Using anomaly detection to search for technosignatures in Breakthrough Listen observations

Abstract: We implement a machine learning algorithm to search for extra-terrestrial technosignatures in radio observations of several hundred nearby stars, obtained with the Parkes and Green Bank Telescopes by the Breakthrough Listen collaboration. Advances in detection technology have led to an exponential growth in data, necessitating innovative and efficient analysis methods. This problem is exacerbated by the large variety of possible forms an extraterrestrial signal might take, and the size of the multidimensional parameter space that must be searched. It is then made markedly worse by the fact that our best guess at the properties of such a signal is that it might resemble the signals emitted by human technology and communications, the main (yet diverse) contaminant in radio observations. We address this challenge by using a combination of simulations and machine learning methods for anomaly detection. We rank candidates by how unusual they are in frequency, and how persistent they are in time, by measuring the similarity between consecutive spectrograms of the same star. We validate that our filters significantly improve the quality of the candidates that are selected for human vetting when compared to a random selection. Of the ~ 10^11 spectrograms that we analyzed, we visually inspected thousands of the most promising spectrograms, and thousands more for validation, about 20,000 in total, and report that no candidate survived basic scrutiny.