Deep learning-based keypoint detection for electron recoil vertex identification and trajectory reconstruction

Machine learning (ML) techniques that are capable of simultaneously classifying and localizing objects in image data have shown promise in their ability to identify particle tracks with significant spatial overlap. Object detection is one method that is trained to detect tracks with bounding boxes and assign classifications. Recently, the MIGDAL experiment has used object detection in their rare event search for the Migdal effect, which consists of a nuclear recoil (NR) and low energy electron recoil (ER) sharing a vertex. While object detection can detect heavily overlapping ERs and NRs, it does not provide directional information of the tracks, which is crucial for verifying the topological signature of the Migdal effect. We therefore extend on MIGDAL’s work by utilizing object keypoint detection, a method that, in addition to detecting tracks, is also trained to identify key points within each detected object. We explore the use of object keypoint detection both for vertex reconstruction and trajectory fitting using a large sample of simulated 5.9 keV Fe55 tracks. We report preliminary results comparing 2D vertex position identification using keypoint detection to previously established non ML-based methods, and share our progress on fitting 2D trajectories of these tracks. Directional reconstruction of ERs is of broad interest to applications beyond the MIGDAL experiment, including Xray astronomy, astrophysical neutrino observations, 0νββ and dark matter searches.