Machine learning for fast transients for the Deeper, Wider, Faster programme with the Removal Of BOgus Transients (ROBOT) pipeline

Abstract

The Deeper, Wider, Faster (DWF) programme is optimized to detect fast transients that show luminosity changes on time-scales of sub-second to days using fast cadence simultaneous observations and rapid response follow up. One of the significant bottlenecks in DWF is the time required to assess candidates for rapid follow up and to manually inspect candidates prior to triggering space-based or large ground-based telescopes. In this paper, we present the Removal Of BOgus Transients (ROBOTs) pipeline that uses a combination of machine learning methods, a Convolutional Neural Network (CNN), and Decision Tree (CART), to analyse source quality and to filter in promising candidates. The ROBOT pipeline is optimized for ‘lossy’ compressed data required by DWF for fast data transfer to find these candidates within minutes of the light hitting the telescopes. Preliminary testing of the ROBOT pipeline on archival data showed to reduce the number of candidates that require a manual inspection from 69 628 to 3327 (a factor of ∼21 times), whilst simultaneously sorting candidates into categories of priority, with potential for further improvement. Recent real-time operation of the ROBOT pipeline in DWF-O10 showed to further reduce manual inspections from ∼155 000 to ∼5000 (a factor of ∼31 times).