Automation of the γ-ray spectrometry setup of the Environmental Radioactive Laboratory at NRF-iThemba LABS

Abstract

Environmental γ-ray spectrometry plays a critical role in radioactivity monitoring, radiation protection, and nuclear safety assessments. Conventional spectrometric analysis relies heavily on manual peak identification and expert interpretation, which can be time-consuming and subjective, particularly when dealing with complex environmental samples and varying measurement geometries. This thesis presents the development of an automated and physics-inspired machine learning framework for γ-ray spectrometry, aimed at improving the accuracy, efficiency, and robustness of radionuclide identification at the Environmental Radioactivity Laboratory (ERL) of NRF-iThemba LABS. The primary objective of this research was to integrate domain-specific knowledge from nuclear spectroscopy with advanced machine learning techniques to enable reliable automated analysis of γ-ray spectra. To achieve this, γ-ray spectral data were acquired from five selected radionuclides under controlled experimental conditions, including different counting geometries and known activity concentrations. The resulting dataset captured both the statistical and physical characteristics of detector responses, providing a solid foundation for model training and evaluation. Data preprocessing, feature handling, and visualization were carried out using Python and ROOT, ensuring consistency and reproducibility throughout the analysis pipeline. Two physics-inspired deep learning models, namely Convolutional Neural Networks (CNNs) and Kolmogorov-Arnold Networks (KANs), were developed and optimized for γ-ray spectral classification. These architectures were specifically designed to extract meaningful spectral features by exploiting the physical structure of γ-ray interactions, including peak shapes, Compton continua, and energy-dependent detector responses. By embedding physical intuition into the learning process, the models demonstrated strong generalization capabilities when exposed to previously unseen spectra. The performance of the proposed deep learning models was systematically compared with traditional machine learning algorithms, including k-Nearest Neighbours, Artificial Neural Networks, Support Vector Machines, Random Forests, Decision Trees, and AdaBoost. Evaluation metrics such as accuracy, recall, and area under the receiver operating characteristic curve revealed that the CNN and KAN models consistently outperformed conventional approaches across all radionuclides and geometries. Traditional algorithms exhibited limitations in handling spectral complexity and variability, underscoring the advantage of deep learning methods for high-dimensional nuclear spectroscopy data. To facilitate practical deployment, a Gradio-based interactive dashboard was developed, enabling real-time γ-ray spectrometry analysis. The dashboard allows users to upload spectra and receive immediate radionuclide identification results, along with visual feedback on spectral features and model confidence. This interface enhances accessibility and operational efficiency, bridging the gap between advanced machine learning models and routine laboratory workflows. Overall, this research demonstrates that physics-inspired deep learning provides a powerful and reliable approach to automated γ-ray spectrometry. The proposed framework represents a significant advancement in environmental radioactivity analysis and establishes a foundation for future extensions involving additional radionuclides, higher-resolution detectors, and adaptive learning strategies for real-world monitoring applications.