Enhanced gamma-ray spectrum transformation: NaI(Tl) scintillator to HPGe semiconductor via machine learning

Abstract

Thallium-activated sodium iodide scintillation (NaI(Tl)) and high-purity germanium semiconductor (HPGe) detectors are two commonly employed gamma spectroscopy devices. NaI(Tl) detectors are preferred for their cost-effectiveness, efficiency, and ease of construction, while HPGe detectors have superior resolution but face challenges in temperature operation and they are expensive. This article investigates the application of machine learning algorithms, specifically K-Nearest Neighbors (KNN) and a Multi-Channel Output Regression based on Support Vector Regression (MCO-SVR), to enhance the performance of NaI(Tl) detectors by transforming its gamma spectrum into HPGe spectrum. The model was trained using datasets generated from a limited radioisotope library and demonstrated excellent performance across a diverse range of measured experimental test data. The evaluation included various scenarios, such as low-count spectra and background effects. The KNN model exhibited optimal performance, achieving an accuracy of 98.69% with a Manhattan distance metric. In contrast, the MCO-SVR model, employing both direct and chained approaches, exhibited varied results with different kernel types, with the polynomial kernel in the direct approach yielding the value 97.45% accuracy. Overall, the results indicate that machine learning algorithms have the potential to improve the performance of NaI(Tl) detectors and expand their applications in various fields of nuclear security.

Graphical Abstract