{"id":"3d764272-a30d-493d-be36-b457b5a34d18","arxiv_id":"2411.17734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A planar monopulse comparator network using port-transformation rat-race couplers and a 360-degree crossover is measured at 2 GHz, and a DNN is trained to correct remaining angle errors.","lead":"This paper builds and tests a low-cost radar direction-finding receiver using a new flat circuit layout, then adds a small neural network to correct the angle readings. The main hardware contribution is a planar comparator network that avoids the expensive multi-layer boards normally used in monopulse radar.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DNN 'enhanced accuracy' claim rests on a 90/10 random split of 145 correlated trajectory samples, with no independent test set or calibration baseline; the 1.1 mm error may be interpolation, not generalization.","rationale":"The paper's concrete hardware contribution—the port-transformation rat-race coupler with a 360° crossover and the symmetric comparator network—is supported by measurement (Figs. 7-10) and a comparison table, so it would be wrong to treat this as a fabricated or fundamentally unsound design. The serious soft spot is the DNN evaluation in Section III-C. The reader's weakest assumption correctly identifies that the 145-sample single-trajectory dataset cannot support a general accuracy claim; my stress test adds that the random 90/10 split is especially problematic because consecutive samples along a continuous trajectory are near-duplicates, so the validation loss of 2e-6 (1.1 mm) is likely interpolation. The lack of any calibration baseline makes it impossible to know whether DNN adds value over a simple polynomial correction. A spatially disjoint holdout test with a baseline comparison would settle this. This does not move the overall conditional verdict: the hardware should be accepted, but the DNN claim needs revision or the scope needs narrowing. Therefore verdict remains UNCHANGED relative to the reader.","tokens_in":11759,"tokens_out":5440,"duration_ms":51933,"concrete_test":"Collect a new, spatially disjoint validation dataset: place the target on a grid with 60 mm spacing, train the DNN only on one checkerboard parity of grid positions and test on the other parity, including two distances outside the original 0.62 m trajectory. Fit a simple 2D polynomial or lookup-table calibration baseline to the same training parity. If the DNN RMSE on the held-out parity is not significantly lower than the baseline, or if errors grow sharply away from training positions/distances, the claimed 1.1 mm accuracy is not a general result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'enhanced estimation accuracy' claim is supported only by Section III-C. A single DNN (3 fully connected layers of 20-50-10 neurons, Eq. (15)) is trained on 145 samples recorded along one continuous target trajectory; 90% are used for training and the remaining 10% for validation. Because the samples come from a continuous trajectory, neighboring samples are strongly correlated, so a random 10% split leaks near-duplicate points into the validation set, and the reported loss (2e-6, ~1.1 mm at 0.62 m) is an in-sample interpolation metric. There is no independent test set, no comparison against a conventional calibration baseline (e.g., a polynomial or lookup table fit to the same data), no error bars, and no code/data release. The DNN input also includes distance, but the training description does not establish that the distances shown in Fig. 12(b)-(c) (0.66 m, 0.86 m) are covered by training, so the DNN may be extrapolating in distance. If the DNN merely memorizes the training trajectory, the claimed enhanced accuracy does not carry over to arbitrary target positions. The hardware comparator network itself has independent measured support (S-parameters, radiation patterns), so the concern is isolated to the DNN portion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 2 GHz low-cost monopulse receiver combining a 2x2 patch array, a fully planar symmetrical monopulse comparator network built from four port-transformation rat-race couplers and a 360° phase-delay crossover, a downconversion link, and a DNN-based post-processing stage for DoA estimation. The comparator design is described through even-odd mode analysis and validated by simulated and measured S-parameters and radiation patterns. The DNN (three fully connected layers with 20-50-10 neurons) takes estimated azimuth, elevation, and distance as input and outputs corrected angles; it is trained on 145 measured samples from a moving target antenna and reports a validation error equivalent to about 1.1 mm at 0.62 m. The paper claims fast real-time tracking with high accuracy in a low-cost planar implementation.","tokens_in":12039,"tokens_out":14538,"duration_ms":136227,"significance":"The hardware contribution is credible and useful: the proposed comparator network achieves a fully planar, symmetric monopulse feed with about 10% measured bandwidth, and the measured S-parameters and radiation patterns are consistent with simulation. If the DNN accuracy claim were rigorously validated, the system would be a practical low-cost calibration approach for monopulse tracking. However, the current DNN evidence is an in-sample validation with no independent test set, no baseline comparison, and no error bars, so the enhanced-accuracy claim is not yet established. The paper's potential significance is therefore moderate and concentrated in the comparator network; the DNN contribution needs substantial additional evaluation.","major_comments":[{"comment":"The central DNN accuracy claim rests on an invalid generalization test. The DNN is trained on 145 samples collected along one continuous target trajectory with a 30 mm grid spacing, and the reported validation loss (2e-6, about 1.1 mm at 0.62 m) is obtained from a random 90/10 split of those samples. Because neighboring samples on a continuous trajectory and lattice are strongly correlated, a random split places near-duplicate points in both training and validation sets, so the validation loss is an interpolation metric rather than a measure of generalization. With only about 14 validation samples, no error bars, no regularization or hyperparameter search, and a comparatively high-dimensional hidden-layer parameterization, the 1.1 mm figure does not support the claim that the DNN estimates arbitrary target positions. The authors need an independent test set of positions and distances not seen in training, repeated splits, and error bars on the reported error.","section":"III-C, Eq. (15)"},{"comment":"The distance coverage of the training set is not established, so the results at D = 0.66 m and 0.86 m may be extrapolation. The validation error is quoted at a separation of 0.62 m, but the text does not state whether the 145 training samples were collected at that distance or over a range of distances that includes 0.66 m and 0.86 m. Since distance is one of the three DNN inputs, the network may use distance to memorize the measurement geometry; since the far-field monopulse ratio in Eqs. (13) and (14) does not depend on range, the inclusion of this feature also needs physical justification. The authors should specify the training distance distribution, report errors separately for each test distance, and confirm that the test distances are within the training support or explicitly evaluate extrapolation.","section":"III-C, Fig. 12"},{"comment":"No baseline comparison is provided for the claimed accuracy enhancement. The paper reports only the DNN validation loss and a qualitative visual agreement in Fig. 12; it does not quantify the uncorrected monopulse-ratio error, nor does it compare the DNN with a conventional calibration method such as a polynomial or lookup-table fit to the same 145 training samples. The introduction explicitly contrasts DNNs with lookup tables, but no empirical comparison is given. Because the DNN is trained and validated on measurements from the same receiver, the method is currently best described as a fitted calibration function; to support the title, the authors should report RMSE or MAE with error bars for the raw estimates, the DNN-corrected estimates, and a calibration baseline on an independent test set.","section":"III-C"},{"comment":"The DNN input includes the distance between the radar and the target, but the manuscript never states how this distance is measured by the proposed monopulse receiver. The receiver chain described in Fig. 1 and Section III-C (array, comparator, downconverter, VSA) provides only the sum and difference channels used to derive angles in Eqs. (13)-(14); no ranging scheme (e.g., FMCW, time-of-flight, or received-signal-strength) is presented. If distance is assumed known from the test setup, the claim of a real-time tracking system is incomplete; the authors should either remove distance from the input, describe a ranging method, or clearly limit the DNN to scenarios where the range is externally supplied.","section":"III-C and Fig. 11"}],"minor_comments":[{"comment":"The even-odd ABCD matrices in Eqs. (1)-(2) are garbled in the submitted PDF and cannot be checked; please ensure the final typeset equations are legible.","section":"II-A, Eqs. (1)-(2)"},{"comment":"The operating frequency is given as 2 GHz in the abstract and Section II, while the measured comparator and array results in Figs. 8-10 are centered at 1.95 GHz; the text should state this fabrication-tolerance shift consistently when introducing the measurements.","section":"Abstract and Section III-A"},{"comment":"Table I and the introduction should explicitly differentiate the present comparator from [17], which also reports a fully symmetric planar comparator, and from the conference versions [24] and [25]; as written, the incremental novelty is not clear.","section":"Table I and Introduction"},{"comment":"Section III-C reports a training time of about 6 minutes but gives no inference latency; the 'real-time' and 'milliseconds' claims should be supported by a measured or estimated per-query inference time.","section":"III-C"},{"comment":"Fig. 12(b)-(c) shows only qualitative agreement; please add tabulated RMSE/MAE and maximum error per distance, and label the axes with angle units and the DNN input/output quantities.","section":"Fig. 12"},{"comment":"For reproducibility, please report the random split seed, optimizer learning rate, batch size, and whether any regularization was used; the current description of the training procedure is incomplete.","section":"III-C"},{"comment":"The statement that the DNN 'requires small size of dataset' is not supported by any experiment; a comparison with a reduced training-set size or with a polynomial calibration baseline is needed.","section":"III-C"},{"comment":"The conversion from the angle loss in Eq. (15) to a 1.1 mm position error at 0.62 m should be stated explicitly, including the units of the angles and the geometry used for the conversion.","section":"III-C and Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The DNN validation is the main obstacle to publication. The hardware portion is sound and could be published on its own. I would encourage the editor to require an independent test set and a calibration baseline before acceptance; if the authors cannot provide these, the DNN-related claims should be substantially weakened or reframed as a calibration demonstration on the measured trajectory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the comparator network is a real circuit contribution, well supported by measurements; the DNN part is a calibration fit on a tiny dataset and the headline accuracy claim does not survive scrutiny. The two should be separated.\n\nWhat is actually new and good: the port-transformation rat-race coupler with the 360-degree crossover is a neat way to avoid port crossing in a planar comparator. The even-odd mode design equations are parameter-free, and the measured S-parameters and radiation patterns agree with simulation. Getting a fully planar, symmetric monopulse comparator on a single-layer PCB at 1.95 GHz with about 10% bandwidth and null depths above 22 dB is a legitimate engineering result. The comparison table with prior work is useful and fair.\n\nThe soft spot is Section III-C. The DNN is trained on 145 samples taken along one continuous trajectory. A random 90/10 split leaks strongly correlated neighbors into the validation set, so the reported loss of 2e-6 (about 1.1 mm at 0.62 m) is an interpolation metric, not a generalization result. There is no independent test set, no baseline comparison (polynomial or lookup-table calibration), no error bars, and no code or data. The input includes distance, but the figures show test distances of 0.66 m and 0.86 m, and the text does not establish that those are covered by the training grid. So the claim that the DNN enhances estimation accuracy for arbitrary target positions is not supported. This is a load-bearing flaw for the paper's title and abstract, though it does not affect the hardware derivation.\n\nA separate small point: the paper says the DNN handles noise and uncertainty better than a lookup table, but no evidence is given for that claim, and with 145 samples a well-chosen LUT would likely do as well.\n\nOverall: the hardware deserves peer review, the DNN section does not. The paper is acceptable only if the DNN claims are either substantially revised with proper validation or removed entirely and the paper is reframed as a comparator-network design. As it stands, the accuracy claim is overreach.\n\nI would send this to peer review because the circuit contribution is real and referees can help fix the DNN framing. But I would not accept it in the current form.","headline":"The planar comparator network is a solid, measured hardware contribution; the DNN accuracy claim is an under-validated calibration fit and should not be accepted as-is.","tokens_in":12555,"tokens_out":1649,"would_cite":true,"duration_ms":16695,"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":"Moving one input port with a 360° phase-delay crossover makes the rat-race comparator planar and symmetric on one PCB layer, and a small neural network then cuts validated position error to about 1.1 mm at 0.62 m.","keywords":["monopulse receiver","direction of arrival estimation","rat-race coupler","microwave crossover","planar comparator network","deep neural network","patch antenna array"],"falsifier":"Move the probe target to a position that is not on the 145-sample training line, for example a different range or an untrained quadrant, and compare the DNN-corrected coordinates with measured truth; the accuracy claim holds only if the error stays near 1.1 mm instead of reverting to the raw monopulse error.","tokens_in":11573,"feed_emoji":"📡","tokens_out":15580,"duration_ms":129089,"temperature":0.7,"pith_summary":"This paper tries to show that a monopulse tracking receiver—one that locates a target by comparing sum and difference signals from an antenna array—can be built on low-cost hardware. The analog comparator network that forms those sum and difference signals is redesigned so it lies flat on a single PCB layer, with no crossing lines, multilayer board, or backside feed. The redesign uses four port-transformation rat-race couplers in which a 360° phase-delay crossover relocates one input port, and a measured 1.95 GHz prototype keeps amplitude imbalance within 0.5 dB and phase imbalance within 10° over about 10% bandwidth. On top of this hardware, a small fully connected neural network takes the raw azimuth, elevation, and distance estimates and corrects them toward the true target coordinates, reaching a validation position error near 1.1 mm at 0.62 m range after training on only 145 samples. If correct, this gives a path to low-cost, real-time monopulse tracking without expensive fabrication steps.","feed_headline":"Planar coupler plus DNN cuts tracking error to 1.1 mm","feed_subtitle":"Single-layer comparator replaces multilayer boards; a neural network corrects the residual angle.","key_machinery":"The load-bearing component is the port-transformation 180° rat-race coupler, a modified hybrid ring that normally combines two inputs into sum and difference outputs but whose ports cross in a planar layout. The mathematical fact that carries the design is that in even/odd-mode analysis the 180° phase shifter is open or short, so its impedance $Z_\\eta$ drops out of the ABCD matrices and can be chosen freely; the crossover itself is designed from four-way even/odd input impedances to satisfy $|S_{xx'}| = |S_{yy'}| = 1$ with zero leakage between the other port pairs. This identity lets four modified couplers and another crossover tile into a symmetric $1\\lambda \\times 1\\lambda$ planar comparator. The second mechanism is a fully connected neural network with hidden layers of 20, 50, and 10 neurons, whose loss is the mean squared error between corrected and true azimuth and elevation angles; it is trained on 145 measured samples of a moving target.","core_discovery":"The central claim is that the port-crossing problem of a conventional rat-race coupler can be removed within a single layer: move one input port to the same side as the other and insert a 360° phase-delay crossover, and the coupler still behaves as a proper 180° hybrid. The even/odd-mode ABCD derivation shows that the added 180° phase shifter appears as an open or short and therefore contributes no impedance value, so its characteristic impedance can be chosen purely for layout convenience. Four such couplers connected with another crossover form a fully planar, symmetric monopulse comparator whose sum and difference ports align naturally with a 2×2 patch array; measurement gives better than 17 dB return loss, 7.2 ± 0.5 dB transmission, and the required 0°/180° phase relationships over about 10% bandwidth. The DNN part then treats the mismatch between monopulse-ratio angles and true coordinates as a learnable systematic error, with a three-hidden-layer network mapping azimuth, elevation, and distance to corrected angles.","pith_inferences":["The 1.1 mm validation error is measured on a 10 percent holdout from the same 145-sample trajectory; a stronger test would be a disjoint target path or an untrained grid of positions, since a small network could partly memorize a single line of samples.","The port-reallocation identity should transfer to other 180° hybrids and other transmission media, so the single-layer symmetric comparator concept may scale to higher bands without a new topology.","Because the neural network learns one receiver's specific systematic distortion, practical deployment would likely need occasional retraining when the antenna, down-converter, or surrounding environment changes."],"forward_implications":["A monopulse comparator network can be fabricated on a single PCB layer, avoiding multilayer lamination and backside feeds, while keeping about 10 percent measured bandwidth at 1.95 GHz.","The resulting 2×2 patch array produces sum gain near 11 dBi with sidelobes better than 15 dB and difference-pattern null depths better than 22 dB.","A neural network trained on fewer than 150 samples can correct residual angle errors in real time, with validation position error near 1.1 mm at 0.62 m.","The design equations for the coupler and crossover supply a repeatable procedure for building symmetric comparators at other center frequencies."],"supporting_citations":[{"why":"Presents the prior planar symmetric 180° coupler design whose unequal group delay the new topology is meant to overcome.","marker":"[17]"},{"why":"Describes the earlier conference study of DNN-boosted monopulse estimation that this work extends to a complete receiver.","marker":"[24]"},{"why":"Conference version introducing the port-transformation rat-race comparator network that this paper expands with array, receiver, and DNN.","marker":"[25]"},{"why":"Supplies the ABCD and even/odd-mode transmission-line analysis used to derive the coupler's port-transformation property.","marker":"[26]"},{"why":"Provides the crossover analysis from which the 360° phase-delay crossover design equations are taken.","marker":"[27]"}],"fun_headline_variants":["Single-layer coupler plus DNN sharpens DoA accuracy","Monopulse receiver slashes error with planar coupler and neural net","DNN-backed monopulse coupler hits 1.1 mm accuracy","Low-cost planar monopulse gets neural boost for precise angles","Neural network fixes monopulse angle error in flat coupler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a neural network trained on 145 samples of continuous target locations corrects angles for targets anywhere else, since it is validated on a 10 percent split of the same dataset rather than on an independent set of positions or distances.","fun_headline_variants_meta":{"raw":{"variants":["Single-layer coupler plus DNN sharpens DoA accuracy","Monopulse receiver slashes error with planar coupler and neural net","DNN-backed monopulse coupler hits 1.1 mm accuracy","Low-cost planar monopulse gets neural boost for precise angles","Neural network fixes monopulse angle error in flat coupler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2354,"prompt_tokens":984,"completion_tokens":1370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":1281}},"tokens_in":600,"tokens_out":1370,"duration_ms":9843,"temperature":1.0,"reasoning_tokens":1281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:23:50.531026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Move the probe target to a position that is not on the 145-sample training line, for example a different range or an untrained quadrant, and compare the DNN-corrected coordinates with measured truth; the accuracy claim holds only if the error stays near 1.1 mm instead of reverting to the raw monopulse error.","supporting_citations":[{"cited_title":"A Fully Symmetrical Uni - Planar Microstrip Line Comparator Network for Monopulse Antenna,","cited_arxiv_id":null,"evidence_quote":"Presents the prior planar symmetric 180° coupler design whose unequal group delay the new topology is meant to overcome."},{"cited_title":"Boosting Estimation Accuracy of Low -Cost Monopulse Receiver Via Deep Neural Network,","cited_arxiv_id":null,"evidence_quote":"Describes the earlier conference study of DNN-boosted monopulse estimation that this work extends to a complete receiver."},{"cited_title":"A Planar Monopulse Comparator Network Design from Port - Transformation Rat-Race Coupler,","cited_arxiv_id":null,"evidence_quote":"Conference version introducing the port-transformation rat-race comparator network that this paper expands with array, receiver, and DNN."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ABCD and even/odd-mode transmission-line analysis used to derive the coupler's port-transformation property."},{"cited_title":"A Novel Dual -Band Zero-Phase True Crossover with Arbitrary Port Impedance,","cited_arxiv_id":null,"evidence_quote":"Provides the crossover analysis from which the 360° phase-delay crossover design equations are taken."}],"review_version":1}