Machine Learning Approaches in Room Temperature Semiconductor Detectors

Room Temperature Semiconductor Radiation Detectors (RTSDs) such as CdTe, CdZnTe, CdZnTeSe are becoming popular in novel space, homeland security and other medical applications (for example in Computed Tomography (CT) Imaging) which require sub-millimeter position information of interacting high energy (in the keV range) X-rays and γ-rays at excellent spectroscopic performance. These detectors are often pixelated requiring cumbersome post interaction 3D event reconstruction algorithms, which can benefit by detailed material characterization at the micron level. Characterization of these RTSD with respect to the charge transport and material defects is a vital step before implementing such detectors in real world systems. Several RTSD crystals are typically used in a detector module and current state-of-art characterization is done either as a whole or at most pixel-by-pixel over each RTSD crystal. Characterizing each RTSD crystal itself is a time consuming and labor-intensive process, and thus characterizing the entire RTSD module would involve much more labor and time. In order to address all these limitations, we present a novel machine learning model to characterize the RTSD crystals at micron length scales which is faster than the current detector characterization methods. The model is based on physical transport of charges (electrons and holes) with trapping centers for electrons and holes as in the actual RTSD crystal. The model discretizes the detector volumetrically as voxels and different material properties such as charge drift, trapping, detrapping, and recombination are modeled as learnable parameters in each voxel. Our model replicates the behavior of an actual physical RTSD. The input to our model is the high energy photon (X-ray or γ-ray) induced electron-hole pairs at various voxels and the outputs are the signals at the electrodes at either end of the RTSD and distributed charges (electrons and holes) in the voxels. Several reduced order models using fewer outputs such as using only the charges and/or the signals at the electrodes have also been developed which charcterize the RTSD either completely or in an equivalent manner. Our method can be applied to not only RTSDs but also to semiconductor detectors such as High purity Ge (HPGe) and others.