{"id":"5b0fb30f-f28a-47f0-adcf-46b87a10b90a","arxiv_id":"2507.06826","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A physics-informed neural network enforces momentum and continuity equations on all four Ambisonic channels, outperforming data-only and W-channel-only baselines for room impulse response interpolation.","lead":"This paper adds physics-based constraints to a neural network that interpolates spatial room impulse responses, and shows the constraints improve accuracy on Ambisonic audio data. The method could make immersive audio and virtual reality applications work with fewer microphone measurements.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The momentum and continuity losses are only valid if HARP2's FOA channels satisfy u = -v/(rho0 c0) exactly; the paper cites this mapping but never verifies the simulator's normalization, sign, or near-field behavior.","rationale":"The reader's weakest assumption correctly identifies the FOA-to-particle-velocity mapping as the load-bearing premise. My independent derivation of Eqs. (11)-(12) from Eqs. (1)-(2) confirms that both priors depend on u = -v/(rho0 c0) without any free constant; if the simulator uses a different normalization (e.g., sqrt(3) in SN3D first-order channels or sqrt(2) in FuMa) or a different sign, the priors are inconsistent with the data. I considered whether the evaluation could still be valid because the data term anchors measured points, but this does not rescue the central claim: the paper explicitly asserts the priors 'more precisely reflect the physical principles of FOA RIRs' and credits them for the improvement. The proposed test of computing the PDE residuals on HARP2's ground truth directly settles whether the priors are physically correct for the data generator. I find no other more fragile step: the PDE derivation is internally sound, the network architecture is standard, and the empirical comparison, while lacking seeds and statistics, is a secondary concern. Thus the reader's CONDITIONAL verdict stands; no change is needed, but the HARP2-mapping check should be the condition for acceptance.","tokens_in":9282,"tokens_out":14339,"duration_ms":171294,"concrete_test":"Download HARP2 and simulate the paper's configuration (10 shoebox rooms, 5 cm grid, 8 kHz, 100 ms). Compute the residuals of Eqs. (11) and (12) directly on the ground-truth FOA RIRs using the paper's assumed mapping, via finite differences or a separate automatic-differentiation field, and inspect HARP2's channel definitions for the actual scaling/sign relative to pressure and particle velocity. If the residuals are not near zero (relative to the norms of the terms), re-run the PI-DANF comparison with the corrected mapping to see whether the reported NMSE/PCC gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eqs. (11)-(12) are derived solely from the mapping in Eqs. (7)-(8): under SN3D, w is pressure and v = [x,y,z] is taken to be -rho0 c0 u. If HARP2's FOA output uses a different convention (ACN/SN3D has a sqrt(3) gain on first-order channels, FuMa sets W = p/sqrt(2), or the image-source directivity model produces X/Y/Z as p times direction cosines instead of true pressure-gradient/particle-velocity), the proposed priors are not zero on the ground-truth data they are meant to regularize. The paper treats Eq. (7) as a given (cited to [31]) and never checks it against HARP2. This is load-bearing because the improvement of PI-DANF over the Naive W-only wave-equation prior is attributed entirely to these two priors. If the mapping is off by any scalar factor or sign, the network must trade off data fidelity against an incorrect physical penalty, and the reported gains could be an artifact of regularization strength rather than evidence that the FOA-specific physics prior is correct. The risk is concrete: HARP2 is an image-source simulator, so near-field gradient terms and the exact Ambisonic normalization are exactly the details that are easy to get wrong and are not discussed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PI-DANF, a physics-informed extension of the direction-aware neural acoustic field (DANF) for first-order Ambisonic (FOA) room impulse responses. Starting from the linearized momentum equation and the continuity equation, the authors derive two additional losses, Eqs. (11) and (12), that couple the predicted W channel to the (X, Y, Z) channels through spatial and temporal derivatives. The key physical input is the standard FOA relation in Eqs. (7)-(8), where W is pressure and the vector of first-order channels is proportional to particle velocity. The training objective in Eq. (17) combines data fidelity with these two physics priors using adaptive weights. Experiments on 10 simulated shoebox rooms from HARP2, with 250 or 500 measured training positions, compare vanilla DANF, a wave-equation prior applied only to the W channel, and the proposed method. The paper reports NMSE and Pearson correlation and claims that PI-DANF performs best because of the FOA-specific priors.","tokens_in":9487,"tokens_out":14275,"duration_ms":175207,"significance":"If the results hold, the contribution is a simple and well-motivated extension of PINNs from scalar pressure fields to first-order Ambisonic fields, giving two cheap PDE residuals that couple all four FOA channels. The algebraic derivation of Eqs. (11)-(12) from Eqs. (1)-(2) and (7) is correct, and the method is easy to reimplement. The experimental comparison covers 10 rooms, two training-set sizes, and two strong baselines. I see no circularity problem: the priors are standard PDEs rather than quantities fitted to the target, and the HARP2 ground truth is generated from the same wave-propagation physics. However, the quantitative support for the central claim is moderate: the evaluation is simulation-only, there is one training run per condition, no statistical significance tests are reported, and the mapping from FOA channels to particle velocity in Eq. (7) is assumed rather than verified against the simulator.","major_comments":[{"comment":"The relation u(r,t) = -v(r,t)/(rho0 c0) is load-bearing for both proposed priors in Eqs. (11) and (12), but it is taken as given from [31] and is never checked against the actual data generator HARP2. Different Ambisonic conventions differ in the scaling of the first-order channels relative to W (for example, some SN3D formulations introduce a sqrt(3) factor, and FuMa uses different gains), and different image-source directivity implementations can differ in sign. If the simulator's FOA output is not exactly Eq. (7), the physics priors are not zero at the ground truth they are intended to regularize, and the reported gain over the naive W-only prior could be an artifact of regularization strength rather than evidence that the FOA-specific physics prior is correct. Please add a direct numerical verification: simulate a known plane-wave or single-image-source scenario in HARP2 and confirm that w = p and v = -rho0 c0 u, or explicitly state and implement the exact normalization and sign convention of HARP2's FOA output.","section":"Section 2.2, Eqs. (7)-(8)"},{"comment":"The headline claim that PI-DANF \"consistently performed best\" is supported only by boxplots over 10 rooms, with a single random room-source configuration and a single training run per condition. There are no confidence intervals, standard deviations over random seeds, or paired significance tests. The differences between the proposed method and the naive W-only prior appear modest in the figures, and without significance testing the reader cannot rule out that the differences are within random variation. Please report per-room paired differences with bootstrap confidence intervals or a paired test, and ideally repeat training with several seeds.","section":"Section 4.2, Figures 4 and 5"},{"comment":"The comparison against the \"Naive\" baseline does not isolate the channel-coupling property of the proposed priors. Because the wave-equation prior is applied only to the W channel, the improvement of PI-DANF could stem simply from regularizing the otherwise-unconstrained (X, Y, Z) channels rather than from the cross-channel momentum/continuity structure. A fair baseline is to apply the wave-equation prior separately to each of the four channels (or to all channels) with the same training budget. If the proposed method still outperforms this baseline, the claim that the FOA-specific coupled priors are the source of the gain would be on much firmer ground. An ablation that removes Lmomentum or Lcontinuity individually would also clarify the contribution of each prior.","section":"Sections 4.1-4.2"},{"comment":"The adaptive weighting in Eq. (17) is initialized from epsilon_data = 1.0 and epsilon_prior = 0.1, and the authors do not examine sensitivity to these initial values or to the number of Latin-hypercube samples used for the priors. Since the central comparison is between methods that use different physics losses, a brief sensitivity study or at least a statement that the results are stable over a range of initialization values would help rule out tuning artifacts.","section":"Section 4.1, Eq. (17)"}],"minor_comments":[{"comment":"The sentence saying that W matches sound pressure and the later statement that under SN3D w is equivalent to p should be reconciled with the exact scaling of the first-order channels; please state explicitly whether the (X, Y, Z) channels include any normalization gain relative to W.","section":"Section 2.2"},{"comment":"The manuscript should specify the version of HARP2 and the exact settings used to generate FOA RIRs, including the Ambisonic normalization, so that the experiments are reproducible.","section":"Section 4.1"},{"comment":"The time range T used in the physics priors is not explicitly defined relative to the 100 ms window of the RIRs; please state that T corresponds to the same early-time window and clarify how automatic differentiation treats the discrete time samples.","section":"Section 4.1, Eq. (17)"},{"comment":"The colorbar label is given as \"x10^2\" without units; adding pressure units and using a common color scale across rows for each time step would make the visual comparison easier to interpret.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the core derivation is sound. The most important revision is to verify Eq. (7) against HARP2 and to strengthen the experimental evidence with significance testing and additional baselines. If those points are addressed, I would be willing to recommend acceptance. The self-citation [39] is appropriate and does not load the central result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: the paper does one clean thing, and it does it correctly. Prior PINN work on RIRs regularized scalar pressure with the wave equation. This paper writes down the linearized momentum and continuity equations for first-order Ambisonic channels, tying the W and X/Y/Z channels through partial derivatives. That is new relative to the cited DANF and to [27], [28]. The derivation in Section 3.2 is correct: substituting w for p and -v/(rho0 c0) for u gives equations (11) and (12). I checked signs and units and they are consistent.\n\nWhat it does well: the experiments are clean. Ten simulated shoebox rooms, 250 or 500 training points, 9,261 grid positions, early 100 ms at 8 kHz. The boxplots show PI-DANF consistently below vanilla DANF and below the W-only wave-equation prior on both NMSE and Pearson correlation, and the improvement also shows up on the X/Y/Z channels. That last point is interesting because the naive prior never touches those channels directly; the shared network carries some of the constraint. The paper does not overclaim. It says simulated shoebox rooms and early RIRs, and the conclusion stays inside those bounds.\n\nSoft spots, in order.\n\n1. The load-bearing assumption is the mapping u = -v/(rho0 c0), cited to [31] but never checked against HARP2. HARP2 is an image-source simulator with its own directivity and normalization choices. If its X/Y/Z channels are not exactly particle velocity under SN3D with that sign and scale, then equations (11)-(12) penalize the correct ground-truth data. This is a real concern, not a fatal one. It is checkable in one experiment or by reading the HARP2 normalization code, and it should be done in revision.\n\n2. The stress-test note about near-field terms is worth taking seriously. HARP2's high-order Ambisonics may include distance-dependent terms that make the simple plane-wave relation (7) inexact at some positions. The paper does not discuss this. Again, checkable.\n\n3. There are no statistical tests and no multiple seeds. The boxplots are over 10 rooms, but each room gets one training run. I would not reject a paper for this alone, but the consistency claim rests on a visual pattern rather than a confidence statement.\n\n4. Simulation-only validation. The authors say realistic rooms are future work. That is an honest limitation, not a flaw, but it bounds the practical value.\n\nThe math, baseline choice, and citation pattern look solid. Self-citation of DANF is appropriate, and the central result is not circular: the priors are standard PDEs, not fitted to the target.\n\nWho this is for: people working on neural acoustic fields, spatial RIR interpolation, or PINNs in acoustics. A competent referee can verify the main claim in an afternoon. I would send this to peer review with the request that the authors check the FOA normalization in HARP2 and add seeds or significance testing.","headline":"Clean, narrow extension of DANF with correctly derived FOA momentum and continuity priors; the main open question is whether HARP2's X/Y/Z channels actually match the assumed particle-velocity convention.","tokens_in":10084,"tokens_out":2160,"would_cite":true,"duration_ms":25189,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding momentum and continuity loss terms to a neural acoustic field improves interpolation of first-order Ambisonic room impulse responses at unmeasured positions.","keywords":["physics-informed neural network","first-order Ambisonics","room impulse response","neural acoustic fields","sound field interpolation","particle velocity","momentum equation","continuity equation"],"falsifier":"Evaluate the two residuals in Eqs. (11) and (12) on the simulator's ground-truth FOA RIRs at the training positions: if the true recordings do not have near-zero residuals under the assumed sign and scaling, the physics priors are penalizing the data themselves and the reported gains would be artifacts of a mismatched convention.","tokens_in":9040,"feed_emoji":"🔊","tokens_out":10801,"duration_ms":163817,"temperature":0.7,"pith_summary":"This paper proposes PI-DANF, a physics-informed extension of the direction-aware neural acoustic field (DANF) for first-order Ambisonic room impulse responses (FOA RIRs). The authors aim to show that two extra training penalties, derived from the linearized momentum equation and the continuity equation, force the four FOA channels (W, X, Y, Z) to respect the physical coupling between sound pressure and particle velocity. In simulated shoebox rooms, the resulting model reconstructs the early 100 ms of FOA RIRs at unmeasured positions with lower normalized mean squared error and higher Pearson correlation than a data-only network and than a network with a wave-equation prior applied only to the W channel. This matters for immersive and binaural audio because FOA RIRs encode direction, so more accurate interpolation could make spatial audio rendering reliable at arbitrary listener positions.","feed_headline":"Two physics losses sharpen directional room-impulse prediction","feed_subtitle":"Momentum and continuity constraints make neural acoustic fields interpolate FOA room impulse responses more faithfully.","key_machinery":"The load-bearing machinery is the analytical correspondence between the FOA channels and the acoustic field: with the standard SN3D Ambisonic normalization, W equals pressure and (X, Y, Z) encode particle velocity up to the factor $-1/(\\rho_0 c_0)$. Feeding this correspondence into the linearized momentum equation and the continuity equation produces the two residuals in Eqs. (11) and (12), the momentum loss and the continuity loss, which tie all four channel outputs together through first-order partial derivatives. The residuals are evaluated at stochastically sampled positions and times by automatic differentiation and combined with the data-fidelity term using adaptively learned weights. The network is a multilayer perceptron with sinusoidal activations and an input-dependent gating mechanism, chosen so that the derivatives needed for the physics terms remain well behaved.","core_discovery":"The paper's central claim is that FOA RIRs should be treated as a coupled pressure-velocity field rather than four independent outputs. Under SN3D normalization, the W channel is the sound pressure and the (X, Y, Z) channels are related to particle velocity by $\\mathbf{u}(r,t) = -\\mathbf{v}(r,t)/(\\rho_0 c_0)$. Substituting this relation into the linearized momentum and continuity equations yields the two physics-informed residuals, $\\nabla \\hat{w} - c_0^{-1} \\partial_t \\hat{\\mathbf{v}} = 0$ and $\\nabla \\cdot \\hat{\\mathbf{v}} - c_0^{-1} \\partial_t \\hat{w} = 0$, which are added to the training objective as penalties. Across ten simulated shoebox rooms and two training-set sizes (250 and 500 measured positions), the method consistently attains the best NMSE and Pearson correlation on both the W channel and the averaged (X, Y, Z) channels, outperforming the vanilla DANF and the naive W-channel wave-equation prior.","pith_inferences":["Beyond the paper's own comparisons, the two priors may be acting partly as a spatial regularizer; an ablation with only one loss at a time would reveal whether momentum or continuity carries most of the gain.","The assumed sign and scaling relation between (X, Y, Z) channels and particle velocity is taken from a citation; an implicit implication is that any other FOA encoder or normalization would need a consistency check before these losses are applied.","Because the conservation laws hold in the late field as well, a testable extension is whether the same losses improve interpolation of full-length FOA RIRs rather than only the first 100 ms.","The pressure-velocity view suggests scaling the construction to higher-order Ambisonics by deriving analogous derivative constraints for each spherical-harmonic order, although the current losses are first-order only."],"forward_implications":["A physics-informed objective for FOA RIRs should couple all four channels through the momentum and continuity residuals instead of regularizing only the pressure-like W channel.","Because the wave equation follows from the momentum and continuity equations, PI-DANF is a natural extension of existing wave-equation PINNs and keeps their grid-less query capability.","With as few as 250 measured positions, the early part of FOA RIRs can be interpolated at unmeasured locations with lower error, and the gains appear in reflected energy that a W-only prior misses.","The reported NMSE and Pearson-correlation improvements are tied, by the metric reference the paper uses, to listening quality and perceptual localization, so the gains are likely audible in binaural rendering.","Each model is trained from scratch per room and can then be queried at arbitrary positions, so no dense measurement grid is needed at deployment."],"supporting_citations":[{"why":"Establishes the mapping from FOA (X, Y, Z) channels to particle velocity that the two priors are built on.","marker":"[31]"},{"why":"Provides the network architecture, training scheme, and W-channel wave-equation prior used as the naive baseline.","marker":"[28]"},{"why":"Introduces the direction-aware neural acoustic field (DANF) that PI-DANF extends, including the data-fidelity loss.","marker":"[39]"},{"why":"Supplies the simulated FOA room impulse responses used for training and evaluation.","marker":"[48]"},{"why":"Provides the sinusoidal activation function whose derivatives are used to compute the physics residuals.","marker":"[46]"},{"why":"Supplies the adaptive weighting that balances data fidelity and physics-prior losses during training.","marker":"[47]"},{"why":"Justifies Pearson correlation as a metric by linking it to listening quality and perceptual localization.","marker":"[50]"}],"fun_headline_variants":["Pressure-velocity coupling sharpens neural room impulse prediction","Two physics priors improve directional sound field interpolation","Momentum and continuity equations enhance acoustic field learning","Coupling pressure and velocity improves immersive audio prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole improvement rests on the assumption that the X, Y, and Z channels of the simulated FOA recordings are exactly proportional to the three components of particle velocity, with the W channel as pressure, including the sign and scale convention in Eq. (7); the paper takes this mapping from a cited reference rather than verifying it against the simulator's output.","fun_headline_variants_meta":{"raw":{"variants":["Pressure-velocity coupling sharpens neural room impulse prediction","Two physics priors improve directional sound field interpolation","Momentum and continuity equations enhance acoustic field learning","Coupling pressure and velocity improves immersive audio prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000768,"raw_usage":{"total_tokens":3411,"prompt_tokens":959,"completion_tokens":2452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2391}},"tokens_in":575,"tokens_out":2452,"duration_ms":18877,"temperature":1.0,"reasoning_tokens":2391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:53:30.687192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the two residuals in Eqs. (11) and (12) on the simulator's ground-truth FOA RIRs at the training positions: if the true recordings do not have near-zero residuals under the assumed sign and scaling, the physics priors are penalizing the data themselves and the reported gains would be artifacts of a mismatched convention.","supporting_citations":[{"cited_title":"Spatial impulse response rendering I: Analysis and synthesis,","cited_arxiv_id":null,"evidence_quote":"Establishes the mapping from FOA (X, Y, Z) channels to particle velocity that the two priors are built on."},{"cited_title":"Room impulse response reconstruction with physics-informed deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the network architecture, training scheme, and W-channel wave-equation prior used as the naive baseline."},{"cited_title":"Direction-aware neural acoustic fields for few-shot interpolation of ambisonic impulse responses,","cited_arxiv_id":null,"evidence_quote":"Introduces the direction-aware neural acoustic field (DANF) that PI-DANF extends, including the data-fidelity loss."},{"cited_title":"HARP: A Large-Scale Higher-Order Ambisonic Room Impulse Response Dataset","cited_arxiv_id":"2411.14207","evidence_quote":"Supplies the simulated FOA room impulse responses used for training and evaluation."},{"cited_title":"Implicit neural representations with periodic activation functions,","cited_arxiv_id":null,"evidence_quote":"Provides the sinusoidal activation function whose derivatives are used to compute the physics residuals."},{"cited_title":"Self-adaptive loss balanced physics-informed neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive weighting that balances data fidelity and physics-prior losses during training."},{"cited_title":"Towards an objective quality metric for interpolated directional room impulse responses,","cited_arxiv_id":null,"evidence_quote":"Justifies Pearson correlation as a metric by linking it to listening quality and perceptual localization."}],"review_version":1}