{"id":"d1e898ee-5a7f-40d4-877e-20640fe627e1","arxiv_id":"1908.03702","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An iterative neural-network-based optimizer discovers a silicon L3 photonic crystal nanocavity design with a theoretical quality factor above 11 million, using 8,070 simulations.","lead":"The authors trained a deep neural network in an iterative loop to propose, test, and add new photonic crystal cavity designs, finding a silicon L3 cavity with a theoretical Q factor above 11 million. The result more than doubles previous record Q values found by other automated methods, using a similar total number of simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The record Q=1.10e7 claim lacks FDTD convergence evidence; without a mesh/PML convergence test, the factor-of-two improvement over prior records is not yet established.","rationale":"The reader's verdict identified the same weakest assumption: the published Q values are treated as exact first-principles numbers, while no convergence or error estimate is provided. My independent reading of the paper finds this to be the single most load-bearing issue because the abstract's central quantitative claim ('exceeding 11 million', 'more than twice of the previously reported record values') is a comparison of FDTD Q values across different simulation settings. If the FDTD discretization at the reported settings overestimates Q, the 1.10e7 value could be an artifact rather than a real radiative-quality-factor improvement. I also considered the lack of a random-search or baseline comparison, since the phrase 'efficiently found' implies a comparison against generic exploration; however, the paper does include comparisons with a genetic algorithm and a leaky-mode visualization method, and the efficiency claim is secondary to the numerical record claim. The parameter count is different (25 vs. 5 and 9), but this is a possible source of unfairness, not a fatal flaw, because more parameters could also hurt optimization. The absence of code/data further limits reproducibility, but the reader already noted this. My proposed test, a convergence study plus a same-solver recomputation of the reference structures, would directly settle whether the record claim is numerically robust. Since the reader's conditional verdict already requires such validation, my analysis does not move the verdict; it supports keeping CONDITIONAL rather than full acceptance.","tokens_in":12465,"tokens_out":2185,"duration_ms":27972,"concrete_test":"Recompute the Q of the strategy-(A+C) best structure (Fig. 5(c)) with the same FDTD solver at two mesh resolutions differing by a factor of 2 and with at least double the PML distance; also compute the Q of the reference designs from [16] and [17] using the identical solver, mesh, and boundary settings. If the Q of the best structure remains above 1.0e7 and is still more than twice the reference Q values under identical refined settings, the record claim is supported; if Q shifts by more than about 20% or the gap to the references narrows materially, the central comparison needs to be re-qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central headline claim is the numerical value Q = 1.10e7 found by strategy (A+C) in Section 3.4, and the comparison claiming this is more than twice the previous records from [16] and [17]. Both assertions depend on treating 3D-FDTD Q values as exact across different structures and different simulation settings. The paper reports no mesh-convergence study, no domain/PML sensitivity test, and no error estimate for the FDTD Q values in Section 3.1 or Section 3.3 ('calculation conditions are the same as in [19]'). For cavities with Q above 10^7, FDTD results can be sensitive to spatial discretization and boundary placement; numerical leakage or insufficient grid resolution can change Q by tens of percent or even produce spuriously high values. The comparison with references [16] and [17] also assumes that the authors' FDTD settings, hole radii, slab thickness (0.5366a vs. 0.55a in [16]), and mesh resolution give Q values directly comparable to those references, which is not demonstrated. Because the record claim is explicitly quantitative and factor-of-two, this is the weakest load-bearing assumption in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative optimization method for photonic-crystal nanocavity designs. A deep neural network is trained on a dataset of random hole-displacement structures and their FDTD-computed Q factors; candidate structures are then generated by gradient search on the network's predicted Q, subject to artificial loss terms that constrain exploration. The candidates are validated by 3D-FDTD and appended to the training set for the next iteration. The method is demonstrated on a silicon L3 cavity with 25 free parameters (after enforcing mirror symmetries), using 101 iterations and 8070 total FDTD evaluations. The reported best design has Q = 1.10 × 10^7, obtained with their 'A+C' exploration strategy, which the authors claim is more than twice the previous record L3-cavity Q values from a genetic algorithm [16] and from leaky-mode visualization [17]. The paper also compares the approach with Bayesian optimization and argues that the gradient-based search over a NN surrogate is better suited to high-dimensional parameter spaces.","tokens_in":12745,"tokens_out":3401,"duration_ms":37563,"significance":"If the numerical Q values are reliable, this is a valuable contribution: the method is not circular in the sense that the final Q is computed by FDTD rather than predicted by the network, and the iterative accumulation of validated samples is a sound and easily reproduced idea. The reported comparison across three exploration strategies and the discussion of Bayesian optimization are useful. The central quantitative claim, however, rests on two assumptions that are not verified in the manuscript: that the 3D-FDTD results are converged and accurate at the level of a factor of two, and that the earlier literature Q values [16,17] are directly comparable to the authors' own simulator settings. Because these assumptions are load-bearing for the headline record claim, the paper needs additional numerical evidence before the claim can be accepted.","major_comments":[{"comment":"The central result, Q = 1.10 × 10^7, is presented without any convergence or error analysis for the 3D-FDTD calculation. For Q values above 10^7, small numerical leakage due to finite grid resolution, PML placement, or simulation domain size can change the computed Q by a large fraction, and a factor-of-two claim is not robust to such effects. The authors should report a mesh-convergence study (e.g., Q vs. grid size for the final structure), a test of PML distance and boundary conditions, and an estimate of the numerical uncertainty of the reported Q values. Without this, the value Q = 1.10 × 10^7, and its comparison with previous records, is not quantitatively established.","section":"§3.3, §3.4, Fig. 3"},{"comment":"The comparison with the previous record values from [16] and [17] assumes that the authors' FDTD settings produce Q values directly comparable to those in the cited references. The text itself notes that the slab thickness differs (0.5366a here versus 0.55a in [16]), and the hole radius and other simulation conditions are said to be 'the same as in [19]' without giving details. If the reference values were computed with different mesh sizes, domain sizes, or PML parameters, the claimed factors of 2.6 and 2.1 could be artifacts of numerical settings. The authors should either recompute the reference structures with their own FDTD settings or otherwise demonstrate that the cited Q values are directly comparable.","section":"§4.2, §3.1"},{"comment":"The final optimized structure is shown only as a displacement-vector plot; the numerical coordinates of the air holes in the record Q = 1.10 × 10^7 design are not provided. Since the paper's purpose is to demonstrate a useful design tool and a quantitative record, the actual optimized geometry should be made available, either as a table of the 25 displacement parameters or as a data file. This is needed for independent verification and for other researchers to use the result.","section":"§3.4, Fig. 5"}],"minor_comments":[{"comment":"The manuscript contains two sections numbered '3.2', one titled 'Learning phase' and the other 'Structure search phase'; the second should be renumbered.","section":"Headings"},{"comment":"Equation (1) shows 'log 0' where it should read 'log10' in the teacher-data term; this appears to be a typesetting error that should be corrected.","section":"Eq. (1)"},{"comment":"Equation (5) contains a stray '≤ .' and the displayed formula could be misinterpreted as an inequality rather than a penalty term; please restate it as a regular sum of inverse distances.","section":"Eq. (5)"},{"comment":"The claim that the 10 neural networks 'learn the dataset in different orders' is not specified further; a sentence describing how the ordering is randomized would improve reproducibility.","section":"§3.2"},{"comment":"The phrase 'computation costs for first principles calculations increased only by 8 times (from 1070 to 8070 sample cavities)' is unclear because the first round uses 1000 initial plus 70 candidates, and later rounds add 70 per round; the count should be made explicit.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely sound in its methodology, and the central result is not circular because FDTD validates the candidates. However, the record Q value and the comparison with previous records are the main selling points, and the absence of FDTD convergence evidence is a genuine concern for a quantitative factor-of-two claim. I would ask the editor to require the authors to address this before publication, including either a convergence study or a softened claim. The final geometry should also be made available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the Q = 1.10e7 result is a 3D-FDTD value, not a neural-network prediction, so the loop is not circular: the NN proposes, FDTD disposes. Second, the paper lacks the numerical convergence checks that would make a factor-of-two record claim airtight. That is the soft spot, and it is fixable.\n\nWhat is new: the iterative scheme that takes the authors' earlier single-shot deep-learning optimization and closes the loop, adding FDTD-validated candidates to the training set. The three candidate-generation strategies—especially the repulsive artificial loss (C)—are a sensible way to balance exploitation and exploration in a high-dimensional hole-displacement space. The comparison with the genetic algorithm and the leaky-mode method is framed fairly; if anything, their 25-parameter optimization is harder than the 5- and 9-parameter optimizations in [16] and [17], so the benchmark is not rigged in their favor. The method is clearly described, and the computational budget (8070 samples) is documented.\n\nThe main weakness is the absence of any FDTD convergence study. For a cavity with Q above 1e7, the computed Q can depend sensitively on mesh resolution, PML placement, and boundary conditions. The text says the calculation conditions are the same as in [19], but that is a reference to their own prior paper, not evidence of convergence. A referee should ask for a mesh-refinement test at least at the final structure, and ideally for a few structures along the optimization path. Without that, the \"more than twice the previous record\" claim is not fully established, because the prior records came from different simulation settings (slab thickness 0.55a in [16] vs 0.5366a here, for example). That difference is not necessarily fatal, but it needs to be addressed head-on.\n\nTwo smaller points: no code or dataset is released, which hurts reproducibility of the method details; and the claim that the approach generalizes to generic high-dimensional optimization is extrapolated from one cavity type and should be softened or supported with another example.\n\nOverall, this is a serious piece of work with a clearly positive result. I would send it to peer review, not desk reject, but I would require the convergence test and a comparability discussion before accepting the record claim. For the reading group, it is worth a session on how machine learning can be coupled with expensive simulations in photonics. I would cite it in my own work on optimization methods, with a caveat about the unverified FDTD values.","headline":"A credible iterative deep-learning optimization of L3 cavities with an 11M-Q FDTD result, but the record claim needs convergence evidence before it fully lands.","tokens_in":13221,"tokens_out":1673,"would_cite":true,"duration_ms":20553,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.70.Qs","42.60.Da"],"model":"deepseek-v4-flash","headline":"An iterative deep-learning loop finds an L3 photonic-crystal cavity with Q = 11 million, more than twice the previous record.","keywords":["photonic crystal nanocavity","quality factor","deep neural network","iterative optimization","L3 cavity","finite-difference time-domain","gradient-based search","high-dimensional design space"],"falsifier":"Recalculate the Q of the paper's best cavity (Figure 5c) with an independent FDTD solver on a sequence of finer meshes; if the extrapolated Q does not exceed 5.3 million, the claimed improvement is not real. A second decisive check is to rerun the 101-round loop from the same initial dataset and confirm that the $Q=1.10\\times 10^{7}$ structure appears within 8,070 Q evaluations.","tokens_in":12226,"feed_emoji":"🧠","tokens_out":10896,"duration_ms":103681,"temperature":0.7,"pith_summary":"The paper proposes a closed-loop recipe for discovering photonic-crystal nanocavity designs with record-high quality factors: train a neural network to predict Q from air-hole displacements, use the network's gradient to propose new structures, compute their Q with 3D finite-difference time-domain (FDTD) simulation, add those to the training set, and repeat. Applied to a standard silicon L3 cavity with 25 independent displacement parameters, the loop finds a design with $Q = 1.10\\times 10^{7}$ within 101 iterations and 8,070 total simulated samples, more than twice the best Q previously reported for this cavity type. The key ingredient is a search penalty that pushes candidate structures away from already-evaluated designs, which the paper shows yields better results than staying near the current best or adding random exploration. A sympathetic reader should care because the method automates the exploration of the high-dimensional hole-position space that manual, physics-guided optimization cannot cover.","feed_headline":"Neural-network loop finds a cavity with Q over 11 million","feed_subtitle":"An automated train-predict-revalidate cycle beats prior genetic and visual methods within the same FDTD budget.","key_machinery":"The load-bearing mechanism is the iterate: neural-network regression plus a gradient search whose loss function contains artificial penalty terms. The network maps the 25-dimensional hole-displacement vector to $\\log_{10}Q$ and provides a smooth surrogate whose gradient does not vanish in high dimensions. Search is performed by minimizing $L' = |\\log_{10}Q_{\\text{target}} - \\log_{10}Q_{\\text{NN}}|^2 + \\text{artificial loss}$, with $Q_{\\text{target}}=10^{8}$, starting from random initial structures. Three artificial losses are tested: (A) squared distance to the best previous structure, (B) squared distance to a random initial structure, and (C) a sum of inverse distances to all dataset structures; (C) makes the search avoid already-known regions. The paper's demonstration that strategy (A+C) finds the highest Q shows that forcing exploration of unexplored parameter space is what carries the optimization to record values.","core_discovery":"On its own terms, the paper's central claim is that iteratively augmenting the training data with the structures the network itself proposes turns a rough regression model into a practical search engine for very high Q cavities. Starting with 1,000 random L3 structures (Q spread from $10^{3}$ to $10^{5}$), ten networks are trained to predict $\\log_{10}Q$ from the 25 independent in-plane hole displacements (the 50-hole optimization region reduced by mirror symmetry). Each round, 70 candidates are produced by gradient descent on a loss that targets $Q=10^{8}$ while one of three artificial penalty terms constrains the explored region, and their true 3D-FDTD Q values are appended. After 101 rounds, the best structures found are $Q=5.75\\times 10^{6}$, $Q=9.12\\times 10^{6}$, and $Q=1.10\\times 10^{7}$ for strategies (A), (A+B), and (A+C), respectively. The $Q=1.10\\times 10^{7}$ result is presented as more than twice the previous records of $4.2\\times 10^{6}$ (genetic algorithm) and $5.3\\times 10^{6}$ (leaky-mode visualization) for an Si L3 cavity, with a comparable number of FDTD samples.","pith_inferences":["A natural extension the authors leave implicit is applying repulsive exploration (C) to multi-objective design, where the network predicts several metrics and the same 'avoid known points' pressure maps trade-offs between Q and modal volume.","The inverse-distance penalty is a crude stand-in for prediction uncertainty; a testable variant would replace it with the variance across the ten networks and use that variance to guide exploration.","If the FDTD record survives an independent mesh-converged check, the same loop should be carried through fabrication; a measured Q near $10^{7}$ would turn the theoretical record into a device claim."],"forward_implications":["The loop should transfer to other cavity geometries and other figures of merit, since the network output and the validation simulator can be changed independently.","The 8,070-sample dataset accumulated under strategy (A+C) is a reusable resource; later optimization runs can warm-start from it instead of sampling randomly.","The result sets a new benchmark for L3 nanocavity design: a method should beat $Q = 1.10\\times 10^{7}$ with no more FDTD evaluations to claim an improvement.","The linear training cost of the neural network versus the cubic cost of Bayesian optimization makes this kind of iterative surrogate loop the more scalable option for high-dimensional design spaces."],"supporting_citations":[{"why":"Supplies the base deep-learning regression setup: network architecture, random initial dataset, and log10 Q as the learning target; the iterative loop extends this prior method.","marker":"[19]"},{"why":"Provides the genetic-algorithm benchmark: a Q of 4.2 million with about 8,000 sample cavities, which the new Q must beat.","marker":"[16]"},{"why":"Provides the leaky-mode visualization benchmark: a Q of 5.3 million with about 200 manually chosen samples.","marker":"[17]"},{"why":"Establishes the momentum-space description of leaky components that motivates why the hole pattern controls the Q factor.","marker":"[12]"},{"why":"Supplies the convolutional-layer construction used in the neural networks that map hole-displacement maps to Q.","marker":"[21]"},{"why":"Defines Bayesian optimization and its cubic cost scaling, the baseline against which the paper argues its linear-cost NN loop scales better.","marker":"[28]"}],"fun_headline_variants":["Neural net iteration finds nanocavity with Q >11M","AI-driven search cracks 11M Q barrier for nanocavities","Iterative ML search doubles Q record to 11M","Neural network iterates to 11M Q cavity, 2x record"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that the simulated Q factors from 3D-FDTD are accurate enough to rank cavities at Q near 10 million, and that the authors' FDTD conditions match those used for the earlier record values; neither is verified by an error or convergence test.","fun_headline_variants_meta":{"raw":{"variants":["Neural net iteration finds nanocavity with Q >11M","AI-driven search cracks 11M Q barrier for nanocavities","Iterative ML search doubles Q record to 11M","Neural network iterates to 11M Q cavity, 2x record"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001148,"raw_usage":{"total_tokens":4835,"prompt_tokens":1095,"completion_tokens":3740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":3664}},"tokens_in":711,"tokens_out":3740,"duration_ms":28114,"temperature":1.0,"reasoning_tokens":3664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:04:41.859649+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recalculate the Q of the paper's best cavity (Figure 5c) with an independent FDTD solver on a sequence of finer meshes; if the extrapolated Q does not exceed 5.3 million, the claimed improvement is not real. A second decisive check is to rerun the 101-round loop from the same initial dataset and confirm that the $Q=1.10\\times 10^{7}$ structure appears within 8,070 Q evaluations.","supporting_citations":[{"cited_title":"Opti izatio of photo ic crystal a ocavities based o deep lear i g,","cited_arxiv_id":null,"evidence_quote":"Supplies the base deep-learning regression setup: network architecture, random initial dataset, and log10 Q as the learning target; the iterative loop extends this prior method."},{"cited_title":"Auto ated opti izatio of photo ic crystal slab cavities,","cited_arxiv_id":null,"evidence_quote":"Provides the genetic-algorithm benchmark: a Q of 4.2 million with about 8,000 sample cavities, which the new Q must beat."},{"cited_title":"I prove e t i the quality factors for photo ic crystal nanocavities via visualization of the leaky co po e ts,","cited_arxiv_id":null,"evidence_quote":"Provides the leaky-mode visualization benchmark: a Q of 5.3 million with about 200 manually chosen samples."},{"cited_title":"Mo e tu space design of high-Q photo ic crystal optical cavities,","cited_arxiv_id":null,"evidence_quote":"Establishes the momentum-space description of leaky components that motivates why the hole pattern controls the Q factor."},{"cited_title":"Ha dwritte Digit Recog itio with a Back-Propagatio Networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional-layer construction used in the neural networks that map hole-displacement maps to Q."},{"cited_title":"Taki g the hu a out of the loop: A review of Bayesia opti izatio ,","cited_arxiv_id":null,"evidence_quote":"Defines Bayesian optimization and its cubic cost scaling, the baseline against which the paper argues its linear-cost NN loop scales better."}],"review_version":1}