{"id":"d609c8a5-c194-4fee-bf93-384d727e26cd","arxiv_id":"2505.03354","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of AI and physics-informed neural network methods for electromagnetic and nanophotonic design, with no original results.","lead":"This preprint surveys recent work applying deep neural networks and physics-informed neural networks to electromagnetic and nanophotonic design. It is a narrative review, not a new research result.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Load-bearing concern: the review's central comparative claim that PINNs outperform DNNs rests on unsourced quantitative entries in Table 1 and a single citation (ref 11) that does not support the stated comparison.","rationale":"This is a narrative review, not a research preprint; its central claim is a comparative synthesis rather than a novel result. The most load-bearing assumption is that the synthesis is faithful to the cited literature and that comparative claims are supported by those references. The reader identified a citation mismatch in Section 2.2.1, and I agree that is a real symptom. However, I find a more consequential problem: the paper's key comparative conclusion that PINNs outperform DNNs is not backed by a controlled benchmark. Section 2.3.2 attributes this superiority to ref 11, which is an inverse-problem PINN paper without a head-to-head DNN comparison. Section 4's Table 1 provides quantitative parameter ranges (100k-1M for PINNs, 1M-20M for DNNs, etc.) that are not sourced anywhere; these numbers appear to be fabricated or at least unverifiable from the cited works. Since the comparative evaluation in Section 4 and the conclusion in Section 5 rely on these numbers, the central claim's evidentiary basis is shaky. That said, the paper remains a survey with no novel claim to accept or reject. The proper verdict is still UNVERDICTED: it should not be treated as a definitive reference, but it does not fail a specific test of correctness. I therefore recommend UNCHANGED, while flagging that the unsourced table and unsupported comparison are significant reliability concerns for any reader who cites this review.","tokens_in":30730,"tokens_out":3140,"duration_ms":29046,"concrete_test":"Compile Table 1's parameter counts from the actual cited papers (refs 11-15,169 for PINN; 1,3,4,7,70,72,75,152,156-159 for DNN; etc.). For each row, check whether the cited paper states the number of parameters used. If, as the text suggests, these ranges are not reported in any cited source, then recompute the comparative evaluation without them; if the conclusion \"PINNs are superior ... with small size of dataset\" (Section 5) still holds without Table 1's numbers, the concern is moot; otherwise the central claim weakens. Additionally, search ref 11 for any head-to-head comparison between PINN and a purely data-driven DNN on the same task; absence of such a benchmark would confirm the \"outperformed DNNs\" sentence is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that PINNs provide a more reliable, data-efficient alternative to pure DNNs—is a comparative judgment, not just a summary. The strongest support for this judgment appears in Section 2.3.2's closing paragraph: \"PINNs have demonstrated robustness in handling noisy data and complex parameter retrieval tasks, which outperformed DNNs in various studies\" citing only ref 11 (Chen et al. 2020). That paper solves inverse scattering problems with PINNs; it does not contain a controlled benchmark against data-driven DNNs, so the claimed superiority is not established by that citation. Similarly, Section 4's Table 1 lists model complexity ranges (e.g., PINN: 100k–1M parameters; DNN: 1M–20M; Transformer: 10M–100M) without any source. These numbers are load-bearing because the table is the paper's comparative evaluation, and the conclusion \"PINNs are superior ... with small size of dataset\" (Section 5) leans on them. In addition, Section 2.2.1 attributes the SOI power splitter work to \"Mohammed et. al.\" while citing ref 71 (Tahersima et al.), and Section 2.2.2 correctly attributes the same work to Tahersima et al. with the same reference—an internal inconsistency that signals the survey's summaries are not reliably checked. Since all conclusions are drawn from cited literature, a single demonstrably wrong attribution plus unsourced quantitative comparisons undermine the central claim's evidentiary basis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of artificial intelligence, deep neural network (DNN), and physics-informed neural network (PINN) methods for electromagnetic and nanophotonic design. It reviews DNN and PINN fundamentals, forward and inverse design in nanophotonics, electromagnetic scattering and antenna applications, nonlinear optical modeling, and a comparative architecture overview in Table 1, concluding that PINNs offer a more reliable and data-efficient alternative to pure data-driven DNNs.","tokens_in":31033,"tokens_out":2872,"duration_ms":28052,"significance":"If its claims were properly supported, this survey would be a useful entry point for researchers seeking an overview of AI-based photonic and electromagnetic design, covering a broad range of architectures and applications. The paper's strengths are its wide thematic scope, recent references, and clear organization. However, because this is a review with no new mathematical results, its value rests entirely on the accuracy of its citations and summaries; the attribution errors and unsourced quantitative comparisons identified below currently undermine that basis.","major_comments":[{"comment":"The quantitative model-complexity ranges listed in Table 1 (e.g., PINN: 100k-1M parameters; DNN: 1M-20M; Transformer: 10M-100M) are presented without any source or derivation. These numbers are load-bearing because the table is the paper's central comparative evaluation and the conclusion in Section 5 that 'PINNs are superior ... with small size of dataset' rests on them. The authors should either cite specific sources for each range, explain how they were estimated, or remove the quantitative entries and replace them with qualitative statements.","section":"Section 4, Table 1"},{"comment":"The claim that 'PINNs have demonstrated robustness in handling noisy data and complex parameter retrieval tasks, which outperformed DNNs in various studies' is supported only by reference 11 (Chen et al., Optics Express 2020). That paper solves inverse scattering problems with PINNs; it does not report a controlled benchmark against data-driven DNNs. The claimed superiority is therefore not established by this citation. The authors should either cite actual comparative studies or qualify the claim as an expectation rather than an established result.","section":"Section 2.3.2, closing paragraph"},{"comment":"The SOI power splitter work is attributed to 'Mohammed et. al.' in Section 2.2.1 with citation 71, but Section 2.2.2 correctly attributes the same work to Tahersima et al. with the same reference 71. This internal inconsistency indicates that the survey's summaries have not been checked against the cited sources. A related mismatch occurs earlier in Section 2.2.1, where Peurifoy et al. are cited with references 4 and 70, but reference 70 is So et al. (ACS Applied Materials & Interfaces 2019), not Peurifoy et al. These errors erode confidence in the accuracy of the survey's account of the literature.","section":"Section 2.2.1 and Section 2.2.2"}],"minor_comments":[{"comment":"The chain-rule expression in Eq. (2) omits the summation over neurons in each layer; the standard backpropagation formula includes a sum over the units receiving the weight. This is a minor notation issue but should be corrected for pedagogical accuracy.","section":"Section 1.3.2, Eq. (2)"},{"comment":"The two forms of the generalized Snell's law use inconsistent subscripts ('ni' in Eq. (5) and 'nI' in Eq. (6)), and the phase-gradient term dφ/dx is not defined. Please unify the notation and define all symbols.","section":"Section 2.1.1, Eqs. (5)-(6)"},{"comment":"The four panels in Figure 3 describe specific PINN variants but do not include citation numbers in the caption text, making it difficult for readers to locate the corresponding sources. Please add reference pointers.","section":"Figure 3 captions"},{"comment":"The word 'Appartently' appears at the beginning of the closing paragraph; this should be 'Apparently'.","section":"Section 2.3.2, text"},{"comment":"The column headed 'PINN Model Complexity (Params)' is confusing because the table lists many non-PINN architectures; presumably this column indicates whether the architecture is a PINN (Yes/No) and then lists typical parameter counts. Please rename the column or split it into two columns.","section":"Section 4, Table 1 header"},{"comment":"The sentence 'Bayesian Optimization (BO) and Support Vector Machines (SVMs) provide d lightweight alternatives' contains a typo ('provide d' -> 'provided' or 'offer').","section":"Section 4, text"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains a large block of self-citations (e.g., refs 37, 40-59, and 69) and several demonstrable citation mismatches. For a review article, whose reliability depends on faithful summary of the literature, the editorial bar should include a full citation audit. The topic fits the journal's scope, but the current evidentiary basis for the central comparative conclusion is too fragile to accept as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a narrative review of DNNs and PINNs in nanophotonics and electromagnetics. It has no new results, which is fine for a review, but its central comparative claim—that PINNs are more reliable and data-efficient than DNNs—is not backed by the evidence it cites. The stress-test note is on target.\n\nWhat it does well: the survey covers a wide range of applications: forward and inverse design, scattering, antenna design, beamforming, nonlinear optics, and quantum optics. The introductory sections on DNN and PINN fundamentals explain loss functions, automatic differentiation, and several architectures clearly enough for a reader new to the field. Organizing the literature by application area makes it easy to get a bird's-eye view.\n\nSoft spots: first, the claim in Section 2.3.2 that PINNs \"outperformed DNNs in various studies\" cites only ref 11, a PINN inverse-scattering paper that does not contain a controlled benchmark against data-driven DNNs. Second, Table 1 lists parameter ranges (e.g., PINN: 100k–1M; DNN: 1M–20M; Transformer: 10M–100M) with no sources, and the conclusion leans on these numbers. Third, the same SOI power splitter work is attributed to \"Mohammed et al.\" in Section 2.2.1 and to Tahersima et al. in Section 2.2.2, both citing ref 71. That is an internal contradiction that makes you doubt the other summaries. Fourth, the reference list is stuffed with the first author's own papers (refs 37, 40–59, 69), many in a block; even if each is relevant, the density looks like padding and a reviewer should check.\n\nThe central argument that PINNs embed physical laws and can be data-efficient is already established in the literature, so the review doesn't need to prove it. But the specific comparative evaluation in Section 4 and the conclusion are built on unsourced or misattributed evidence. That weakens the review's value as a synthesis.\n\nWho this is for: a newcomer wanting a quick map of AI techniques in photonics and EM, but they should follow the actual references rather than trust the text. It is not a definitive reference.\n\nRecommendation: I would send this to peer review with a clear request for major revision: fix the misattribution, add sources to Table 1 or remove it, hedge the comparative claims, and trim the self-citations. If the authors make those changes, the review becomes a competent entry point. If not, better to let it go.","headline":"A broad but sloppy review: useful as a literature map for newcomers, but the central comparative claim about PINNs outperforming DNNs is unsupported and the citation errors need fixing before this can be trusted.","tokens_in":31555,"tokens_out":2230,"would_cite":false,"duration_ms":22861,"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":"Physics-informed nets beat data-only nets for photonic design","keywords":["physics-informed neural networks","nanophotonics design","inverse design","electromagnetics","deep learning","metasurfaces","electromagnetic field equations","forward design"],"falsifier":"A controlled benchmark would settle the central claim: take a fixed set of nanophotonic inverse-design tasks—for example, power splitters and a metasurface beam deflector—and train a PINN and a data-only DNN with matched compute budgets and identical training data, measuring accuracy as the dataset shrinks. If even one well-posed task shows the data-only network matching the PINN's accuracy with no more data, the review's comparative claim would need substantial qualification.","tokens_in":30544,"feed_emoji":"⚡","tokens_out":6922,"duration_ms":66206,"temperature":0.7,"pith_summary":"This review surveys recent work applying deep neural networks and physics-informed neural networks (PINNs) to electromagnetic and nanophotonic design. Its central claim is that embedding governing physical laws—the electromagnetic field equations, energy conservation, or nonlinear wave equations—directly into the network's training objective makes PINNs more reliable and data-efficient than purely data-driven DNNs for both forward prediction and inverse design. The paper supports this claim by cataloging applications in metasurface design, light scattering, antennas, nonlinear optics, and electromagnetic simulation, and by comparing architectures on speed, dataset needs, physical consistency, and accuracy. The stakes are practical: if the claim holds, the default design workflow shifts from generating large simulation datasets to training physics-constrained models that respect conservation laws and boundary conditions with far fewer samples.","feed_headline":"Physics-informed nets beat data-only nets for photonic design","feed_subtitle":"Embedding the governing field equations into neural networks can cut data needs and speed up device design, a survey argues.","key_machinery":"The load-bearing mechanism is the physics-augmented loss function, $\\mathcal{L}=\\mathcal{L}_{\\mathrm{data}}+\\lambda\\mathcal{L}_{\\mathrm{PDE}}$, in which the PDE residual—for electromagnetics, the residual of the governing field equations evaluated at collocation points—penalizes predictions that violate the embedded physics. Automatic differentiation supplies the spatial and temporal derivatives needed to compute that residual without numerical meshes. This single mechanism carries the argument: it is what makes PINNs data-efficient, physically consistent, and applicable to inverse problems, because the same loss can be minimized over unknown material parameters or geometries as well as over network weights.","core_discovery":"The central claim is that physics-informed neural networks unify data-driven learning with physical laws and thereby overcome the main weaknesses of pure data-driven DNNs: large data requirements, unphysical predictions, and poor handling of ill-posed inverse problems. The survey presents evidence that PINNs retrieve material parameters from scattered or near-field data without iterative solvers, predict field distributions in milliseconds, accelerate design cycles by orders of magnitude, and model nonlinear and time-dependent effects such as soliton dynamics and the optical Kerr effect. It also reports that PINNs typically need far fewer parameters than DNNs or CNNs—roughly 100k–1M versus 1M–20M—while enforcing physical consistency. The paper's contribution is synthetic rather than a new algorithm: it assembles the surveyed results into a comparative argument that physics-constrained learning, not raw data size, is the more robust route to photonic and electromagnetic design.","pith_inferences":["If the surveyed pattern is correct, the practical bottleneck shifts from data collection to choosing the right physics residual and balancing it against data; a natural testable extension is a standardized benchmark that reports accuracy versus training-set size for matched compute budgets across PINN, DNN, and adjoint or full-wave baselines.","The physics-loss advantage should be largest when inverse solutions are non-unique, because the PDE residual selects physically admissible branches, and smallest in well-posed forward regression with abundant data, where pure data-driven networks may match or surpass PINNs—the review does not test this boundary.","The reported parameter-count contrast (PINNs at 100k–1M versus DNNs at 1M–20M) is suggestive but uncontrolled for accuracy and dataset size; a controlled scaling study could turn it into a quantitative design rule.","Because many surveyed PINN successes are demonstrated on single devices or narrow geometry classes, the strongest next test would be generalization across a broad, unseen geometry distribution—the regime where the review claims PINNs should most clearly outpace data-only networks."],"forward_implications":["Forward design—predicting spectra or field distributions from geometry—can be done from smaller datasets when physics residuals are part of the loss, because the governing equations act as a regularizer.","Inverse design can retrieve geometries, permittivity, or permeability from scattered or near-field data without iterative full-wave solvers, addressing problems that are ill-posed or non-unique.","Hybrid physics-informed frameworks, such as physics-augmented CNNs and physics-informed reinforcement learning, can cut design-cycle times by large factors while keeping fabrication constraints such as minimum feature size in the loop.","The same physics-constrained approach extends to nonlinear optics: self-focusing, soliton propagation, and power-dependent scattering can be modeled with less data than split-step or pure data-driven methods.","The review's own comparison identifies remaining bottlenecks—stiff nonlinear loss landscapes, spectral bias at high frequencies, and the difficulty of balancing data and physics loss terms—so adaptive loss weighting and hybrid architectures are the stated next steps."],"supporting_citations":[{"why":"Supplies the core evidence that embedding the governing field equations in the loss enables inverse parameter retrieval in nano-optics and metamaterials.","marker":"11"},{"why":"Supplies the foundational PINN formulation of data loss plus PDE residual on which the survey's mechanism rests.","marker":"105"},{"why":"Supplies the DNN baseline for integrated photonic power splitters that PINNs are claimed to improve upon in data efficiency.","marker":"71"},{"why":"Supplies the physics-augmented CNN result claiming orders-of-magnitude design-cycle speedup for diffractive devices.","marker":"96"},{"why":"Supplies the physics-informed reinforcement learning result claiming sample-efficiency gains and fabrication-feasibility constraints.","marker":"97"},{"why":"Supplies the multi-receptive-field PINN result claiming millisecond field reconstruction in complex electromagnetic media.","marker":"94"},{"why":"Supplies the PINN result for time dynamics in optical resonances used for real-time wave manipulation.","marker":"12"},{"why":"Supplies the PINN result for three-dimensional light diffraction from metasurfaces used for holographic design.","marker":"15"}],"fun_headline_variants":["Physics-informed nets cut data and speed photonic design","Survey: PINNs beat pure data nets for photonic devices","Embedding physical laws in AI accelerates device design","Physics-constrained learning: less data, faster photonics","AI that knows physics designs photonics without trial-and-error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review assumes that its summaries of the cited papers are faithful and correctly attributed, because every comparative conclusion is built on those summaries; if a key study is misattributed or its reported numbers are wrong, the comparison loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed nets cut data and speed photonic design","Survey: PINNs beat pure data nets for photonic devices","Embedding physical laws in AI accelerates device design","Physics-constrained learning: less data, faster photonics","AI that knows physics designs photonics without trial-and-error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2875,"prompt_tokens":961,"completion_tokens":1914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":1835}},"tokens_in":577,"tokens_out":1914,"duration_ms":13629,"temperature":1.0,"reasoning_tokens":1835,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:52:34.987013+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled benchmark would settle the central claim: take a fixed set of nanophotonic inverse-design tasks—for example, power splitters and a metasurface beam deflector—and train a PINN and a data-only DNN with matched compute budgets and identical training data, measuring accuracy as the dataset shrinks. If even one well-posed task shows the data-only network matching the PINN's accuracy with no more data, the review's comparative claim would need substantial qualification.","supporting_citations":[{"cited_title":"Physics -informed multi -LSTM networks for metamodeling of nonlinear structures","cited_arxiv_id":null,"evidence_quote":"Supplies the PINN result for time dynamics in optical resonances used for real-time wave manipulation."}],"review_version":1}