 2.80–2.86× as fast as analytical PO across the three tested frequencies, including preprocessing and output writing. The table reports mean seconds per target. See “Inference Efficiency” in the main text and “Additional Inference Timing Results” in the appendix.

Q02. Analytical and accelerated PO literature

We added an appendix subsection on analytical and accelerated physical optics. It cites Altintas and Çelik (2004) for planar triangle integration, Boag (2004) for multilevel acceleration across angular and frequency queries, and Xue et al. (2020) for multilevel PO with quadratic patches. The benchmark uses analytical triangle integration and a planar-triangle monostatic adaptation of the multilevel scheme. The Inference Efficiency subsection reports the hardware and the full OBJ-to-RCS timing protocol.[^1][^2][^3]

Q03. Meaning of the airplane geometries and validation scope

By “real-world targets,” we meant geometries representative of aircraft shapes encountered in practice. We now refer to them as airplane geometries from ModelNet. The dataset section specifies that the numerical reference responses are computed by PO on wavelength-refined meshes. Accordingly, the reported accuracy quantifies agreement with those PO reference responses.

Q04. Inference time compared with accurate PO and other methods

Across the two efficiency benchmarks below, EFSA records a shorter runtime in 11 of the 12 comparisons against a baseline at a given frequency. In the PO benchmark, all three methods start from the same airplane OBJ meshes and produce complete monostatic RCS maps. Timing includes preprocessing, postprocessing, and writing the numerical output.[^1][^3]


Method3 GHz4.5 GHz6 GHz
Analytic PO30.5230.6329.49
Multilevel PO9.1612.4014.37
EFSA10.7710.7010.52

In the end-to-end PO benchmark, EFSA is 2.80–2.86× as fast as analytical PO across the tested frequencies and also outpaces multilevel PO at 4.5 and 6 GHz. Multilevel PO is fastest at 3 GHz. The table reports mean seconds per target. See “Inference Efficiency” in the main text and “Additional Inference Timing Results” in the appendix.

EFSA also has a lower measured wall time on the airplane subset used to benchmark SC-RNet and our GraphSolver adaptation on the same GPU. This comparison uses their respective timing windows. SC-RNet starts from precomputed voxels, GraphSolver starts from OBJ, and both load weights before timing. EFSA is timed from process launch with OBJ input. The “Additional Inference Timing Results” appendix reports the learning baselines' full-angle runtimes and timing conditions, including a breakdown showing that scattering-center extraction accounts for most of SC-RNet's measured runtime. Its network inference time is close to the published result for preprocessed inputs.[^4][^5]


Method3 GHz4.5 GHz6 GHz
EFSA11.1711.4211.22
SC-RNet72.0372.2671.77
GraphSolver188.40190.37192.91

For the common six targets, EFSA produces a full-angle map in about 11 s per target at each frequency, with a shorter measured runtime than either learning baseline at every tested frequency. See “Additional Inference Timing Results” for the timing windows and the SC-RNet stage breakdown.

AuthorsOpenReview ↗

# Response to Reviewer QxJS part2

Q05. Contributions of the model components

The matched ablation shows that the full EFSA model lowers MAE on both datasets compared with direct prediction of the real and imaginary components of each facet contribution. It also lowers MAE relative to PO-init, demonstrating the benefit of the learned correction. See “Ablation Study” and the full-angle maps in the appendix. In a separate fixed-checkpoint diagnostic on ModelNet-Airplane, the amplitude-only correction lowers RCS MAE relative to PO-init at every tested frequency. The phase-only result is close to PO-init on this scalar metric.

Electrical-size encoding supplies a dimensionless descriptor of facet size relative to the wavelength. The ablation targets the main contribution, the PO-based parameterization of complex facet responses. The complex facet readout supplies the amplitudes and relative phases required for coherent field summation. Removing that representation would change the field summation model itself, so both trained variants retain the same complex output and summation. Both are also trained with the same facet and total-field complex supervision.


DatasetMethodMAE (dB)wMAERMSE (dB)
BasicFull EFSA6.138.438.81
Basicw/o PO-init8.8414.2812.23
BasicPO-init13.0913.9617.79
AirplaneFull EFSA4.637.726.48
Airplanew/o PO-init5.199.596.58
AirplanePO-init8.6912.6811.29

Method3 GHz4.5 GHz6 GHz
Amplitude only4.554.845.14
Phase only7.018.7510.23
PO-init6.958.7610.27

Full EFSA achieves the lowest MAE, wMAE, and RMSE on both datasets in the matched ablation. The fixed-checkpoint diagnostic separately shows that enabling the amplitude correction improves scalar RCS MAE across the tested frequencies. See “Ablation Study” and the appendix subsection “Additional Baseline and Correction Diagnostics.”

Q06. Sensitivity to triangulating the same surface

EFSA outperforms PO-init in MAE at every tested mesh resolution, and coarser input meshes reduce peak inference memory. The same checkpoint is evaluated directly on all tiers. Midpoint subdivision changes triangulation density while preserving the underlying piecewise-planar surface and its reference response. EFSA's error increases as the mesh is coarsened. The trade-off and angular cuts are reported in “Robustness to Mesh Discretization” and its appendix.


Facets (approx.)Mean electrical sizePO-init MAE (dB)EFSA MAE (dB)EFSA peak memory (GB)
3k3.313.156.882.0
1k6.519.488.760.5
0.6k8.723.579.630.3

On the coarsest mesh, EFSA runs with 0.3 GB of peak memory and keeps MAE below half that of PO-init on the same mesh. See “Robustness to Mesh Discretization” in the main text and appendix for the setup and angular cuts.

Thank you for your response. However, my major concern remains that the work is "limited to PO-generated data." I fully agree with the other reviewer's assessment that "the results show agreement with that approximation rather than validation against the actual scattering problem. There is no full-wave or experimental comparison, and no runtime evidence that the network is preferable to accurate PO integration."

Still, no real-world wave experimental validation was provided in the response. Today, using coding agent to train a model to fit simulated data is easy; the critical question is whether the proposed method can genuinely solve the physical problem in practice. Currently, there is insufficient evidence to demonstrate that this approach works in real-world scenarios.

Thank you for clarifying your concern. We would like to emphasize that physical optics (PO) is an established engineering approach to far-field radar cross-section (RCS) prediction for electrically large targets. In this regime, high-frequency approximations offer a practical balance between computational efficiency and engineering accuracy when their underlying assumptions hold. Mesa and Quevedo-Teruel [1] discuss the accuracy and efficiency of PO for electrically large microwave problems, particularly when far-field contributions dominate. Bai et al. [2] likewise describe its widespread application to electrically large target scattering and demonstrate efficient PO calculations for large ship and aircraft models. These studies support the practical relevance of PO-based modeling for the problem regime addressed in our work.

Full-wave methods, while providing a more comprehensive treatment of electromagnetic interactions, incur substantial computational and memory costs as electrical size increases [1, 2]. Consequently, their availability does not diminish the engineering value of efficient high-frequency methods. Our use of PO-generated data is motivated by this established accuracy–cost tradeoff. With our available computational resources and measurement facilities, completing full-wave simulations at the target scale and physical scattering measurements within the rebuttal period is not feasible.

Regarding the question of whether the network is preferable to direct PO integration, the manuscript already reports comparisons with the conventional PO algorithm identified by the reviewer, and our method achieves faster inference in the vast majority of the evaluated configurations. Thus, the reported results provide evidence of both agreement with the PO reference and a computational advantage over the compared implementation. We respectfully submit that an efficient surrogate for an established, practically relevant scattering approximation constitutes a meaningful contribution to electrically large RCS computation.

References

[1] F. Mesa and O. Quevedo-Teruel, “The validity of physical optics for microwave analysis: accuracy through simplicity,” Journal of Optics, vol. 27, no. 2, article 022501, 2025. doi:10.1088/2040-8986/ada6d8.

[2] J. Bai, S. Yang, and D. Su, “An Improved Z-Buffer Accelerated PO Method for EM Scattering from Electrically Large Targets,” Chinese Journal of Electronics, vol. 34, no. 2, pp. 475–482, 2025. doi:10.23919/cje.2024.00.025.

Reviewer_Zy8GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduced the EFSA, a representation that represents each mesh facet's PO scattering contribution analytically as a center-point prior x GNN predicted cross patch interference correction. The author claimed the EFSA is able to predict RCS in high-frequency regimes (coarser mesh) with less memory consumption, and achieve higher precision. The conclusion is partially supported by the experiments. The author show that EFSA method achieve better quantitative results in the geometry shape over the baseline from 3GHz - 6GHz. However, the evidence needs to be further justified to support the claim.

Critical strengths

Physical Interpretability The author proposed the EFSA, a method that write each facet’s contribution as an analytic PO prior times a learned complex correction, and lets graph nodes represent facets of any electrical size (with limited evidence). The model then predicts high-frequency angular RCS accurately on the original coarse meshes, with no wavelength-dependent re-meshing. The paper positions EFSA as a practical intermediate representation that does not tie graph resolution to the wavelength.

It shows that with an orthogonal decomposition of the contributions: amplitude contributed from the geometry together with the interference between facets, the RCS can be more efficiently calculated with less graph mesh when the frequency is high.

Critical weaknesses

The motivation is questionable. The paper argues that "mesh GNNs inherit a sub-wavelength constraint" (line 12). However, the mesh GNN is not the only method to predict the RCS with learning based method. Moreover, the author claimed that EFSA is able "to fundamentally decouple neural representations from these strict physical size requirements"(line 14). However, EFSA is also neural network based representation. The related work section is also missing regarding what the methodology of EFSA is based on and how is it different from the previous work.

The conclusion is weakly supported. The author claims that EFSA to be able to "models the RCS contribution of facets at arbitrary scales" (line 15), and "achieves high predictive accuracy while avoiding the graph-size explosion". However, the baselines are compared at 3, 4.5 and 6 GHz only, with no frequency sweep, or extrapolated frequencies showing that EFSA is able to perform better over arbitrary frequency range. Finally, the memory benefit is shown only for EFSA across its own mesh tiers; no baseline memory or runtime comparison is provided to support that it avoids the memory explosion while the baseline does not.

Questions for authors

Motivation

  1. The paper attributes the sub-wavelength constraints to neural representation and mesh GNN. Is it a property of the network architecture, or is it a property of the facet labelling (like MoM) ?
  2. Direct prediction methods or hybrid methods, does not use GNN, and therefore should not grow with f^2. Is it true that the proposed method outperforms over all other baselines that use direct or hybrid prediction method? How to justify the benefits of EFSA?

Related Work

  1. Please provide relate work section. How is the EFSA different from the other PO-based or hybrid neural approaches?

Experiments

  1. Please define what frequency range would be seen as "high frequency" that EFSA is targeting at. 15. Why 3 to 6 GHz and why these three evaluation frequencies? Are they in the “high-frequency” regime?

  2. To see the continuous range of how the GNN based method explode in memory and lost precision, I would expect the author to present the EFSA and baseline methods over a wider, continuous frequency range, with denser sample of frequency points, and report the memory, precision over the frequency.

  3. Given plot requested in 2, would the author explain where does each method stop working? In table 2, what does the MAE, wMAE, RMSE represents for?

  4. Would the author present the comparison of precision and memory for more baselines, including different GNN based methods, direct prediction methods, hybrid prediction method?

Presentation

  1. In Figure 3 and figure 4, what does the number of the color bar indicates for?
  2. The author proposed that "our framework accurately reconstructs these high-frequency variations" - in line 412, and "accurately reproduces the coherent superposition of scattering contributions from different facets." - line 445. Would the author elaborate how to tell those conclusions from the figure 3 and 4? It looks that every method's output in the figure 3&4 looks different from the ground truth.
  3. Table 1 seems to appear before table 2 but cite after table 2. Would be great if the table 1 is placed at the right place.
  4. In table 1, how is the * with and without different? The asterisk is not defined in the paper.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Zy8G part1

Response to Reviewer Zy8G

Thank you for asking us to clarify the source of the discretization constraint and quantify the scope of the results. We revised the motivation and added related work, an end-to-end timing comparison with PO, and full-angle timings for two learning baselines.

Z01. Where the subwavelength constraint comes from

The constraint comes from the representation and the supervision used by the current-based models discussed in our paper. Numerical current labels are generated on a mesh fine enough to resolve the current. When one centroid value represents a whole facet, the radiation approximation also assumes limited phase variation within that facet. A GNN that uses this discretization inherits its resolution. For a fixed surface area, the quadratic dependence on frequency in our appendix is a lower bound on the number of facets under a maximum-edge constraint proportional to wavelength. We clarified these conditions in the Introduction and the section on the analytical physical-optics prior.

Z02. What facet sizes the evidence supports

EFSA reduces peak inference memory as the input mesh is coarsened. On Basic Geometry, the same checkpoint outperforms PO-init in MAE at every tested mesh resolution, including the coarsest tier. The MAE increases as the mesh is coarsened, so we report both the memory savings and the accuracy cost in “Robustness to Mesh Discretization” and its appendix table.


Facets (approx.)Mean electrical sizePO-init MAE (dB)EFSA MAE (dB)EFSA peak memory (GB)
3k3.313.156.882.0
1k6.519.488.760.5
0.6k8.723.579.630.3

At the coarsest resolution, EFSA uses 85% less peak memory than it does on the finest tested mesh and attains less than half the MAE of PO-init on that same coarse mesh. See “Robustness to Mesh Discretization” in the main text and the appendix.

Z03. Comparison with direct and hybrid prediction

EFSA achieves lower mean MAE than SC-RNet on ModelNet-Airplane. EFSA also achieves lower MAE on Basic Geometry at 3 and 6 GHz. EFSA also explicitly represents complex facet contributions and sums them coherently before computing target RCS. The hybrid result in Z09 shows that a target-level predictor can achieve lower scalar RCS error in a separate setting. The facet-count bound applies to methods that refine the mesh to satisfy a wavelength-dependent edge limit.


DatasetMethod3 GHz4.5 GHz6 GHz
BasicSC-RNet7.577.297.80
BasicEFSA6.517.426.71
AirplaneSC-RNet5.995.976.06
AirplaneEFSA4.514.915.22

EFSA achieves lower RCS MAE (dB) in five of the six dataset–frequency comparisons with SC-RNet. SC-RNet has a slightly lower MAE on Basic Geometry at 4.5 GHz. See “Performance on Basic Geometry” and “Performance on ModelNet-Airplane” for the full results.

Z04. Related work

We added a Related Work subsection to the appendix. It discusses graph-based surface-current models such as PhiGRL, GraphSolver, and PdEgatSCL, whose outputs are used to synthesize the scattered field.[^1][^2][^3]

It also discusses direct predictors such as RCS-PointNet++, SC-RNet, ECHO, and GTC-3DEM, including architectures that use a graph encoder to predict a target-level response.[^4][^5][^6][^7]

The appendix distinguishes PO-guided correction of the target RCS from EFSA's correction of complex facet contributions before coherent summation.[^8]

A separate appendix subsection discusses analytical triangle integration and accelerated PO.[^9][^10][^11]

Z05. Frequency range and electrical size

We characterize the electrical size of a triangular facet by dmax⁡/λd_{\max}/\lambda, with dmax⁡d_{\max} the longest edge. The tested Basic Geometry tiers include facets several wavelengths across, as quantified in the mesh robustness analysis. The main comparison uses the endpoints and midpoint of the benchmark band defined in “Dataset Preparation” to sample the lower, middle, and upper parts of that band.


SettingFrequency (GHz)
Benchmark band3–6
Frequency step0.3
Main comparison frequencies3, 4.5, 6

The frequency grid is specified in “Dataset Preparation.” The tested electrical sizes are in the “Robustness to Mesh Discretization” table.

AuthorsOpenReview ↗

# Response to Reviewer Zy8G part2

Z06. Denser frequency evaluation and extrapolation

The main evaluation focuses on generalization to held-out geometries within the benchmark band. EFSA maintains similar MAE across the denser in-band sweep on ModelNet-Airplane. We also evaluated the existing checkpoints without retraining at an out-of-band frequency on both datasets, using refined-mesh PO as the reference. Frequency extrapolation is a supplementary check, with results at the tested frequency reported in “Supplementary Frequency Evaluation of EFSA.”


Frequency (GHz)EFSA MAE (dB)
34.59
3.34.63
3.64.67
3.94.74
4.24.77
4.54.89
4.85.02
5.15.21
5.45.19
5.75.13
65.17

DatasetOut-of-band frequency (GHz)EFSA MAE (dB)
Basic Geometry7.27.58
ModelNet-Airplane7.25.50

EFSA's MAE on ModelNet-Airplane varies by only 0.62 dB across the 11 in-band frequencies. The appendix subsection “Supplementary Frequency Evaluation of EFSA” gives the protocol for both the in-band sweep and the single out-of-band check.

Z07. Graph cost of wavelength-dependent mesh refinement

EFSA keeps the input graph fixed across the tested frequencies, avoiding the graph growth required by wavelength-dependent mesh refinement. For a surface-current GNN subject to a maximum edge length proportional to the wavelength, the appendix derives a facet-count lower bound that grows quadratically with frequency before rounding. With a node per facet and a fixed feature width, the required node-feature storage inherits this lower bound. The appendix gives the graph-size estimates for the airplane meshes. The fixed-mesh GPU memory comparison is discussed in Z09.


Frequency (GHz)Median facet-count lower bound over 30 airplanes
3131,119
6524,473.5

The table shows the theoretical bound at h=λ/10h=\lambda/10. The derivation is in the appendix subsection “Comparison of Facet Counts After Electrically Small Mesh Discretization at Different Wavelengths.”

Z08. Conditions under which accuracy degrades

EFSA achieves lower MAE than PO-init and the coarse-mesh GraphSolver adaptation at every tested frequency on both datasets. On the fixed meshes in the main accuracy table, the errors of PO-init and GraphSolver increase with frequency. The mesh robustness experiment also shows that EFSA's error increases as the input mesh is coarsened. An operational failure threshold depends on the tolerated error, mesh policy, and available hardware.


DatasetMethod3 GHz4.5 GHz6 GHz
BasicPO-init9.9013.2816.29
BasicGraphSolver51.3255.5958.90
BasicEFSA6.517.426.71
AirplanePO-init7.008.7910.38
AirplaneGraphSolver27.5630.1332.30
AirplaneEFSA4.514.915.22

At 6 GHz, EFSA keeps MAE at roughly half that of PO-init or lower on both datasets, with errors of 6.71 dB on Basic Geometry and 5.22 dB on ModelNet-Airplane. See the main accuracy table and “Robustness to Mesh Discretization” for the frequency and mesh effects.

Z09. Additional learning baselines

The main accuracy table shows that EFSA has lower mean MAE than the tested SC-RNet and GraphSolver implementations on ModelNet-Airplane. The supplementary fixed-mesh comparison with the PdEgatSCL adaptation also shows that EFSA achieves lower MAE across the tested frequency sweep while using less peak allocated GPU memory. Both methods use the same test entries, directions, outer batch size, and GPU in that comparison.[^5][^2][^3]

In the separate full-angle comparison, adding a learned target-level residual to analytical PO achieves lower scalar RCS MAE than EFSA. EFSA's facet-level output provides complex contributions that can be combined coherently to form the scattered field. For related work on target-level PO correction, see Rong & Wan (2026)[^8].


MethodPer-frequency MAE range (dB)Peak allocated memory (MiB)
EFSA4.48–5.36839
PdEgatSCL adaptation10.61–12.082191

Method3 GHz4.5 GHz6 GHz
Analytic PO + residual MLP3.944.034.09
EFSA5.415.736.01

In the fixed-mesh sweep, EFSA uses about 62% less peak allocated GPU memory than the PdEgatSCL adaptation while improving MAE at every tested frequency. The separate full-angle comparison quantifies the hybrid baseline's scalar MAE advantage. See “Additional Baseline and Correction Diagnostics” in the appendix for both protocols.

AuthorsOpenReview ↗

# Response to Reviewer Zy8G part3

Z10. Meaning of the error metrics

Section 3.2 defines MAE as the average absolute RCS error and RMSE as the square root of the average squared RCS error, both evaluated in the logarithmic domain. Equation (8) defines wMAE as the average absolute error weighted by max⁡(yi,1)\sqrt{\max(y_i,1)}, with yiy_i the numerical reference RCS value in dBsm. This gives stronger reference responses more weight.

Z11. Color scales in Figures 3 and 4

The colors show monostatic RCS in dBsm at each angular sample. The predictions and reference in each row share a color scale. We labeled the color bars and stated the shared-scale convention in both figure captions.

Z12. What the RCS maps show

Figure 3 shows where EFSA captures sharper angular structure than SC-RNet. In the cylinder example in the top row, EFSA reproduces several bright and dark bands around the central peak in the PO reference, while the SC-RNet map is smoother. Some contrasts and null depths still differ from the reference. In the top row of Figure 4, both EFSA and the PO reference show the broad central low-RCS region and adjacent bands. These patterns illustrate rapid RCS variation with direction at a fixed frequency. The field summation used to produce EFSA's RCS maps is defined in Section 2.1.

Z13. Order of the tables

We moved the mesh-discretization table to the Robustness to Mesh Discretization subsection, next to its first detailed discussion. The main accuracy table now precedes it in the revised source.

Z14. Meaning of the asterisk

We now explain the notation in the main text next to the mesh-discretization table. An asterisk denotes PO-init using the facet-center estimate without the learned correction. Unmarked rows report EFSA, and a slash indicates that neural inference memory is unreported for PO-init.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.