{"id":"f5312173-3d61-4bfe-84ae-3350b9fd09c3","arxiv_id":"1909.02262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single CNN trained on one 150 Hz frequency band detects hour-to-day neutron star transient signals across bands with sensitivity comparable to the Generalized FrequencyHough, and was applied to real LIGO data from the GW170817 remnant search.","lead":"This paper tests whether convolutional neural networks can find long-lasting gravitational wave signals from spinning neutron stars in noisy LIGO data. The networks match an established search method in sensitivity, run far faster, and were used in a real search for the remnant of the 2017 neutron star merger GW170817.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-detector transfer of the CNN is asserted but not tested: all calibration and robustness results use Livingston O2 noise, while the GW170817 search applies the same network to Hanford at a Livingston-derived threshold.","rationale":"The paper is a genuinely useful empirical study: it reports a real-data search, quantifies false alarms, retrains the network 1255 times to estimate error bars, and shows that modest training sets achieve efficiencies comparable to a standard method. Those elements independently support the claim that CNNs can work for long transients in real data. The load-bearing weakness is the transfer step. Frequency transfer is tested directly in Fig. 3, so the more fragile part of the generalization claim is detector transfer: all calibration, robustness, and efficiency measurements in Section IV use Livingston O2 noise, yet the real search in Section VI applies the network to Hanford and uses the same pthr=0.9 with only a time-coincidence requirement. A direct per-detector false-alarm and efficiency test would settle whether the search's trigger count and upper limits are interpretable. This is exactly the weakest assumption identified by the reader, and my assessment does not move the verdict: the concern is real but addressable, and the appropriate outcome remains CONDITIONAL, pending the per-detector validation check.","tokens_in":13113,"tokens_out":4274,"duration_ms":46196,"concrete_test":"Apply the trained CNN to about 500 Hanford O2 noise maps drawn from the same epoch as the Livingston calibration, without injections, and measure the false-alarm probability versus pthr; compare with the Livingston curve in Fig. 4. Then inject the same n=7, fixed-amplitude power-law signals used in Fig. 6 into Hanford maps and measure detection efficiency at pthr=0.9, comparing with Livingston efficiency at equal false-alarm probability. If Hanford false-alarm or efficiency differs by more than the 90% error bars in Fig. 5, the Section VI search and upper limits must be re-derived with per-detector thresholds before the real-data claim is accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that one CNN trained on signal/noise maps in a single 150 Hz band of O2 Livingston data can be applied to any band and to both LIGO detectors, with only per-band threshold recalibration. Section IV C calibrates pthr on 500 Livingston noise maps, and Fig. 3 tests frequency transfer only on Livingston data. Section VI then applies the same network to Hanford data and accepts p>0.9 triggers, relying on a time-coincidence step to control false alarms. No Hanford-only false-alarm rate or efficiency curve is reported. If Hanford noise artifacts, lines, or non-stationarity differ in a way the persistency veto does not remove, the per-map false-alarm probability at pthr=0.9 is unknown; the 50 CNN triggers in Section VI B and the CNN-based upper limits in Section VI C then have no calibrated sensitivity. The non-detection does not certify the transfer, because a null result is compatible with both a working and a miscalibrated classifier. This is a concrete, addressable gap rather than a contradiction: the paper's own analysis shows threshold selection is noise-dependent (Fig. 4), and no measurement establishes that Hanford noise is equivalent after cleaning.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an empirical study of convolutional neural networks (CNNs) for detecting long-duration transient gravitational waves from isolated neutron stars. It trains CNNs on reduced time/frequency maps built from O2 Livingston data, characterizes detection efficiency and false-alarm probability as functions of training-set size, braking index, frequency band, and output threshold pthr, compares the CNN against the Generalized FrequencyHough (GFH) and other architectures, and applies the network to a one-week search for a GW170817 remnant. The central claims are: a CNN trained on a single 150 Hz band can be applied to other bands and to both LIGO detectors; CNN sensitivity is comparable to the GFH while being orders of magnitude faster; CNNs can detect signals with time-varying braking index that the GFH cannot; and the authors have performed the first machine-learning-based search for an isolated-neutron-star gravitational-wave signal.","tokens_in":13361,"tokens_out":6582,"duration_ms":73125,"significance":"If the claims hold, this is a useful proof-of-principle: it shows that a small CNN can act as a fast trigger generator for long-transient searches, and it includes a concrete search design and a real-data demonstration. Strengths include the 1255-retraining error analysis, the explicit treatment of pthr and persistent-line vetoes, the robustness tests across frequency bands and signal morphologies, and the final search with GFH follow-up. I found no circularity: training and threshold selection use simulated injections in real noise, and the final search is on previously unseen data. The significance is, however, conditional on two calibration questions: transfer of the network to Hanford data and fairness of the CNN-versus-GFH sensitivity comparison.","major_comments":[{"comment":"The real search applies to Hanford data a network whose false-alarm thresholds were calibrated exclusively on O2 Livingston data, without any Hanford-only false-alarm or efficiency measurement. Section IV C determines pthr from 500 Livingston noise maps, and Section VI B accepts Hanford triggers with p > 0.9, relying on time coincidence to control false alarms. A null result is compatible with both a well-calibrated and a miscalibrated classifier, so the CNN upper limits in Section VI C are not fully supported as presented. Please add noise-only Hanford maps to measure the false-alarm probability at the chosen threshold, and ideally Hanford injection studies to show that the detection efficiency also transfers.","section":"VI B, VI C (cf. IV C)"},{"comment":"The headline CNN-versus-GFH sensitivity comparison is made at different false-alarm levels: the text reports about 1% false-alarm probability for the CNN/ANN and about 0.01% for the GFH at the operating points used. 'Similar efficiencies' is therefore not an apples-to-apples statement. Please provide efficiency curves at matched false-alarm probability, or ROC-style results, before concluding that the CNN is comparable in sensitivity to the GFH.","section":"IV E, Fig. 6"},{"comment":"The paper states that the GFH cannot detect signals with time-varying braking index and delta-n/delta-t in [-1e-4, 1e-4] /s, but no GFH efficiency on those injections is shown in this manuscript. This claim is load-bearing for the 'signals to which the GFH is blind' novelty. Please quantify GFH performance on the same varying-braking-index test set, or restrict the statement to the power-law model of Eq. (3) with a clear caveat.","section":"IV B, Fig. 3"}],"minor_comments":[{"comment":"The 32 amplitude curves in Fig. 2 are difficult to read without a legend or colorbar; please label the amplitudes explicitly or use a color scale.","section":"IV A, Fig. 2"},{"comment":"Please clarify whether the 1255 retrainings differ only by random seeds or also by data batching and dropout realizations, and state which of the resulting networks is used for the thresholds applied in Sections IV E and VI.","section":"IV D"},{"comment":"Please report the total number of time/frequency maps analyzed in the search so that the reader can convert the 50 CNN triggers into a false-alarm rate rather than only an event count.","section":"VI B"},{"comment":"The blue 'CNN-only' upper-limit curve needs an explicit detection criterion: which pthr is used, and what false-alarm probability is assumed, since pthr is not by itself a calibrated significance.","section":"VI C"},{"comment":"Please state explicitly for each characterization result that the noise maps used for testing were disjoint from the noise maps used for training; the paper warns about this in Section V but does not always make the train/test separation explicit for the individual figures.","section":"IV and V"},{"comment":"If AlexNet is used as a comparison architecture, please cite the original AlexNet reference in addition to the in-house implementation reference.","section":"IV E"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the core idea is promising. The two main gaps—Hanford threshold calibration and matched-false-alarm sensitivity comparison—are concrete and fixable, and addressing them would substantially strengthen the paper. I do not see a basis for rejection, but the current presentation overstates the CNN/GFH comparison and the cross-detector robustness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it actually does something new: it shows that a single CNN trained on signal/noise maps in one 150 Hz band of O2 Livingston data can pick out long-duration neutron-star signals in other frequency bands, including signals with time-varying braking indices that the Generalized FrequencyHough (GFH) cannot see. Second, it ends with a real search for a GW170817 remnant, which is the first ML-based search for an isolated neutron star signal. Those are real contributions, not incremental noise.\n\nThe paper does several things well. The training-efficiency study is useful: you do not need huge injection sets, which matters for low-latency use. The 1255-network error analysis is a serious attempt to quantify the black-box randomness, and the false-alarm characterization across frequency bands is careful. The comparison to GFH, showing comparable efficiencies at orders-of-magnitude lower cost, is credible as a scaling statement. The authors are also honest about the GFH grid issue in the upper-limit section and about the importance of line cleaning.\n\nNow the soft spots, in order of severity. The headline CNN-versus-GFH sensitivity comparison is made at different false alarm rates: about 1% for the CNN versus 0.01% for GFH. That is an apples-to-oranges comparison; a matched-FAR or ROC-style curve would make \"similar sensitivities\" a real claim rather than a suggestive one. The stress-test note about cross-detector transfer is correct: the CNN is trained and threshold-calibrated on Livingston noise, then applied to Hanford in the GW170817 search without reporting a Hanford-only false-alarm rate or efficiency curve. The null result does not certify the transfer, since a miscalibrated classifier can also produce no candidates. This is addressable with one extra experiment, but it is a genuine gap. The \"first time ever\" claim is also a bit narrower than it sounds: refs [19] and [32] have applied ML to continuous waves and r-mode signals from neutron stars, so the novelty is the long-transient real search, not ML for neutron stars per se. Finally, no code or trained models are released, which hurts reproducibility but is a minor point for a paper of this type.\n\nThe central empirical claims are not circular and are broadly supported; the flaws are all fixable. This paper deserves serious peer review, and I would accept it with moderate revision. The cross-detector calibration gap should be closed before publication, and the sensitivity comparison should be put on equal false-alarm footing. If those are done, this becomes a solid reference for anyone building ML trigger stages for long transients.\n\nFor a reading group, this is worth a look if the discussion is on ML in GW data analysis or long-transient search pipelines. I would cite it in that context.","headline":"A genuinely useful empirical study of CNNs for long-duration neutron-star transients: the cross-band generalization and real-data search are new and mostly well supported, but the headline sensitivity comparison and the cross-detector transfer both need tightening before the numbers are taken at face value.","tokens_in":13921,"tokens_out":1366,"would_cite":true,"duration_ms":16715,"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":"One CNN trained in one band finds neutron-star signals in any band, matching the standard search's sensitivity and catching signals it misses.","keywords":["gravitational waves","neutron stars","long-duration transients","convolutional neural networks","machine learning","Generalized FrequencyHough","GW170817 remnant","time-frequency maps"],"falsifier":"Train the same network on one 150 Hz band of O2 Livingston data, then run it on a large sample of Hanford noise maps from a different epoch (or on O3/O4 data) without retraining; if the false alarm probability at the calibrated threshold is far above 1% even after per-band recalibration, or if the detection efficiency on time-varying-braking-index injections falls below the GFH's, the transferability claim is refuted.","tokens_in":12945,"feed_emoji":"📡","tokens_out":7381,"duration_ms":57800,"temperature":0.7,"pith_summary":"This paper asks whether a convolutional neural network (CNN) can be trusted to find long-lived gravitational-wave signals from isolated neutron stars in real detector data. The authors establish that a single CNN trained on signal/noise maps in one 150 Hz band generalizes to other frequency bands and to signal shapes it never saw, with detection efficiencies comparable to the established Generalized FrequencyHough algorithm and at orders-of-magnitude lower computational cost. They show that false-alarm probability can be tuned by choosing a frequency-dependent threshold on the network's output, and that only a modest amount of training data is required. Using these results, they run the first real-data machine-learning search for an isolated neutron star signal, looking for a remnant of the neutron-star merger GW170817, and find no significant candidate.","feed_headline":"One CNN trained in one band finds neutron-star signals in any band","feed_subtitle":"Matches the standard search's sensitivity, runs orders of magnitude faster, and sees signals the standard method misses.","key_machinery":"The load-bearing object is the reduced time/frequency map: each 2000 s stretch of whitened, line-cleaned detector data is converted into a peakmap and then downsampled by taking the maximum value in each 16x16 block above a 2.5-sigma threshold. These maps are fed to a CNN with five convolution blocks (zero-padding, convolution, ReLU, max-pooling) and two fully connected layers ending in a softmax that outputs a signal probability. The GFH, the comparison method, maps power-law curves in the time/frequency plane to lines in a parameter plane of initial frequency and spindown. The persistency veto, which removes frequency bins that persistently show excess peaks, is what keeps the CNNs from treating noise lines as signals.","core_discovery":"The central discovery is that CNNs can be a practical search tool for long transients from neutron stars, not just a laboratory curiosity. Trained on roughly 20,000 injections plus about 2,000 real noise maps from LIGO Livingston in a single 150 Hz band, the network distinguishes signal from noise in any band across 100-1900 Hz, including bands and braking-index behaviors it never saw. Its sensitivity curves match the Generalized FrequencyHough, and it detects signals with time-varying braking index that the GFH cannot recover because they violate the power-law model. The false alarm probability is controlled by selecting a threshold on the softmax output; the required threshold varies by band. Applied to one week of O2 data after GW170817, the CNN plus GFH pipeline found no significant candidate and produced upper limits consistent with the previous search.","pith_inferences":["If the transferability holds across observing runs, a single network pretrained on O2 data could serve for O3/O4 searches with only per-band threshold recalibration, a direct extension the paper leaves for future work.","The CNN's extreme sensitivity to spectral lines could be repurposed: the same network, probed or inverted, might serve as a line-detection and diagnostics tool for detector characterization.","Combining two or three detectors' maps into a single image (RGB synthesis), which the authors mention as future work, should lower false alarms because noise lines are largely detector-specific; this is a natural testable next step.","Extending the CNN to directly estimate the braking index and its time derivative, which the authors say they are working on, would remove the need for a GFH parameter loop entirely and make the pipeline fully machine-learning driven."],"forward_implications":["A CNN can act as a fast trigger generator: it flags time/frequency maps that deserve a GFH follow-up, cutting the compute from minutes per map to microseconds.","Because one trained network covers many bands and signal morphologies, a search can be set up with a few thousand injections rather than exhaustive parameter-space coverage.","The method opens a new observational window: signals with time-varying braking index, invisible to power-law searches, become detectable.","Low-latency searches become feasible: training on about 23 days of noise suffices, so the network can be retrained on fresh data quickly without discarding much data.","False alarms are manageable in practice: with per-band thresholds around 0.7-0.95, the false alarm probability can be held near 1%, which is acceptable for a trigger stage followed by a more precise method."],"supporting_citations":[{"why":"Defines the Generalized FrequencyHough baseline that maps power-law signals to lines; the CNNs are compared against its sensitivity and speed.","marker":"[30]"},{"why":"Supplies the CNN architecture and the idea of classifying time/frequency maps, originally for core-collapse supernovae.","marker":"[11]"},{"why":"Provides the resolution-reduction procedure for time/frequency maps and earlier ANN/Alexnet comparisons used here.","marker":"[32]"},{"why":"The previous GW170817 remnant search whose upper limits and search configuration the authors compare against.","marker":"[7]"},{"why":"Establishes the peakmap construction and frequency-Hough detection statistic used in the follow-up.","marker":"[12]"},{"why":"Describes the Short FFT Database format used to build the time/frequency maps.","marker":"[14]"},{"why":"Identifies the known noise lines that must be zeroed before the CNN sees a map, essential for false alarm control.","marker":"[18]"}],"fun_headline_variants":["Single-band CNN detects neutron-star signals in any band","CNN matches standard search yet is orders of magnitude faster","First machine-learning search for isolated neutron star waves","One-band training, broad detection: CNN for neutron-star signals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"A CNN trained on about 2,000 Livingston noise maps from one observing run in a single 150 Hz band transfers to other frequency bands, other time periods, and the Hanford detector, with only the output threshold recalibrated per band.","fun_headline_variants_meta":{"raw":{"variants":["Single-band CNN detects neutron-star signals in any band","CNN matches standard search yet is orders of magnitude faster","First machine-learning search for isolated neutron star waves","One-band training, broad detection: CNN for neutron-star signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1775,"prompt_tokens":1031,"completion_tokens":744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":680}},"tokens_in":647,"tokens_out":744,"duration_ms":8252,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:55:28.476985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same network on one 150 Hz band of O2 Livingston data, then run it on a large sample of Hanford noise maps from a different epoch (or on O3/O4 data) without retraining; if the false alarm probability at the calibrated threshold is far above 1% even after per-band recalibration, or if the detection efficiency on time-varying-braking-index injections falls below the GFH's, the transferability claim is refuted.","supporting_citations":[{"cited_title":"Codes to create sfdbs and o2 used time segments","cited_arxiv_id":null,"evidence_quote":"Defines the Generalized FrequencyHough baseline that maps power-law signals to lines; the CNNs are compared against its sensitivity and speed."},{"cited_title":"The CNNs are sensitive to time-varying brak- ing indices as well, even though these upper limits are done on fixed braking indices","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN architecture and the idea of classifying time/frequency maps, originally for core-collapse supernovae."},{"cited_title":"Binary neutron star mergers: a review of Einsteins richest laboratory","cited_arxiv_id":null,"evidence_quote":"Provides the resolution-reduction procedure for time/frequency maps and earlier ANN/Alexnet comparisons used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The previous GW170817 remnant search whose upper limits and search configuration the authors compare against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the peakmap construction and frequency-Hough detection statistic used in the follow-up."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the Short FFT Database format used to build the time/frequency maps."},{"cited_title":"Advanced ligo","cited_arxiv_id":null,"evidence_quote":"Identifies the known noise lines that must be zeroed before the CNN sees a map, essential for false alarm control."}],"review_version":1}