Deep learning-based real-time object detection for processing and analyzing optical readout data in the MIGDAL experiment

A number of direct dark matter detection experiments invoke the yet-to-be experimentally verified Migdal effect in nuclear scattering to improve sensitivity to low WIMP-mass DM interactions. The Migdal in Galactic Dark mAtter expLoration (MIGDAL) experiment aims to make the first measurement of the Migdal effect in fast-neutron scattering. The experiment uses an Optical TPC (OTPC) with double glass GEM amplification in 50-Torr CF₄ gas and employs a CMOS camera, 120 indium-tin-oxide strips, and a PMT as readouts. The CMOS camera records 2.4 MP images at rates of up to 120 frames per second (fps), or about 45 TB of image data per day, posing a challenge of efficiently analyzing this data. In this talk we introduce the end-to-end image processing and analysis pipeline employed for the optical readout of the OTPC. This pipeline uses YOLOv8, a state of the art convolutional neural network that’s trained to simultaneously identify and classify particle tracks observed in the OTPC. All together, the pipeline processes and analyzes images at 200 fps on consumer PC hardware, reducing our daily data footprint from 45TB, to about 10GB of processed physics data. YOLOv8 is also an excellent tool for Migdal effect-candidate skimming, which we will detail in this talk.