The Application of Machine Learning to Event Classification in Radiation Detectors

Abstract

Radiation physics is typically troubled by the accompanying of an interesting radiation field with an uninteresting one, presented as the signal and the background respectively. In some instances this background is a consequence of the signal of interest and its interaction with the world around it. These issues can be alleviated by clever experimental design that take advantage of the highly predictable way in which radiation interacts with the world. This predictability can lead to structured data which lends itself to data driven techniques which enable the identification of signal and background. One such technique that can exploit the structure of radiation detection data is machine learning, which is the focus of this thesis. The first application was event selection onboard a satellite, called Advanced Energetic Pair Telescope (AdEPT), being developed for photon polarimetry. This application utilizes an adaptation of Google’s GoogLeNet which was trained off of simulated data produced in a simulation developed in Geant4. The performance of this adaptation, named GammaNet, was investigated and found to achieve the desired 99.99 % background rejection while maintaining signal sensitivity ranging from 0.1±0.1 % to 69±2 %. The signal sensitivity range depends on the down sampling rate implemented and the energy of the incident photon. The other application explored was neutron and photon separation in EJ-301 and EJ-309 neutron sensitive scintillators. This application, being experimental and not simulation based, required the implementation of an already existing classification routine to generate the labelled data necessary for machine learning. This was accomplished by utilizing Tail-To-Total (TTT) Pulse Shape Discrimination (PSD) and algorithmically fitting the resultant photon and neutron populations. With this technique it was possible to correctly identify 75.3±0.5 % of the neutrons while removing 99.9999 % of the photons. The neutron identification rate was found to depend on the neutron source used, but was mostly accounted for when considering the differences in neutron energies for those sources. Both machine learning applications had the feature extraction technique, Gradient weighted Class Activation Mapping (Grad-CAM), applied to them. This technique produces a spatially correlated activation mapping for each class inside the neural network, identifying regions of the input that produced the greatest confidence to an output classification. The features observed from both applications aligned with the intuition of what comprises the dominant features of each signal and background.