Detecting short gravitational-wave bursts in real LIGO noise

Frequentist, Bayesian and machine-learning analysis of simulated bursts injected into GWOSC O1 data

Frequentist, Bayesian and machine-learning analysis of simulated bursts (sine-Gaussian, Gaussian pulse, linear chirp) injected into real LIGO-Hanford (H1) strain from the first observing run (O1), public data from GWOSC. Final project for the Statistics and Data Analysis course (Università di Milano-Bicocca, A.Y. 2025–2026), Giacomo Moretti. The complete report is Docs/report/main.pdf.

The organisation follows the earlier projects on LHCb open data and on AFM force curves: one shared data set, one notebook per method, one common set of figures of merit, a report that says where each number comes from.

Research question. At a fixed false-alarm probability on real LIGO noise, how do a frequentist matched-filter bank, a Bayesian analysis of the same likelihood and machine-learning detectors compare in detection efficiency, in the quality of their uncertainty statements, and in robustness to the non-Gaussian noise and to signal morphologies they were not designed for?

Repository layout

Code_General/
  00_waveform_exploration.ipynb          exploratory prototype: GW150914, PSD, whitening, Q-transform, first injections
  01_data_whitening_validation.ipynb     8 O1 chunks, PSD, whitening, 1 s windows, Gaussianity / stationarity tests, split by chunk
  02_waveforms_and_injections.ipynb      exact analytic burst transforms, optimal SNR, validation of the injection chain
  03_frequentist_analysis.ipynb          matched-filter bank, calibration, efficiency / ROC, MLE + Fisher, toy MC, systematics
  04_bayesian_analysis.ipynb             exact grid posterior + emcee, Bayes factors, prior dependence, SBC, credible vs confidence
  05_ml_detection.ipynb                  features + Ridge/Lasso/RF/GBT, 1-D CNN, ensemble, MC dropout, autoencoder, calibration
  06_sequence_models_hypothesis_testing.ipynb   Transformer / GRU detectors, masked-patch pre-training, 4-class morphology vs frequentist and Bayesian model comparison
  07_tuning_sequence_models.ipynb        stress test of the Transformer-vs-CNN result: fresh injections, learning rate, patch size, convolutional stem
  08_realistic_inspiral_signals.ipynb    binary-black-hole (IMRPhenomD) injections: chirp-mass mapping of the linear chirp, IMR template bank, detection of realistic signals
  09_two_detector_coincidence.ipynb      H1+L1: coincident statistic, time-slide background down to FAP 1e-6, zero-lag check, efficiency vs single detector
  10_open_set_goodness_of_fit.ipynb      flagging signals that belong to none of the hypotheses: matched-filter residual, network open-set scores
  11_bootstrap_error_bars.ipynb          bootstrap errors on rho50/rho90 and paired model comparisons (nb 06-07 detectors)
  12_threshold_convention.ipynb          bank rho50 with thresholds from train / validation / pooled / test background; fixes the report convention
  13_real_events.ipynb                   GW150914 / GW151012 / GW151226 in H1+L1 through the unchanged banks and coincident statistic
  14_continuous_search.ipynb             mini continuous search on 2.25 h of H1+L1: sliding windows, clustering, time-slide FAR in yr^-1, candidate list, injection recovery
  15_cnn_vs_bank_ablation.ipynb          why the CNN beats the bank: Gaussian vs real noise, 640-template bank, amplitude veto
  16_chi2_veto_ranking.ipynb             Allen chi2 and new SNR for the bank; logistic / boosted ranking statistics
  17_hard_negative_mining.ipynb          fine-tuning the CNN with hard negatives vs control (no measurable gain)
  18_glitch_families.ipynb               unsupervised families of high-statistic noise windows; which ones alarm bank / chi2 / CNN
  19_unfolding.ipynb                    unfolding of the total-mass distribution: inversion, IBU, OmniFold on IMR bank output
  20_prior_mismatch_base_rate.ipynb     CNN and bank outside the training prior; PPV of an alarm vs astrophysical base rate
  21_injected_glitches_robust_bayes.ipynb  labelled injected glitches (blip, whistle, arch) and a glitch-aware Bayes factor
  22_imr_parameter_estimation.ipynb      Mc, M, q, merger time of IMR signals: bank MLE, grid posterior, CNN regression with Gaussian head
  23_background_rigour.ipynb             leave-one-chunk-out epoch dependence, time-local background, GPD tail extrapolation, Feldman-Cousins rate limits
  24_rare_event_mc.ipynb                 subset simulation of the Gaussian-noise tail of the bank statistic down to FAP 1e-10
  25_robustness_conformal_interpretability.ipynb  systematics, conformal thresholds/sets, saliency, latency, FGSM attacks on the CNN
  26_scaling_augmentation_adaptation.ipynb  CNN scaling with training windows, augmentation, label noise, H1->L1 transfer
  27_bank_design_local_whitening_psd.ipynb  bank design (number/placement of templates), time-local whitening, PSD-estimation uncertainty
  28_physical_units_range.ipynb          h_rss benchmarks, ringdowns/cusps/white-noise bursts, physical BBH injections: sensitive range and volume
  29_precession_higher_modes.ipynb       IMRPhenomPv2 / XPHM libraries (precession, higher modes) vs the detectors
  30_seeds_paired_tests.ipynb            5 CNN / 3 hybrid / 3 Transformer seeds, seed scatter, permutation and Welch tests, 5-CNN ensemble
  31_coherent_network.ipynb              coherent (lambda_1) vs incoherent vs Fisher statistics, physical network injections (antenna patterns, delay), 10 background chunks
  32_bayesian_extensions.ipynb           exact-model MCMC (emcee), two-detector signal-vs-glitch evidence, SBC of the grid posterior, hierarchical population inference
  33_unsupervised_flow_vae.ipynb         RealNVP flow and VAE on noise only as anomaly detectors
  34_other_architectures.ipynb           ResNet-1D, TCN, 2-D CNN on spectrograms, EdgeConv graph network
  35_data_quality_flags.ipynb            GWOSC CAT1/CAT2 flags as vetoes: dead time vs glitch floor
  36_o2_o3_generalisation.ipynb          unchanged pipeline on O2, O3a, O3b chunks (download_runs.py)
  37_limits_low_mass_bns_nsbh_ecc.ipynb  SNR retained by the 1 s / 30-500 Hz window (low masses, BNS), NSBH and eccentric libraries vs the detectors
  38_virgo_snr_budget.ipynb              SNR budget and sensitive volume of HLV against HL (O2 August 2017); download_virgo.py
  39_transformer_tuning_round2.ipynb     Transformer depth / width / 30 epochs (variants tf16_d6/w128/long_fresh of seq_models.py)
  pe_tools.py, network_tools.py          likelihood/sampler and network injection helpers of notebook 32
  download_coinc_extra.py                6 more H1+L1 chunks -> results/coinc_chunks_extra.json
  precession_tools.py                    waveform generation of notebook 29 (needs lalsuite to build)
  glitch_tools.py                        glitch shapes, injection and GlitchBank of notebook 21
  download_events.py                     downloads the three event chunks (NaN gaps handled in nb 13)
  make_summary_figures.py                summary rho50 figure and pipeline diagram of the report (fig13_*)
  gwosc_tools.py                         functions shared by notebooks 01-10
  seq_models.py                          models, dataset and resumable training loop of notebooks 07-08 (run it to train the variants in parallel processes)
  imr_tools.py                           IMRPhenomD signal libraries (train and held-out test) and the IMR template bank (needs lalsuite to build the libraries)
  open_set.py                            family fits and goodness-of-fit scores of notebook 10
  download_data.py                       downloads the 8 raw H1 chunks (cache/raw/) and writes results/chunks.json
  download_coinc_data.py                 downloads the 6 H1+L1 chunks of notebook 09 and writes results/coinc_chunks.json
  results/                               summary JSON of every notebook (tracked)
  figures/                               figures of every notebook, PNG + PDF (tracked)
  cache/                                 raw strain, whitened windows, scans, trained models (NOT tracked, regenerable)
Docs/report/                             the report (main.tex, sections/, biblio.bib, figures/, main.pdf)
environment.yml                          conda environment of all the code

Data

Eight chunks of 4096 s of H1 strain at 4096 Hz from O1 (Sept. 2015 – Jan. 2016), inside the H1 science segments and at least 2000 s from every catalogued event, selected and downloaded by Code_General/download_data.py (about 0.5 GB in float32, needs internet access to gwosc.org). The split is by chunk (train 0,2,4,6 · validation 1,5 · test 3,7), never by window.

Tests and reproducibility

python -m pytest Code_General/tests -q (19 checks: whitening, exact SNR of injections, bank, oracle, bootstrap, χ², splits, glitch/PE/network helpers; synthetic data only); bash Code_General/run_all.sh regenerates data, libraries and every notebook in order; python Code_General/make_checksums.py --check verifies the raw strain files against results/data_checksums.json.

Environment

conda env create -f environment.yml
conda activate gwosc-burst
python -m ipykernel install --user --name cern

All notebooks use the kernel cern. Then, from Code_General/:

python download_data.py
python download_coinc_data.py        # only for notebook 09 (H1+L1 chunks)
python imr_tools.py                  # only for notebooks 08-10 (IMRPhenomD libraries, needs lalsuite: pip install lalsuite)
jupyter nbconvert --to notebook --execute --inplace --ExecutePreprocessor.kernel_name=cern NN_*.ipynb      # NN = 01 ... 10, in order

Run times on a 10-core laptop on AC power: 01 ≈ 1 min, 02 ≈ 1 min, 03 ≈ 10–15 min, 04 ≈ 20 min, 05 ≈ 30–40 min with training from scratch (a few minutes with the saved models), 06 ≈ 1–2 h, 07 ≈ 3 h of training on four processes, 08 ≈ 10 min plus 20 min of training, 09 ≈ 10 min, 10 ≈ 10 min. All random numbers use fixed seeds; long blocks cache their output and resume on re-run. Notebooks 03–05 read the whitened windows cached by notebook 01 (rebuilt automatically from cache/raw/ if missing); notebook 04 reads results/03_summary.json.

The report is compiled with tectonic -X compile main.tex from Docs/report/.

Pipeline

Notebook	Content	Output
01	PSD (median Welch) of each chunk, whitening, 30–500 Hz band, 1 s windows; tests of the Gaussian-noise assumption	results/01_summary.json
02	sine-Gaussian / pulse / chirp as exact Fourier transforms, analytic optimal SNR, injections as whitened bins, validation (non-central χ²)	results/02_summary.json
03	160-template bank, profile-likelihood statistic, look-elsewhere (N_eff), calibration on unseen epochs, efficiency vs oracle and excess power, iminuit MLE vs Fisher, Rician toy MC with Neyman/Feldman–Cousins belts, Barlow systematics	results/03_summary.json
04	analytic amplitude marginalisation, exact grid posterior, emcee cross-check, Bayes factors and prior dependence, simulation-based calibration, credible vs confidence intervals	results/04_summary.json
05	22 features with Ridge/Lasso/RF/GBT (bias–variance, importance, PCA), 1-D CNN, 5-member ensemble, MC dropout, autoencoder, calibration, comparison on the same injections	results/05_summary.json
06	patch Transformer and GRU against the CNN (efficiency at FAP 10⁻², 10⁻³, glitches, attention), masked-patch pre-training and label efficiency, 4-class (noise / sg / pulse / chirp) detectors against frequentist and Bayesian model comparison, an out-of-family burst	results/06_summary.json, figures/fig06_*
07	tuning of the sequence models, one lever at a time (fresh injections each epoch, learning rate, patch size, convolutional stem); diagnosis by training vs validation loss; same test injections and thresholds as 05–06; four-class networks with fresh injections	results/07_summary.json, figures/fig07_*
08	mapping of the linear chirp to a binary chirp mass; library of IMRPhenomD signals (4000 train + 2000 held-out); IMR matched-filter bank and oracle; detection of realistic signals by every method; mass dependence	results/08_summary.json, figures/fig08_*
09	H1+L1 chunks; Fisher combination of per-detector p-values; time slides ($6.6\cdot10^7$ pairs) down to FAP 10⁻⁶; zero-lag calibration; network-SNR efficiency, coincident against single detector	results/09_summary.json, figures/fig09_*
10	goodness-of-fit residual of the best-fit family, maximum softmax, energy score and Mahalanobis distance of the networks; stochastic bursts and IMR signals as out-of-family	results/10_summary.json, figures/fig10_*
11	bootstrap of ρ₅₀/ρ₉₀ (resampling background and injections), paired differences between detectors	results/11_summary.json
12	which background sets the threshold: bank ρ₅₀ with train / validation / pooled / test thresholds	results/12_summary.json
13	the three O1 events: IMR / sine-Gaussian bank SNR, best-fit mass, H–L delay, p-values, coincident FAP	results/13_summary.json
14	sliding-window search with trigger clustering; FAR in yr⁻¹ from time slides of the same chunks (2.08 yr); zero-lag 62 vs 63.1 expected; injection recovery	results/14_summary.json
15	ablation of the CNN-vs-bank gap: real vs Gaussian noise (bank 5.38 vs CNN 5.63 in Gaussian), richer bank (no gain), amplitude veto (helps at 1e-3 only)	results/15_summary.json
16	Allen χ² and re-weighted SNR; learned ranking statistics: at FAP 10⁻³ the χ² veto takes the bank from 7.76 to 6.35 (CNN 5.48), no gain at 10⁻²	results/16_summary.json
17	hard-negative mining vs control: mining − control = +0.10 (−0.15, +0.33) at FAP 10⁻², no effect	results/17_summary.json
18	clustering of the 4 % highest-statistic windows: no real structure except a small extreme family (real glitches) that the χ² test and the CNN reject	results/18_summary.json
19	unfolding of the detected total-mass spectrum: inversion unstable, IBU 3–5 iterations best, OmniFold close but biased; bootstrap spread < bias	results/19_summary.json
20	prior mismatch: bank robust, CNN fails at f0 400–500 Hz (ρ₅₀ 8.6 vs 5.1) and on loud short pulses (eff 0.02 at ρ=40); base rate: PPV 0.001 (FAP 10⁻², 1/day) vs 0.90 (FAP 10⁻⁶)	results/20_summary.json
21	labelled glitches: bank/Bayes/CNN alarm on 97 % of blips at ρ=8; glitch-aware Bayes factor cuts blip alarms to 0.35 and whistles to 0.04 at a cost of 0.08 in signal efficiency; χ² is not a glitch veto	results/21_summary.json
22	IMR parameter estimation: bank/posterior ≈7 % error in Mc and 1 ms in time, CNN 10–17 % and 15–85 ms; grid posterior overconfident (68 % interval covers 9–32 %), CNN calibrated (0.65–0.82)	results/22_summary.json
23	background rigour: LOO observed/nominal FAP 0.3–0.9 (chunk 2: 4.3), local thresholds cut the spread 0.013 → 0.001, GPD ξ=0.91 unsafe below 1e-3, FC 90 % limit 13 signals	results/23_summary.json
24	subset simulation: bank threshold agrees with 1−(1−e^{−λ/2})^{N_eff} (N_eff=6794) to 0.1–2 down to FAP 1e-10	results/24_summary.json
25	CNN robust to PSD estimate and calibration but not to a 60 Hz line (FAP 0.4 % → 17 %) nor FGSM (eff 1.0 → 0.33 at 0.1σ); conformal guarantee fails across epochs; CNN 20× faster than the bank per batch	results/25_summary.json
26	CNN is data-limited (ρ₅₀ 6.5→4.5 from 10³ to 1.6·10⁴ windows, no saturation); flip/shift augmentation no gain; H1-trained CNN transfers to L1 (5.55) while scratch on 2 L1 chunks fails (8.45)	results/26_summary.json
27	bank design: ≥80 templates all equal (fitting factor 0.97→0.996, ρ₅₀ 5.6), random = regular; time-local whitening does not cure epoch dependence; PSD from ≥256 s is negligible, 32 s inflates FAP ×1.5–2.3	results/27_summary.json
28	h_rss50 ≈ 2.5·10⁻²³ Hz⁻¹ᐟ² (H1, optimal orientation); CNN ≥ bank on unseen shapes except broadband bursts; BBH range 1.3 Gpc (IMR bank) / 1.6 Gpc (CNN with IMR), Euclidean	results/28_summary.json
29	precessing (Pv2) and higher-mode (XPHM) signals cost ≤0.3 in ρ₅₀ for all detectors in the 30–100 M☉ window (fitting factor 0.97 → 0.93)	results/29_summary.json
30	seed scatter of the fresh CNN is 0.40 at FAP 10⁻² (0.03 at 10⁻³); architecture gaps (hybrid +1.0, Transformer +2.0) significant; 5-CNN ensemble 3.79 / 5.32	results/30_summary.json
31	coherent λ₁ does not beat the Fisher statistic (ρ_net50 10.2 vs 9.4 at FAP 10⁻⁴); with physical antenna patterns ρ_net50 at FAP 10⁻⁶ is ≈15, not 11	results/31_summary.json
32	exact-IMRPhenomD MCMC calibrated (68 % coverage 0.6–0.8) with 5–8 % error in Mc; two-detector evidence odds no better than min ρ² (AUC 0.85 at norm 8); SBC KS 0.05 (f0, Q), finer grid no help; hierarchical fit recovers (45, 5, 0.30) as (44.8, 4.7, 0.30) while the naive fit gives μ=49.6	results/32_summary.json
33	noise-only models: VAE local error ρ₅₀ 6.7/6.5/8.5 (sg/pulse/chirp, IMR 7.7), flow 7.5/7.2/9.3; ELBO useless; alarms only 20–26 % shared with the bank	results/33_summary.json
34	ResNet-1D 5.01, TCN 5.09, 2-D CNN 5.20, EdgeConv graph network 5.48 (FAP 1e-2, one seed): none beats the 1-D CNN (4.31 ± 0.40)	results/34_summary.json
35	CAT2 veto: 0.9 % dead time removes 40 % of the loudest bank windows, LOO FAP spread 0.74 → 0.15, threshold at 1e-3 110 → 57; no effect on the CNN	results/35_summary.json
36	O2/O3 transfer: O1 thresholds give FAP 0.9–2.1 % at nominal 1 %; ρ₅₀ unchanged (bank 5.55, CNN 3.7–3.9 with O1 threshold); O3 glitches (kurtosis 10⁵–10⁶) make the bank threshold at 10⁻³ unusable (>16), CNN robust	results/36_summary.json
37	1 s/30–500 Hz keeps 93–97 % of the SNR for ≥10 M☉ components but 49 % at 3+3 and 8.7 % for a BNS; NSBH costs the IMR bank 5.5 in ρ₅₀ (CNN with IMR: 0.9); eccentricity ≤ 0.9	results/37_summary.json
38	adding Virgo (O2): V1 median ASD 3.7× H1; 9.6 % of the network SNR²; ideal volume +6.3 % (HLV/HL); no 3-detector search	results/38_summary.json
39	Transformer: depth 6 → 6.23, width 128 → 6.11, 30 epochs → 5.81 (FAP 1e-2; baseline 6.29, CNN 4.31)	results/39_summary.json

Main results

All SNRs are optimal SNRs; the false-alarm probability (FAP) is per one-second window (of H1 alone, except in notebook 09).

Statistic (FAP 10⁻², validation background)	sine-Gaussian ρ₅₀ / ρ₉₀	pulse ρ₅₀ / ρ₉₀	chirp ρ₅₀ / ρ₉₀
matched-filter bank (frequentist)	4.87 / 6.53	5.34 / 6.82	6.28 / 8.21
excess power	6.54 / 9.10	5.77 / 7.54	10.10 / 12.21
gradient boosting (22 features)	5.61 / 7.30	5.70 / 7.24	6.06 / 7.73
CNN ensemble (5)	5.05 / 7.05	4.21 / 5.90	5.28 / 6.92
CNN trained on sine-Gaussians only	5.09 / 6.87	6.50 / 12.93	6.80 / 10.82
autoencoder (noise only), local max	6.89 / 8.80	6.75 / 8.53	8.30 / 11.08

(ρ₅₀, ρ₉₀: SNR at 50 % and 90 % efficiency on test-chunk injections.)

• Real noise. Gaussian in the bulk (N(0,1) down to ~10⁻⁴), with a glitch floor of a few 10⁻³ per second and strong epoch-to-epoch variations. It sets the price of a low FAP: lowering it from 10⁻² to 10⁻³ moves ρ₅₀ of the bank from ≈5.5 to ≈11 with the training-background threshold (4.9 → 6.8 with the validation one: notebook 12), and the threshold calibrated on the training chunks gives an observed FAP 4.5× below nominal on the test chunks.
• Frequentist. The look-elsewhere effect is quantified by N_eff ≈ 6800 independent trials; the oracle with known parameters needs ρ₅₀ = 3.0 against ≈5 for the bank; the MLE saturates the Fisher bound for ρ ≳ 10; the profile-likelihood interval of the amplitude covers 39 % at s = 0 (Neyman / Feldman–Cousins cover 68 % everywhere); the dominant systematic is the PSD estimator (+0.025 on ε(ρ=8)).
• Bayesian. The grid posterior is validated against a closed form, emcee and simulation-based calibration (uniform ranks in Gaussian and in real noise); the Bayes factor ranks windows like the frequentist statistic (same efficiency within 0.01), depends on the prior as −2 ln σ_ρ and exceeds 100 on 0.2–0.9 % of pure-noise windows because of glitches; credible and confidence intervals agree for ρ ≳ 3.
• Machine learning. The CNN ensemble matches the bank on sine-Gaussians and is ≈1 SNR unit better on the pulses and chirps it was trained on; a CNN trained on one morphology loses efficiency on the others; an autoencoder trained on noise adds nothing to an energy statistic. Calibration is good (ECE ≲ 3 %), the ensemble is better than a single network in every proper score, and rejecting the 20 % most uncertain windows raises the AUC from 0.992 to 0.9997.
• Sequence models and hypothesis testing. With the same data and budget a patch Transformer (114 k parameters) and a bidirectional GRU are worse than the 17.8 k-parameter CNN (ρ₅₀ at FAP 10⁻² on sine-Gaussian / pulse / chirp: CNN ensemble 5.05 / 4.21 / 5.28, GRU 6.27 / 5.16 / 6.36, Transformer 6.38 / 4.88 / 7.35) and do not reject glitches better; masked-patch pre-training on whitened noise learns almost nothing (it helps only at ≈1000 labels per class). For the morphology, Bayesian model comparison is the best classifier (3-way accuracy 0.97 at ρ = 8) among the networks trained on the static set, the CNN is close at high SNR, the Transformer last (but see notebook 07: with fresh injections the CNN ties with it, 0.977 at ρ = 8); no family decision flags a burst that belongs to none of the three families.
• Tuning the sequence models (notebook 07). The patch Transformers under-fit (training ≈ validation loss ≈ 0.3, against 0.15–0.17 for the CNN) and no single lever fixes them: fresh injections, a 3× learning rate, patches of 8 samples and the 25-epoch run all leave mean ρ₅₀ at 6.3–6.6 (FAP 10⁻²). A convolutional stem followed by two Transformer blocks is the lever that works (5.7 static, 5.4 with fresh injections) but does not overtake the CNN. The biggest effect is on the CNN itself: fresh injections at every epoch cut its mean ρ₅₀ from 5.1–5.3 to 4.6 (FAP 10⁻²) and from 6.2–6.3 to 5.65 (FAP 10⁻³), better than the 5-member static ensemble of notebook 05 (4.85, 5.79); the CNN of notebooks 05–06 was data-limited. Two seeds for the main variants, ≤ 12 epochs for the sequence models. With fresh injections the 4-class CNN reaches 0.977 three-way accuracy at ρ = 8, as good as or slightly better than the Bayesian model comparison (0.969).
• Realistic signals (notebook 08). The sweep rate of the linear chirp corresponds to a binary of 2–30 M☉ but without acceleration or merger. On IMRPhenomD binary-black-hole signals (held-out waveforms, total mass 30–100 M☉) the sine-Gaussian bank needs ρ₅₀ = 7.4 (FAP 10⁻²) against 5.4 for an IMR template bank and 3.0 for the oracle; the CNNs trained on the phenomenological morphologies generalise (5.5–6.2), the hybrid Transformer does not (7.6); a CNN trained with IMR signals reaches 4.45 (efficiency 0.98 at ρ = 8, against 0.91 for the bank), at no cost on the three phenomenological morphologies.
• Two detectors (notebook 09). With H1+L1, a Fisher combination of per-detector p-values and time slides the false-alarm floor moves from 10⁻³ to 10⁻⁶ per window; the noises of the two detectors are independent (zero-lag against slides of the same chunks: ratio 0.6–1.1). Network-SNR ρ₅₀ at FAP 10⁻² / 10⁻³ / 10⁻⁴ / 10⁻⁵ / 10⁻⁶ on binary-black-hole injections: CNN 4.1 / 6.8 / 8.2 / 9.7 / 11.0, IMR bank 7.4 / 7.6 / 8.5 / 9.7 / 11.0, sine-Gaussian bank 12.4 / 12.9 / 14.3 / 16.6 / 20.4. A plain sum of the statistics is dominated by glitches of one detector and is not used.
• Open set (notebook 10). The residual energy after subtracting the best-fit family flags stochastic bursts (AUROC 0.94 at ρ = 8, ≥ 0.998 at ρ ≥ 12); the energy score of the 4-class CNN flags binary-black-hole chirps (0.93 / 0.98 / 0.99 at ρ = 8 / 12 / 16); the Mahalanobis distance of the features does not work (AUROC falls with SNR).

Extensions (notebooks 11–39)

Second round of work, one backlog idea per result (see the table above and CLAUDE.md for the numbers):

• Rigour. Bootstrap errors on ρ₅₀ and paired tests (nb 11, 30); seed scatter of the fresh CNN is 0.40 at FAP 10⁻² (5 seeds); convention = validation background for every thresholded comparison (nb 12); leave-one-chunk-out calibration, time-local background, GPD tail, Feldman–Cousins limits (nb 23); subset simulation to FAP 10⁻¹⁰ (nb 24).
• Why the CNN wins and where it fails. Ablation (nb 15), χ² and ranking statistics (nb 16), hard negatives and glitch families (nb 17–18), prior mismatch and base rate (nb 20), labelled glitches and glitch-aware Bayes factor (nb 21), robustness, conformal sets, saliency, latency, FGSM (nb 25), scaling/augmentation/H1→L1 transfer (nb 26), other architectures (nb 34), Transformer tuning (nb 39), noise-only flow and VAE (nb 33).
• Physics and data. Real events (nb 13), continuous search (nb 14), physical units and range (nb 28), precession/higher modes (nb 29), limits for low masses/BNS/NSBH/eccentricity (nb 37), coherent statistic with physical network injections and 10 background chunks (nb 31), Virgo SNR budget (nb 38), GWOSC data-quality flags (nb 35), O2 and O3 (nb 36).
• Parameters and populations. IMR parameter estimation with CNN regression, grid posterior and exact-model MCMC (nb 22, 32), unfolding and hierarchical population inference (nb 19, 32), SBC (nb 32), bank design, local whitening and PSD uncertainty (nb 27).
• Negative or limited results worth keeping: hard-negative mining, flip/shift augmentation, coherent λ₁ without sky scan, time-local whitening, a denser bank, and extra Transformer depth/width give no measurable gain; the CNN is brittle outside its training ranges.

Limits

• Detectors. Most notebooks use H1 alone; notebooks 09, 14 and 31 add L1 (incoherent combination and a peak-based coherent statistic, no sky scan, no χ² vetoes; 6–12 chunks), notebook 38 only an SNR budget for Virgo. FAPs are per window, not astrophysical rates.
• Injections at a given SNR. Converted into distances only in notebook 28 (Euclidean, single H1); the three morphologies are phenomenological; the main IMRPhenomD library has aligned spins, total masses 30–100 M☉; precessing, higher-mode, NSBH and eccentric signals are tested in notebooks 29 and 37, BNS and masses < 10 M☉ are out of reach of one-second windows.
• Background statistics. 1.6·10⁴ training windows limit the lowest usable FAP to ~10⁻³; the 10⁻³ efficiencies are uncertain.
• Eight chunks. The epoch dependence is measured with two test chunks.
• Grid resolution. The Bayesian grid under-resolves f₀ and t₀ for ρ ≳ 15 (SBC coverage of those two parameters is 0.50–0.60, not the model).
• ML numbers are efficiencies for the training injection prior (the CNN fails outside it, notebook 20); most variants have one seed and the seed scatter at FAP 10⁻² is 0.4.

Conventions

Notebooks, comments and documentation are in English. Every notebook states at the beginning what it does and what it does not do, and ends with the output produced and the next notebook. cache/ is a regenerable cache; results/ is the source of the numbers quoted here.

License

Released under the MIT License. The LIGO strain data are public (GWOSC) and are subject to their own data-use terms.