{"id":"152d4a89-8cb3-440b-b7e8-7178093c6b89","arxiv_id":"2509.11087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Modeling the complex acoustic scattering field as spherical harmonic coefficients inside a hash-encoded neural network improves 3D synthetic aperture sonar reconstruction over isotropic neural baselines.","lead":"A team at Arizona State and Dartmouth built a neural network that reconstructs 3D sonar scenes by modeling how sound scatters differently in different directions, using spherical harmonic coefficients at each point. The method trains directly from raw time-of-flight signals and beats existing isotropic neural and backprojection baselines on simulated data, with qualitative improvements on real in-air and underwater sonar.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Factorized forward model breaks acoustic reciprocity; synthetic benchmark may not exercise real bistatic scattering","rationale":"The central claim is that SH-SAS improves 3D reconstruction quality on synthetic and real SAS benchmarks by capturing directional scattering. The only quantitative support is synthetic (Tables 1 and 3); real data results are qualitative. The load-bearing concern is whether the forward model can represent real bistatic scattering. The reader identified the factorized SH times Lambertian form as the weakest assumption. I agree, but I sharpen it with a concrete physical law: the factorization violates acoustic reciprocity, meaning the model class excludes physically valid scattering functions. This is not a disagreement with current consensus; it is an internal inconsistency with a known symmetry of the wave equation. The synthetic benchmark is also under-specified: Sec. 5.1 does not state the scattering law used by the CUDA/OptiX renderer. If the simulator uses the same factorized model or a Lambertian model, then the SH advantage on synthetic data may be an artifact of matching the forward model, not evidence of capturing true directionality. The proposed test—simulating with a reciprocal, non-factorized BRDF—can settle whether the factorization actually lands. If SH-SAS still wins, the concern is resolved and the claim is robust. If not, the paper should restrict its claims to the factorized model and provide real-data metrics before claiming general improvements. The reader's conditional verdict remains appropriate; no verdict change is needed.","tokens_in":15389,"tokens_out":15874,"duration_ms":204517,"concrete_test":"Use the released simulator (or modify it) to generate synthetic SAS data for a target whose surface scattering is a reciprocal, non-factorized BRDF, e.g., a specular lobe h(omega_i, omega_r) = c * max(0, n·h)^p / normalization, where h is the half-vector between incident and outgoing directions; this form is symmetric under TX/RX swap. Train SH-SAS and Reed et al. on this data and compare Chamfer distance. If SH-SAS no longer outperforms Reed et al. (or the gap shrinks materially relative to Tab. 1), the factorized ansatz is the limiting factor. As a secondary check, evaluate a trained SH-SAS model on a symmetric TX/RX pair and its swapped counterpart; equal predictions are required for reciprocity, and any asymmetry directly demonstrates the model violates a physical constraint.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (6), the scattering amplitude is written as sigma_s(x, theta, phi) * g(omega_x), where (theta, phi) is the direction from receiver o_R to x, and g(omega_x) = max(0, n(x) · (o_T - x)/||o_T - x||) is the Lambertian cosine of the transmitter direction. This factorized ansatz is not invariant under exchanging o_T and o_R: swapping the two foci changes the argument of sigma_s to the new receiver (old transmitter) direction and the cosine to the new transmitter direction, and these generally differ. A passive linear acoustic medium obeys reciprocity, so the true scattering amplitude must be symmetric under this exchange. The model class in Eq. (6) therefore excludes physically valid reciprocal scattering functions (unless sigma_s is a fixed multiple of the cosine, which is not what is learned). The synthetic data generator (Sec. 5.1) is described only as 'geometric acoustics' with diffraction/sub-wavelength effects excluded; it is not stated what scattering law is used. If the simulator uses a reciprocal law, then the synthetic data lie outside the model class, yet SH-SAS outperforms the isotropic baseline—the improvement could come from the SH terms compensating for non-reciprocal artifacts or normal errors rather than from capturing true bistatic directionality. This directly weakens the abstract's claim that the model 'captures directional scattering' and that improvements extend to real SAS.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SH-SAS, an implicit neural representation for 3D synthetic aperture sonar reconstruction. Instead of modeling each voxel as an isotropic complex scattering density as in Reed et al., SH-SAS outputs spherical harmonic (SH) coefficients up to degree L at each point; the DC term provides an isotropic density proxy and higher-order terms encode directional scattering. Training is end-to-end from 1D time-of-flight signals through a differentiable point-based forward model with ellipsoidal sampling and pulse deconvolution. Experiments on synthetic scenes, in-air AirSAS data, and underwater SVSS data report improved reconstruction quality (Chamfer distance, IoU, precision, F1) over time-domain backprojection and Reed et al., along with faster convergence and better performance under sparse views.","tokens_in":15758,"tokens_out":9042,"duration_ms":113820,"significance":"If the claims hold, this is a useful step for SAS: replacing isotropic scattering with a compact directional SH representation inside a neural volumetric inversion, using a lightweight hash-encoded MLP and training directly from raw ToF signals. The paper includes a credible synthetic evaluation, a thoughtful ablation of SH levels, and a novel-view synthesis experiment. However, the evidence has important gaps: the synthetic simulator's scattering law is not specified, the forward model's factorization breaks acoustic reciprocity, real-data results are only qualitative, and per-scene synthetic metrics are not uniformly better than the baseline. These issues must be addressed before the contribution can be fully assessed.","major_comments":[{"comment":"The factorized forward model sigma_s(x, theta, phi) * g(omega_x) is not invariant under exchanging TX and RX, violating acoustic reciprocity for passive linear media. The model class therefore excludes physically reciprocal scattering functions. The manuscript does not discuss or justify this approximation. Please either provide a physical justification, demonstrate empirically that the learned field is approximately reciprocal (e.g., by comparing predictions for swapped TX/RX on a trained scene), or explicitly state this as a limitation affecting the physical interpretability of the SH coefficients.","section":"Sec. 4.4, Eq. (6)"},{"comment":"The synthetic data generator is described only as 'geometric acoustics' with diffraction and sub-wavelength effects excluded; the scattering law is not stated. If the simulator uses the same factorized model as the reconstruction, the synthetic experiments mainly test inversion self-consistency rather than the ability to capture true bistatic directionality. If it uses a reciprocal scattering law, the data lie outside the model class and the gains over the isotropic baseline need explanation. Please specify the scattering law used in the simulator and, if possible, run a test where the simulator includes a genuinely bidirectional (e.g., specular or angle-dependent) response to verify that the SH terms capture it.","section":"Sec. 5.1"},{"comment":"Real-data evaluations (AirSAS and SVSS) are qualitative only. Since AirSAS uses known 3D-printed objects, quantitative metrics (e.g., Chamfer distance or IoU against ground-truth meshes) should be reported. The SVSS result is validated only against expected geometry. Without quantitative real-data evidence, the claim that SH-SAS generalizes to real SAS is not strongly supported. Please add metrics for at least the AirSAS objects and discuss the SVSS case in terms of measurable criteria where possible.","section":"Secs. 6.2 and 6.6"},{"comment":"The average metrics in Table 1 hide per-scene variability. In Table 3, Reed et al. achieves better or comparable IoU/F1 on Bunny and Dragon point clouds (e.g., Bunny IoU 0.412 vs 0.398; Dragon F1 0.529 vs 0.526). The blanket claim of 'better performance' is therefore not uniformly supported. Report per-scene results with error bars (multiple runs or seeds) and discuss the cases where the SH model does not improve over the isotropic baseline.","section":"Sec. 6.1 and Table 3"},{"comment":"Key hyperparameters are not reported: the occlusion scaling factor zeta in Eq. (4) and the loss weights lambda1-lambda5 in Eq. (8). The text also states that priors are 'disabled for the Simulated and SVSS datasets' without clarifying whether this applies to both SH-SAS and Reed et al. This ambiguity affects the fairness of the comparison and the reproducibility of the results. Please provide all hyperparameter values and state explicitly which regularization terms are used for each method and dataset.","section":"Sec. 5.4 and Eq. (8)"}],"minor_comments":[{"comment":"Typo: 'Armadilo' should be 'Armadillo'.","section":"Fig. 3 caption"},{"comment":"In the ellipsoid definition, 'b=c= sqrt(a^2 - (d/2)^2)' is confusing: c is already used for sound speed. It should be b = sqrt(a^2 - (d/2)^2).","section":"Supplement Sec. 9"},{"comment":"State explicitly whether the integral is over a volumetric shell or a surface and clarify the meaning of dx.","section":"Eq. (6)"},{"comment":"Specify the normalization and convention for the spherical harmonics (real vs complex, Condon-Shortley phase) used in the implementation.","section":"Sec. 4.2"},{"comment":"The transmitter beam pattern b_T(x) is never defined; its value/functional form should be given for each dataset, as it affects the forward model.","section":"Sec. 5.4"},{"comment":"LScatterTV applies total variation to the complex field sigma_s; clarify whether it acts on the magnitude, phase, or complex values.","section":"Eq. (8)"},{"comment":"The paper states code will be released after peer review but also includes code in the supplement; align these statements to avoid confusion.","section":"Sec. 5.4"},{"comment":"The SH-level ablation is reported only qualitatively. A quantitative table (e.g., Chamfer/IoU for L=1,2,3) would strengthen the claim that higher SH degrees are beneficial.","section":"Sec. 6.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a direct extension of Reed et al. with overlapping authorship; the incremental contribution is clear but the validation infrastructure and physical assumptions need substantial strengthening before the claims can be taken as established. The reciprocity issue also affects the baseline, so it should not be framed as a fatal flaw unique to this method, but the authors should acknowledge it and provide empirical evidence about how much it matters in practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper proposes SH-SAS, an implicit neural representation where each point carries complex spherical harmonic coefficients for the acoustic scattering field. It is trained directly on 1D time-of-flight signals, and it outperforms the isotropic volumetric baseline (Reed et al.) and backprojection on synthetic scenes by a clear margin (Chamfer distance roughly halves). The higher-order SH terms demonstrably help, and the method converges in fewer iterations. Those are real, reproducible results.\n\nThe novel piece is tying the SH representation to a complex-valued scattering field inside an INR, and it appears to work well. The synthetic experiments are cleanly described, with open simulator and (promised) code. The ablation on SH level and the threshold robustness study are useful. The claim of improved real-data reconstruction is plausible but only supported qualitatively; there are no numerical metrics for AirSAS or SVSS, no error bars anywhere, and almost no hyperparameter detail. That limits how much you can trust the generalization beyond the specific configurations tested.\n\nThe bigger conceptual problem is the forward model. Eq. (6) factorizes the scattering amplitude as sigma_s(x, theta, phi) * g(omega_x), where the SH term depends only on the receive direction and the Lambertian cosine only on the transmit direction. This factorized form is not invariant under swapping transmitter and receiver; I checked, and it violates acoustic reciprocity. A passive linear medium should be reciprocal. Even if the learned sigma_s is complex and can absorb some of the asymmetry, the model class is excluding physically valid reciprocal scattering functions. That does not necessarily destroy the reconstruction capability—it may still be a useful empirical parameterization—but it contradicts the paper's claim that it 'captures directional scattering' in a physically grounded way.\n\nRelatedly, the synthetic data is generated with the same point-based geometric acoustics forward model that the reconstruction uses, so those experiments test whether the optimization and representation can invert that model, not whether the model matches real scattering physics. The stress-test note raised this, and I think it's right. The real-data results are the only independent evidence, and they are qualitative. The authors do acknowledge several limitations in the discussion: no diffraction, no multiple scattering, and sensitivity to noise and narrowband signals. That is honest.\n\nIs this worth engaging with? For someone working in neural fields for sonar or SAS reconstruction, yes. The representation is a natural and useful step beyond isotropic volumetric methods, and the empirical gains are consistent. But I would not yet treat the method as a validated physical scattering model. A serious referee should push for quantitative real-data metrics, error bars, and ideally a test on data generated by a different forward model (e.g., a wave-based simulator or measured ground truth) to break the circularity. They should also ask whether the factorization can be made reciprocal without losing expressiveness.\n\nMy recommendation: send it to peer review. It is not a fundamentally flawed paper, and the central engineering contribution is solid. The flaws are in the validation and in the physical interpretive claims, not in the core idea.\n\nBest.","headline":"Useful representation for directional SAS, but the forward model breaks reciprocity and the synthetic validation is circular; real claims rest on qualitative images.","tokens_in":16175,"tokens_out":2692,"would_cite":true,"duration_ms":31099,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Parameterizing acoustic scattering as spherical-harmonic coefficients inside a neural field improves 3D sonar reconstruction from raw 1D time-of-flight signals.","keywords":["synthetic aperture sonar","implicit neural representation","spherical harmonics","acoustic scattering field","time-of-flight","3D reconstruction","neural rendering","anisotropic scattering"],"falsifier":"A controlled experiment with a flat, smooth target at a grazing angle, where a strong specular return produces a joint transmit/receive direction dependence that the product of an SH receive function and a Lambertian incident cosine cannot fit, would falsify the central claim if the measured transient residuals remain large at high SNR even as SH degree L increases. More simply: if increasing L from 1 to 3 never improves Chamfer distance on any benchmark, then the directionality modeling is not carrying the improvement.","tokens_in":15324,"feed_emoji":"📡","tokens_out":5122,"duration_ms":59648,"temperature":0.7,"pith_summary":"SH-SAS aims to show that replacing the isotropic scattering-density assumption in neural volumetric sonar reconstruction with a spherical-harmonic (SH) expansion of the complex scattering field produces higher-fidelity 3D reconstructions. The DC term of the SH expansion supplies a stable density proxy, while higher-order coefficients capture direction-dependent scattering with minimal parameter cost. Training runs directly on 1D time-of-flight signals through a differentiable forward model, so no beamformed intermediate images are needed. On simulated, in-air, and underwater benchmarks, the method reports lower Chamfer distance and better IoU/precision/F1 than previous isotropic neural methods and time-domain backprojection.","feed_headline":"Spherical-harmonic fields sharpen 3D sonar reconstruction","feed_subtitle":"Neural field trained on raw 1D echoes captures direction-dependent scattering that isotropic models miss.","key_machinery":"The central object is the spherical-harmonic expansion of the scattering field, σ_s(x,θ,ϕ)=Σ_{l=0}^{L}Σ_{m=-l}^{l} c_{l,m}(x)Y_{l,m}(θ,ϕ), where Y_{l,m} are the spherical-harmonic basis functions on the unit sphere and c_{l,m}(x) are complex learnable coefficients. Spherical harmonics are an orthonormal angular basis, so a small set of coefficients captures smoothly varying directionality. The zeroth-order coefficient supplies the isotropic density and surface normals, while higher orders encode anisotropic scattering. The forward model integrates these values over constant-time-of-flight ellipsoids with transmitter/receiver directivity and a Lambertian cosine incident-angle term, synthesizi","core_discovery":"The central claim is that the complex acoustic scattering field at each spatial point can be compactly represented by complex spherical-harmonic coefficients up to degree L, predicted by a lightweight hash-encoded MLP, and that optimizing this field directly against 1D time-of-flight signals yields 3D synthetic aperture sonar reconstructions with sharper geometry and fewer artifacts than isotropic neural volumetric baselines or time-domain backprojection. The zeroth-order SH coefficient defines the isotropic density, whose gradient gives surface normals; higher orders add angular resolution, which the paper argues is the inductive bias that improves reconstruction under occlusions, sparse ap","pith_inferences":["Inference: If the factorized SH-Lambertian scattering model is accurate, the same parameterization could benefit radar and LiDAR neural reconstruction, where view-dependent coherent scattering also matters; the paper lists this as future work.","Inference: A quantitative test of the directionality claim would compare SH-SAS with the isotropic baseline on targets with known specular or strongly bistatic responses, where the factorized model should be wrong; the paper's real-data experiments are primarily qualitative.","Inference: The stated sensitivity of higher-order SH terms to noise suggests that an adaptive degree per spatial region, or a spectral prior on the SH coefficients, could extend the method to low-SNR and narrowband acquisitions.","Inference: Since the DC term alone reproduces the isotropic baseline, the practical gain of the method could be measured as a function of view diversity; one would expect the largest improvement in sparse-aperture regimes, consistent with the 20%-views experiment."],"forward_implications":["Higher SH degrees (up to L=3) monotonically improve reconstruction quality in the paper's ablation, indicating that directionality, not just density, is what the network learns.","Because training uses raw 1D ToF signals, the same architecture could be applied to other coherent time-of-flight sensors by swapping the forward model.","The representation handles sparse aperture subsampling (20% of views) with less streaking and fragmentation than backprojection or the isotropic neural baseline.","The learned field is stable under iso-surface threshold changes, meaning mesh extraction is less sensitive to an arbitrary threshold.","On a single-look underwater acquisition, the method recovers the expected one-sided flat profile and suppresses returns beneath the target, consistent with occluded geometry."],"fun_headline_variants":["SH-SAS: neural field learns directional sonar from raw echoes","Spherical-harmonic neural field sharpens 3D sonar images","Complex SH field models directional sonar scattering","Neural SH coefficients capture angle-dependent sonar returns"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true acoustic scattering at each point factors into a spherical-harmonic function of the receiver-to-point direction times a Lambertian cosine of the transmitter-to-point direction; scattering that depends on both directions jointly—such as specular reflection or diffraction—cannot be captured, and the paper explicitly excludes diffraction and sub-wavelength wave effects from its forward model.","fun_headline_variants_meta":{"raw":{"variants":["SH-SAS: neural field learns directional sonar from raw echoes","Spherical-harmonic neural field sharpens 3D sonar images","Complex SH field models directional sonar scattering","Neural SH coefficients capture angle-dependent sonar returns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000752,"raw_usage":{"total_tokens":3180,"prompt_tokens":739,"completion_tokens":2441,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2373}},"tokens_in":483,"tokens_out":2441,"duration_ms":21581,"temperature":1.0,"reasoning_tokens":2373,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:05:45.177452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment with a flat, smooth target at a grazing angle, where a strong specular return produces a joint transmit/receive direction dependence that the product of an SH receive function and a Lambertian incident cosine cannot fit, would falsify the central claim if the measured transient residuals remain large at high SNR even as SH degree L increases. More simply: if increasing L from 1 to 3 never improves Chamfer distance on any benchmark, then the directionality modeling is not carrying the improvement.","supporting_citations":[],"review_version":1}