On the search for broadband pulsed beacons towards 1884 stars using neural nets

Radio searches for technosignatures are frequently focused around the narrowband continuous wave (CW) beacons. CW beacons are easier to generate, more obvious to identify as artificial, and can be detected across large interstellar distances. Broadband pulsed signals, with added artificiality, can also be utilized as powerful beacons. We demonstrate that the total power (kW-h) required to operate a CW transmitter exceeds a broadband transmitter over a longer operating timescale. The Breakthrough Listen program, which is conducting one of the most comprehensive searches for the evidence of intelligent life, thus extends radio technosignature searches for such preferable ETI broadband pulsed beacons. We outline results from our recently completed survey towards 1884 stars across 4 to 8 GHz with the Robert C. Byrd Green Bank Telescope. We searched for three different types of broadband pulsed beacon using GPU-accelerated tools, exhibiting non-physical dispersions, and found around 10^5 initial hits. We have locally developed and trained a fully supervised Convolutional Neural Network (CNN) based classifier, which helped us eliminate a large number of false-positives (~97%) of obvious anthropogenic origins. We visually vetted and scrutinized around 3000 remaining CNN classified top hits and were able to reject presence of a true pulsed ETI beacon. As our survey is one of the first-ever surveys to search for such broadband pulsed signals, we will show how we can constrain their existence with an average output power-densities of just 1000 W/Hz in the solar neighborhood. We will also discuss our plans to extend this survey to other classes of artificial broadband pulsed signals.