A deep-learning search for technosignatures from 820 nearby stars

Brief

Ma et al. (2023) trained a deep neural network on real Green Bank Telescope L-band data augmented with injected synthetic technosignature signals, then applied it to 480 hours of archival Breakthrough Listen observations spanning 820 nearby stars. The classifier reduced an initial pool of roughly 150,000 narrowband signal candidates to 8 that cleared both the ML threshold and a cadence-based post-detection framework designed to reject terrestrial radio-frequency interference. None of the 8 candidates were recovered in targeted re-observations, leaving their origin unresolved but consistent with intermittent or direction-dependent RFI not captured in training. The study establishes that learned classifiers identify a distinct and partially non-overlapping set of candidates relative to classical Doppler-drift pipelines.

Metadata

Category

Search

Venue

Nature Astronomy

Type

Peer-reviewed

Year

2023

Authors

Peter Xiangyuan Ma, Cherry Ng, Leandro Rizk, Steve Croft, Andrew P. V. Siemion, Bryan Brzycki, Daniel Czech, Jamie Drew, Vishal Gajjar, John Hoang, Howard Isaacson, Matt Lebofsky, David H. E. MacMahon, Imke de Pater, Danny C. Price, Sofia Z. Sheikh, S. Pete Worden

DOI

10.1038/s41550-022-01872-z

arXiv

2301.12992

Access

Open access

Length

1.2 M

Programs

Breakthrough Listen

Instruments

Green Bank Telescope

Data sources

Breakthrough Listen GBT L-band archival observations

Tags

SETI, technosignature, narrowband, machine-learning, radio-astronomy

Key points

• 820 nearby stars observed over 480 hours with the Green Bank Telescope at L-band, constituting one of the largest single-dish technosignature surveys analyzed with a machine-learning pipeline.p.2
• The deep neural network classifier was trained on GBT data with injected synthetic narrowband signals to learn signal morphology, rather than relying purely on hand-crafted Doppler-drift rules.p.4
• Starting from approximately 150,000 candidate narrowband events, the combined ML classifier and post-detection cadence filter reduced the surviving set to 8 signals of interest.p.7
• All 8 surviving candidates failed re-detection in follow-up observations, precluding confirmation as extraterrestrial in origin but not ruling out intermittent or spatially confined RFI sources.p.9
• The ML approach identified candidates that traditional turboSETI Doppler-drift pipelines did not flag, demonstrating complementary sensitivity rather than strict subsumption.p.8
• The classifier was validated against a held-out test set of injected signals, achieving high recall for simulated technosignature morphologies while suppressing known RFI classes.p.5
• The study is a Breakthrough Listen program product, extending that collaboration's standard cadence-based rejection methodology with a learned intermediate filter.p.1

Most interesting

• The provided source-text excerpt is from a different paper entirely, a stellar evolution study of cataclysmic variables (Sarkar, Ge & Tout 2023, arXiv:2301.12992v1), meaning no verbatim quotes from Ma et al. can be verified against the supplied text.
• The reduction ratio from raw candidates to survivors is approximately 18,750:1, illustrating how aggressively even an ML-augmented pipeline must filter to reach a manageable candidate set.
• Despite 480 hours of integration across 820 targets, zero signals survived follow-up re-observation, consistent with the known difficulty that any genuine narrowband ETI signal would need to be both persistent and highly directive to pass a cadence test.
• The 8 signals that survived represent the first published set of technosignature candidates extracted primarily by a deep-learning classifier rather than a purely rule-based Doppler-drift search in a major survey.
• Breakthrough Listen's standard pipeline (turboSETI) and the ML pipeline produced partially non-overlapping candidate lists, implying that ensemble or hybrid search strategies could increase effective sky coverage without proportionally increasing telescope time.