From Laboratory to Pocket: Development of Deep Learning Algorithms for Ionizing Radiation Detection Systems

Abstract

This thesis focuses on the application and development of deep learning algorithms for ionizing radiation detection systems. The main content is organized into three studies (published or under review) that address distinct radiation detection challenges: (a) the unfolding of beta radiation fluence spectra in mixed radiation fields, accompanied by the design of dedicated detection hardware; (b) accelerating gamma-ray spectroscopic measurement and analysis of High-Purity Germanium (HPGe) detectors; and (c) detecting ionizing radiation using a modern smartphone with no hardware modifications.

In these works, a variety of deep learning models are explored in depth, including one-dimensional convolutional neural networks, physics-informed neural net works, 3D–2D hybrid convolutional neural networks, and dual-branch multilayer perceptrons. The research provides innovative solutions to several key challenges in applying deep learning to radiation detection, such as the scarcity of high-quality datasets, the lack of interpretability of neural network models, and insufficient stability and generalization on independent external test sets. A unified and efficient deep learning development workflow for different detection systems is also proposed to guide future research and implementation.

Beyond methodological innovations, this thesis achieves significant progress in three practical applications: (a) real-time beta fluence spectrum unfolding; (b) a three- to five-fold acceleration of HPGe gamma spectrometry measurement using deep learning, along with effective targeted feature learning in the Compton region of the spectrum; and (c) demonstration that a standard smartphone (with no hardware modifications or camera shielding) can rapidly detect dangerous radiation levels and estimate their dose within six seconds. These successes across diverse application domains clearly demonstrate the effectiveness of the general deep learning–based approach to radiation detection developed in this work.