Probabilistic position reconstruction in XENONnT

The XENONnT experiment uses a dual-phase xenon time projection chamber to search for dark matter and observe solar neutrinos. The sensitivity of the XENONnT detector and future experiments based on this technology is highly reliant on 3D position reconstruction. Position reconstruction in these dual-phase noble element time projection chambers is key to discriminating between signals, that are expected to be uniform in the detector, and surface backgrounds that are concentrated near the detector surfaces. In these detectors, the depth can be directly measured using the time difference between the primary scintillation and secondary signals, but the horizontal position needs to be estimated from the distribution of the secondary signal across the photomultiplier tubes. Multiple approaches have been considered for this role by XENON, LUX, LZ, and PandaX, including multilayer perceptrons, convolutional neural networks, and maximum-likelihood fitters. In this work, we present a probabilistic model based on normalizing flows that exceed the performance of other methods used in XENONnT, and provide a posterior with well-calibrated coverage. We also demonstrate how such a model can be used to improve analysis and reject poorly-reconstructed events in the detector.