{"id":"1ed828e5-00cb-48ba-9057-917bb066cadd","arxiv_id":"2504.16289","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A single-author review paper classifies deep, data-driven room acoustics models, comparing them with traditional physics-based and data-driven models and outlining future research directions.","lead":"This preprint is a structured literature review of deep learning approaches to room acoustics modeling, and it proposes a conceptual framework that organizes those models by how much physics they encode. It matters to audio engineers and acoustics researchers because it maps a fast-moving field and points to open questions such as the gap between geometry-based and wave-based deep models.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed geometry/wave gap in Fig. 1 may be an artifact of the review's non-systematic selection, and the PIBI-Net bridge rests on one submitted paper plus an asymptotic analogy rather than demonstrated hybrid behavior.","rationale":"The reader identified the same broad weakness: the classification axes and selected literature may not be representative enough to support the claimed gap, and the bridge is extrapolated from a limited set including the author's own submitted PIBI-Net work. I agree, and I would add a more specific logical point: even the submitted PIBI-Net is described as 'boundary-based' rather than as a demonstrated hybrid of geometry-based and wave-based DL architectures. The asymptotic equivalence of BIE and geometric acoustics is an analytic statement about the continuous model, so it does not by itself show that a deep network trained with a BIE loss learns the geometric structure in a way that would fill the gap. This concern is real but does not change the reviewer's UNVERDICTED verdict: the paper is explicitly a literature review and research-perspective document, and the gap claim is framed as an 'apparent' gap rather than a theorem. The correct response is to flag the need for a systematic verification of the gap and, once available, an independent benchmark of PIBI-Net against geometry-conditioned and wave-based baselines. The duplicated passages and compilation artifacts in the submitted text further reduce confidence that the survey was finalized, but they are secondary to the selection-bias concern. No alternative verdict is justified because the paper makes no experimental or derivational claims that could be accepted or rejected on standard criteria.","tokens_in":18748,"tokens_out":5054,"duration_ms":53675,"concrete_test":"Run a systematic forward search (Scopus/DBLP, up to April 2025) on deep learning and room acoustics, classify every retrieved method for presence of a geometric prior (room-geometry input, ray/image-source features) and a wave-based prior (wave or Helmholtz equation in the loss or architecture), and count the joint cell. If the geometry+wave cell is nonempty, the gap in Fig. 1 is an artifact of the review's selection rather than a property of the literature.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central research-perspective claim is that there is an apparent gap between geometry-based and wave-based deep room-acoustics models, and that boundary information, as in PIBI-Nets, may bridge it. This claim would be undermined if the gap is an artifact of how the literature was selected and classified. The review reports no systematic search protocol, so the emptiness of the geometry+wave cell in Fig. 1 is asserted, not demonstrated. The only model placed in the bridging cell, PIBI-Net [199], is a submitted, not-yet-public paper by the author, and its description in the text is as a 'boundary-based' model, not as an architecture that explicitly combines geometric and wave-based priors. The conceptual bridge relies on the stationary-phase result [37] that a boundary integral equation asymptotically admits a geometric interpretation; that is a property of the continuous operator, not evidence that a deep network trained with a boundary-integral loss will combine both kinds of information. If existing models, e.g., PINNs with boundary-condition losses or neural fields conditioned on both room geometry and wave-equation residuals, already occupy the joint cell, the central gap claim loses its footing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a literature review of deep, data-driven room acoustics modeling, presented at the 11th Convention of the European Acoustics Association. It introduces a two-dimensional classification scheme along physical/data-driven and geometric/wave-based axes, applies this scheme to both traditional room acoustics models and deep learning (DL) models, and reviews two broad categories of DL models: purely data-driven models and models with geometric or wave-based physical priors. The paper closes with three research perspectives: the need for larger and better labeled datasets, the need to understand why nesting and nonlinearity in deep networks benefit room acoustics modeling, and the claim that there is an apparent gap between geometry-based and wave-based deep room acoustics models. The author suggests that including boundary information in both model structure and training, as exemplified by physics-informed boundary integral networks (PIBI-Nets), may be the key to combining geometric and wave-based information.","tokens_in":19074,"tokens_out":3420,"duration_ms":35305,"significance":"If the proposed classification and the claimed geometry/wave gap are accepted, the paper provides a useful conceptual framework for organizing a rapidly growing literature, and its research perspectives point to concrete directions for future work. The review is broad, with a substantial reference list, and it makes an explicit effort to connect traditional acoustics modeling concepts to the deep learning literature. Strengths include the structured overview in Figure 1, the clear separation of purely data-driven models from models with physical priors, and the honest use of hedged language such as 'apparent gap' and 'may lie.' However, the central research-perspective claim depends on a non-systematically selected literature and on an unpublished, author-authored model description, so the significance of the paper currently rests on a conjecture rather than on demonstrated evidence.","major_comments":[{"comment":"The claimed gap between geometry-based and wave-based deep learning models is asserted on the basis of a literature review whose selection is not described: the paper provides no search protocol, inclusion/exclusion criteria, or explicit category-assignment rules. Since the emptiness of the geometry+wave cell in Figure 1 is the central observation from which the key research perspective is derived, this absence is load-bearing. The authors should either provide a systematic methodology (e.g., search databases, search terms, screening criteria, and a table of the categorized papers) or present the gap explicitly as a tentative observation from a curated selection, with the corresponding limitation stated.","section":"Section 3, Figure 1"},{"comment":"The PIBI-Net model is cited as the 'first and promising result' in the direction of combining geometric and wave-based information, but reference [199] is a submitted, not-yet-published paper by the author, and the text provides only a single-sentence description with no architectural detail, experimental setup, or quantitative results. Because this model is the only occupant of the bridging cell in Figure 1, the reader cannot evaluate whether it actually combines geometric and wave-based behavior. The manuscript should describe the model and its reported results in enough detail to support the claim, cite a publicly available version, or explicitly label the bridge as a conjecture that awaits evidence.","section":"Section 3, reference [199]"},{"comment":"The classification of existing models as geometry-based or wave-based is too coarse to support the gap claim. For instance, INRAS is described as including boundary geometry and NACF as including material properties, and many PINN formulations are trained with boundary-condition losses; these could plausibly be viewed as already combining geometric and wave-based information. The paper does not justify why these models are excluded from the geometry+wave cell. Without clear criteria for what counts as 'geometry-based,' 'wave-based,' and 'boundary-based,' the asserted gap may be an artifact of the author's categorization rather than a property of the literature.","section":"Section 2.3"},{"comment":"The argument that the asymptotic geometric interpretation of the boundary integral equation [37] implies that boundary information in a DL model will combine geometric and wave-based priors is an analogy at the level of the continuous operator. It is not evidence that a deep network trained with a boundary-integral loss will exhibit such hybrid behavior, since the optimization dynamics and approximation properties of the network are not governed by the stationary-phase approximation. The paper should separate the mathematical property of the BIE from the empirical hypothesis about PIBI-Nets, and should note that the latter remains untested in publicly available literature.","section":"Section 3"}],"minor_comments":[{"comment":"The manuscript contains a formatting artifact consisting of a repeated paragraph and the stray header 'van Waterschoot Part B2 DIORAMA 3' in the middle of Section 2.2; this should be removed.","section":"Section 2.2"},{"comment":"The abstract states that 'the majority' of deep data-driven room acoustics models lack intrinsic space-time structure, but the paper does not quantify this. Please provide a rough count from the reviewed papers or soften the claim to 'many' or 'most reviewed models.'","section":"Abstract and Section 2"},{"comment":"The in-text citation numbering appears inconsistent with the reference list in places (for example, the duplicate passage cites geometry inference with different numbers, and 'Transformer-based models observed to perform below expectations' is cited to [84], which in the reference list is a different work). Please reconcile all citation numbers.","section":"References and in-text citations"},{"comment":"Figure 1 would be easier to interpret if the caption clearly stated that the placement of deep learning models in the cells reflects the author's own categorization, and if the meaning of the color coding (blue/green/red/yellow) were explained in the caption itself rather than only in the acknowledgments.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's main research perspective is closely tied to the author's own submitted work [199] and to the prospective ERC-DIORAMA project, which is acknowledged in Figure 1. This is not a disqualifier, but the editor should be aware that the key evidence for the bridge claim is an unreviewed self-citation. I would recommend asking the author to either provide a preprint or a detailed technical appendix for PIBI-Net, or to soften the claim accordingly. Additionally, the manuscript is a conference-style short review; whether it meets the journal's standards for systematicity is a judgment for the editor, given the absence of a search protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it's a genuinely useful literature review: the two-axis classification of deep room-acoustics models against the traditional physical/data-driven and geometric/wave-based categories is a sensible organizing scheme, and the survey is broad and current. Second, the paper's central research-perspective claim — that there's a gap between geometry-based and wave-based deep models and that boundary information might bridge it — is real but rests on the author's own submitted PIBI-Net paper, and on a manuscript that is visibly unfinished.\n\nWhat the paper does well: it gives a comprehensive map of the deep-learning room-acoustics landscape, from T60 estimation and echo cancellation to neural acoustic fields and physics-informed neural networks. The figure alone is worth the price of admission for anyone entering the field. The discussion of why nesting and nonlinearity help model a supposedly linear process is a fair open question, and the call for more high-quality datasets is on target.\n\nWhere it's soft: First, the manuscript quality. The text around Figure 1 contains duplicated reformatted passages, which suggests the arXiv upload is a draft rather than a final version. That's a small fix, but it matters for review. Second, the claimed geometry/wave gap is not demonstrated by a systematic search. The review doesn't report a selection protocol, and it's easy to think of existing models (PINNs with boundary-condition losses, neural fields conditioned on geometry and wave equation residuals) that might already sit in the joint cell. The author should either tighten the definition of the gap or show that those candidates fail. Third, the PIBI-Net bridge is a promising hypothesis, not a result: the stationary-phase analogy is about the continuous boundary integral operator, not about whether a boundary-informed network will combine geometric and wave information. The reliance on the author's own submitted work for the central claim is worth flagging, though self-citation itself is not a problem if the work is later shown to deliver.\n\nWho should read it: people who want a structured entry into deep room-acoustics modeling, and anyone planning research in sound field reconstruction or physics-informed audio. It's not a new method or measurement paper.\n\nRecommendation: send it to peer review. It deserves a serious referee, with a request for major revision: clean up the text, add a systematic literature search description, and either substantiate or soften the gap claim. The review frame itself is solid.","headline":"A useful two-axis review of deep room-acoustics modeling whose 'gap' between geometry-based and wave-based models is an interesting hypothesis that needs sturdier support than the author's own submitted paper.","tokens_in":19440,"tokens_out":2613,"would_cite":true,"duration_ms":24801,"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":"This review argues that deep, data-driven room acoustics models fall on the same physics-versus-data and geometric-versus-wave axes as classical models, and that boundary-aware networks may close the gap between the two deep-learning camps.","keywords":["room acoustics","deep learning","data-driven modeling","literature review","physics-informed neural networks","sound field reconstruction","boundary integral equation"],"falsifier":"Train a boundary-informed network and an otherwise identical network with all boundary inputs removed on the same corpus of rooms; if the boundary-free network matches it on sound field reconstruction accuracy across frequencies, source positions, and unseen geometries, the claim that boundary information is the key bridge between geometric and wave-based deep models would be falsified.","tokens_in":18531,"feed_emoji":"🔊","tokens_out":8182,"duration_ms":76171,"temperature":0.7,"pith_summary":"This paper is a literature review that tries to establish a way of seeing the deep-learning turn in room acoustics: the same two distinctions that organize traditional models, physical versus data-driven and geometric versus wave-based, also organize the new deep models. It argues that most deep models imported from speech and image processing ignore the space-time structure of wave propagation, and that the deep models which do respect physics have so far shown their clearest value in sound field reconstruction. The review then identifies a gap between geometry-based and wave-based deep models and proposes that boundary information, built into both the network structure and the training loss, may be the bridge, citing physics-informed boundary integral networks as the first evidence. A sympathetic reader would care because this reframes a scattered literature into a map with an actionable research direction: put the room's boundaries into the network.","feed_headline":"Boundary-aware nets may bridge the two camps in room acoustics","feed_subtitle":"A review maps deep room-acoustics models onto physics axes and points to boundary information as the key to combining them.","key_machinery":"The machinery is a two-axis classification frame plus a theoretical bridge. The frame sorts room acoustics models, before and after deep learning, by whether they are physical or data-driven and by whether they assume geometric or wave-based sound behavior; the paper uses it to show that deep models with physical priors split into geometry-based and wave-based families, with an empty region between them. The bridge is the boundary integral equation, a wave-based formulation of the interior sound field on the room's boundary whose solution asymptotically reduces to geometric acoustics, so it inherently connects the two regimes. The proposed mechanism for crossing the gap is the PIBI-Net, a physics-informed boundary integral network that folds boundary geometry and material properties into both the model structure and the training strategy, so the network learns with the room's surface rather than treating it as an external condition. The framework's work is to convert a scattered literature into a structured map, and the boundary-integral/PIBI-Net pairing is what makes the map's central research direction concrete.","core_discovery":"The paper's central claim is that deep, data-driven room acoustics models inherit the same conceptual structure as classical physical and data-driven models, and should be classified along the same two axes: physical versus data-driven and geometric versus wave-based. It reviews the field to show that most deep models, borrowed from speech and image processing, lack the space-time structure of acoustic wave propagation, while recent models that include either geometric or wave-based priors have produced their clearest successes in sound field reconstruction. Surveying these, the paper finds an apparent gap: geometry-based deep models and wave-based deep models have developed largely separately, with no deep counterpart to the classical boundary-integral models that already sit between the two regimes. Because the wave-based boundary integral equation asymptotically reduces to geometric acoustics, the review argues that the key to combining geometric and wave-based deep models may be to include boundary information in both the model structure and the training strategy, and it points to physics-informed boundary integral networks as a first promising step in that direction.","pith_inferences":["Editorial inference: the boundary-bridge recipe likely transfers to other wave-physics inverse problems, where ray-based and full-wave models could be reconciled by networks trained with boundary integral losses.","Editorial inference: the apparent gap may be partly a chronological artifact of a young field; as boundary-informed models mature, they could absorb both geometry-based and wave-based approaches rather than remain a third category.","Editorial inference: a controlled benchmark with identical rooms and data, comparing PIBI-Nets against pure geometry-conditioned and pure wave-equation-regularized networks, would quantify when boundary information is what actually improves reconstruction."],"forward_implications":["Deep models that ignore wave-propagation structure will keep underperforming on tasks that require room impulse responses or full sound fields, where time delays and space-time relations are the essence.","A unified class of boundary-informed networks could combine the efficiency of geometric models with the accuracy of wave-based models for sound field reconstruction.","Sound field reconstruction, not parameter estimation or enhancement, is where physics-informed deep models are currently proving themselves.","Progress hinges on datasets that reconcile realistic audio scenes, with moving and directional sources and microphones, with accurate labeling of source and receiver positions.","Understanding why nested nonlinear networks work for a process traditionally modeled as linear and time-invariant remains an open prerequisite for deliberate architecture design."],"supporting_citations":[{"why":"Supplies the stationary-phase result that the wave-based boundary integral equation asymptotically reduces to geometric acoustics, which underpins the argument that boundary information can bridge the two model families.","marker":"[37]"},{"why":"The submitted physics-informed boundary integral network (PIBI-Net) for sound field reconstruction is cited as the first promising result combining boundary information in both model structure and training.","marker":"[199]"},{"why":"The Neural Acoustic Field is the representative geometry-based deep model for continuous source-receiver-to-RIR mapping, anchoring one side of the apparent gap.","marker":"[160]"},{"why":"Physics-informed deep learning for room impulse response reconstruction anchors the wave-based side by showing how wave-equation regularization is applied in deep sound field models.","marker":"[176]"},{"why":"The plane-wave-decomposition GAN represents wave-based priors used inside a generative deep model for sound field reconstruction, completing the wave-based family the review contrasts with geometry-based models.","marker":"[182]"}],"fun_headline_variants":["Boundary info may fuse geometric and wave deep models","Review: Deep acoustics models need boundary awareness","Acoustics review eyes boundary integral deep nets","Key to deep room acoustics: boundary conditions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's load-bearing premise is that the studies it surveys and the two-axis scheme it draws are representative enough to make the gap between ray-based and wave-based deep models real, and that the single boundary-integral-network result generalizes beyond its own experiments.","fun_headline_variants_meta":{"raw":{"variants":["Boundary info may fuse geometric and wave deep models","Review: Deep acoustics models need boundary awareness","Acoustics review eyes boundary integral deep nets","Key to deep room acoustics: boundary conditions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1432,"prompt_tokens":952,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":568,"tokens_out":480,"duration_ms":5139,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:05:28.916555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a boundary-informed network and an otherwise identical network with all boundary inputs removed on the same corpus of rooms; if the boundary-free network matches it on sound field reconstruction accuracy across frequencies, source positions, and unseen geometries, the claim that boundary information is the key bridge between geometric and wave-based deep models would be falsified.","supporting_citations":[{"cited_title":"HOMULA-RIR: A room impulse response dataset for tele- conferencing and spatial audio applications acquired through higher-order mi- crophones and uniform linear microphone arrays,","cited_arxiv_id":null,"evidence_quote":"The submitted physics-informed boundary integral network (PIBI-Net) for sound field reconstruction is cited as the first promising result combining boundary information in both model structure and training."},{"cited_title":"Deep neural room acoustics primitive,","cited_arxiv_id":null,"evidence_quote":"The Neural Acoustic Field is the representative geometry-based deep model for continuous source-receiver-to-RIR mapping, anchoring one side of the apparent gap."},{"cited_title":"Toward learning robust con- trastive embeddings for binaural sound source localization,","cited_arxiv_id":null,"evidence_quote":"Physics-informed deep learning for room impulse response reconstruction anchors the wave-based side by showing how wave-equation regularization is applied in deep sound field models."},{"cited_title":"Spatial extrapolation of early room impulse responses with noise-robust physics-informed neural network,","cited_arxiv_id":null,"evidence_quote":"The plane-wave-decomposition GAN represents wave-based priors used inside a generative deep model for sound field reconstruction, completing the wave-based family the review contrasts with geometry-based models."}],"review_version":1}