Deep learning-based object detection algorithms simultaneously classify and localize any number of objects in image data by placing bounding boxes over each detection. Such algorithms are capable of accurately detecting objects that incur significant spatial overlap and are thus a viable tool for detecting composite event topologies. In cases where a rare event signal is a composite consisting of commonly-observed constituents, an object detection algorithm could be trained to detect these constituents, and then use the locations of the constituents to identify rare event candidates. This opens up the possibility of data-driven training for rare event searches.

In this talk, I will demonstrate this principle for the MIGDAL experiment’s rare event search for the Migdal effect. In particular, I will present an end-to-end object detection pipeline trained on real data to identify electronic and nuclear recoils – the constituents forming the characteristic Migdal effect topology. In addition to benchmarking the pipeline’s detection performance and inference speed for MIGDAL, I will highlight the pipeline’s potential for other rare event searches. Specifically, I will discuss the adaptability of this approach to other applications with composite rare event signals, including neutrinoless double beta decay experiments.