Deep-learning searches of public LIGO/Virgo data for black-hole signatures.
Three real-data gravitational-wave sub-projects — a subsolar / primordial-black-hole merger search, a post-merger echo search, and ringdown spectroscopy (a no-hair test) — plus the black-hole physics notes that started it all. Everything is measured on real O3a/O4 detector noise: sensitivity comes from injections, significance from a measured background, and null results are treated as results.
A computer engineer's deep dive into gravitational-wave astronomy. Built carefully: every load-bearing claim is checked against the literature, every sensitivity number comes from injections into real noise, and a regression gate (
./verify.sh, 36 checks) asserts the headline results never silently change.
Below ~1 M☉ no star can collapse into a black hole, so a subsolar merger would be a smoking gun for primordial black holes (a dark-matter candidate). A CNN trigger on real O3a H1 noise reaches 41–45 % of the ideal matched-filter sensitive distance at a zero-false-alarm threshold — for minutes-long signals, the regime with no published ML search.
- Coincidence is the win. Requiring H1×L1 agreement collapses the noise floor: +1.37× sensitive distance (~2.5× volume) over a single detector at matched FAR, and it stays robust down to a 1/year false-alarm rate. A learned coincidence statistic (256-d embeddings → a small head that asks whether the detectors agree) beats a plain sum by a further significant +0.02–0.05, leakage-free and bootstrap-verified.
- The real matched-filter benchmark — the question the whole project circled. A realizable semi-coherent dense-bank matched filter (1,619 templates at 0.1 % chirp-mass spacing, built + run on a laptop GPU) scores 0.489 vs the CNN's 0.472 on identical injections — a statistical tie. A single CNN forward pass matches a 1,600-template bank. Both sit far below the true-template oracle (0.72), so the dominant loss is template-bank mismatch, not learned-vs-matched-filter. The fully-coherent ideal is genuinely intractable (megatemplate scale, consistent with LVK's real 3,452,006-template O4 subsolar bank).
- Honest negatives that sharpened the picture: score-aggregation across windows (two rungs); a learned semi-coherent classifier (caps ~0.70 AUC); and Virgo — a triple H1×L1×V1 coincidence doesn't help subsolar (V1 is too insensitive at these masses to carry signal).
- Self-supervised pretraining on unlabeled noise is a data-wall win: +0.28 sensitive-distance fraction at scarce labels.
If black-hole horizons have quantum structure, a merger might be followed by faint, repeating "echoes." A full pipeline (injection-calibrated, background-defined p-values) searches the post-ringdown of GW150914 / GW151012 / GW151226 / GW250114.
- A small ML noise-model scorer gives a modest but real ~1.2× sensitivity edge over the classical comb through the production path — band-honest, family-robust, periodicity-specific.
- The echo spacing is predicted from first principles: a Kerr-tortoise round-trip Δt(M,χ) that reproduces the Abedi Table-I values to <2 % (and caught a wrong hard-coded value in the repo).
- Every on-source event is a clean null under both the ML scorer and the comb, verified against an independent ±hour background. The novel angles — conditioning the echo search on the event's own ringdown mass, and stacking across events — are built and gated.
Fit the post-merger ringdown tones (quasinormal modes) and test whether they imply a consistent mass and spin — the no-hair theorem.
- An amortized simulation-based-inference (NPE) model with the ringdown start-time marginalized by construction measures the deviation δ 2.6× tighter than the classical fit, calibration-certified (held-out coverage 0.91). On GW250114 (the loudest event ever recorded): δ = −0.16, consistent with a Kerr black hole.
- Cross-validated against the field-standard pipeline (the Isi/Farr
ringdownpackage): it independently detects the GW250114 overtone (A₂₂₁ bounded from zero) where our simplified machinery could not, and its (M, χ) posterior nests inside our NPE's — the first independent check of the whole SBI arc. - A start-time sweep with that package reproduces the early-time systematic our NPE carries: peak-cropped fits are biased ~+10 % in mass, decaying as the start moves later.
| Search | Result | Status |
|---|---|---|
| Subsolar CNN (single detector) | 41–45 % of ideal-MF distance · 0 false alarms in 6.8 h real noise | ✓ gated |
| Subsolar H1×L1 coincidence | +1.37× sensitive distance, FAR-robust to 1/year; learned statistic beats sum | ✓ gated |
| Subsolar — CNN vs real matched filter | tie (0.489 vs 0.472 on identical injections); both bank-mismatch-limited | ✓ gated |
| Subsolar triple H1×L1×V1 | Virgo does not help at subsolar masses (honest negative) | ✓ gated |
| Echoes ML scorer | ~1.2× over the classical comb; all on-source events null | ✓ gated |
| Ringdown no-hair | σ(δ) 2.6× tighter; GW250114 δ = −0.16, Kerr-consistent | ✓ gated |
| Ringdown overtone (field-standard) | GW250114 A₂₂₁ bounded from zero; NPE cross-validated | ✓ gated |
python3 watch_event.py GW250114_082203In ~48 s this runs all three sub-projects (across three Python environments) and emits a one-page report: the ringdown remnant + overtone, the no-hair δ + Kerr consistency, and the echo Δt prediction + search p-value. Because the no-hair NPE is amortized, it costs seconds per event — a standing instrument, ready to point at the next GW250114-class event from O4b/O5. (See watch_GW250114_082203.md for the reference report.)
primordial_blackhole_search/ subsolar/PBH merger search (CNN + coincidence + dense-bank MF)
echoes/ post-merger echo search (comb + ML scorer)
ringdown_spectroscopy/ QNM fitting + SBI no-hair test + ringdown-package cross-check
watch_event.py (root) the event watcher — orchestrates all three into one report
*.md (root) black-hole physics notes (holography, dimensions, paradox)
*.py (root) figure generators (light cones, Penrose maps, …)
verify.sh regression gate over every headline artifact (36 checks)
dashboard.py live run monitor (stdlib HTTP, no deps)
Each sub-project is self-contained: scripts/ (numbered steps + a shared lib), notes/lab_notebook.md (the raw record), README.md (methods + decisions), and its own uv virtual environment (pins in requirements.txt; the ringdown package pipeline needs Python 3.11 — requirements-py311.txt).
Large, regenerable artifacts — raw LIGO strain, spectrogram shards, model weights, venvs — are not committed; only the code, docs, result records, and the segment manifest.json (so the exact GPS segments can be re-fetched from GWOSC).
cd <subproject> && uv venv && uv pip install -r requirements.txt # per-project venv
.venv/bin/python scripts/01_*.py # fetch data from GWOSC, then run the numbered pipeline
./verify.sh # from the root: 36 checks assert every headline resultDocs: ROADMAP.md (next moves) · PLAN.md (full backlog, done + parked) · JOURNAL.md (dated log) · each sub-project's notes/lab_notebook.md.
- Sensitivity from injections into real noise — never assumed.
- Significance from a measured background — never a theoretical p-value.
- Pre-registration of each analysis before looking at on-source data.
- Null results are results. Several headline outcomes here are clean negatives, documented as carefully as the positives — including why they had to be negative.
- Stress-test before believing. The move that most often changed a conclusion was co-injecting on identical data, enlarging a small background, or refusing a false negative from an inadequate tool — each caught a would-be overclaim.
Built entirely on public data from the Gravitational Wave Open Science Center (GWOSC), produced by the LIGO/Virgo/KAGRA collaborations (not affiliated). All SNRs are band-limited to [50, 1024] Hz; whitening is normalized so whitened-domain energy equals matched-filter SNR². Stack: Python · PyTorch · gwpy · gwosc · pycbc · sbi · the Isi/Farr ringdown package. If you use this work, see CITATION.cff.