Using Unsupervised Neural Networks for Signal Classification in XENONnT

The XENONnT experiment is a dual-phase time projection chamber that utilizes liquid xenon (LXe) as a target for the direct detection of dark matter. Reconstruction of the interactions in our detector is crucial for rare event search of WIMP dark matter interactions. This requires distinguishing scintillation from ionization signals with very high accuracy to keep false positive detections (accidental coincidences) of WIMP dark matter as low as possible. We use unsupervised neural networks, in particular Self-Organizing Maps (SOMs), to aid in improving the classification of signals in XENONnT. SOMs excel at finding similarity groups in multi-dimensional data with great sensitivity. We found improved classification results using the SOM-aided classification over the current classification algorithm. Using the SOM-aided classification vs the current classification algorithm that is restricted to a two-parameter space, the classification accuracy of 37 Ar scintillation signals improved from 99.33% to 99.96% and for photoionization electrons from 99.51% to 99.91%. This improvement will reduce accidental coincidences by a factor of 3.039.