Deep learning-based radioisotope quantification for extreme environments

Abstract

Gamma-ray spectroscopy is a crucial technique for identifying and quantifying radioisotopes, but its reliability is severely compromised in extreme environments such as high temperatures and intense radiation. Traditional analysis methods, which depend on stable reference spectra and post-acquisition corrections, struggle to address the complex and non-linear spectral distortions arising from these conditions. This study introduces a deep learning-based system designed to offer a more robust and direct analytical approach. We developed a solution combining a ruggedized gamma-ray detector with a 2D convolutional neural network (CNN) that estimates radioisotope proportions directly from raw, distorted spectra. The proposed system demonstrated exceptional generalization and robustness. Trained on a sparse subset of data (¹³⁷Cs, ⁶⁰Co, ²²Na, ¹³³Ba, and ¹⁵²Eu) at varying temperatures (25–150 °C) and with a Ce:GPS scintillator exhibiting radiation-induced degradation (0–1.67 MGy), the model accurately estimated isotope proportions even under untrained conditions. It achieved low mean absolute error (MAE) values for both untrained temperatures (1.82 % at 75 °C and 125 °C) and untrained post-irradiation conditions, achieving an average MAE of 1.86 % across the untrained dose steps (with a localized increase for ¹⁵²Eu at dose step 2). These results validate the system's ability to operate effectively without requiring specific environmental information or calibration adjustments, showcasing a significant advantage over conventional methods. This work represents a significant advancement in gamma-ray spectroscopy by providing a reliable solution for isotope quantification in challenging, high-stress environments. The system's strong generalization capabilities pave the way for practical applications in nuclear accident monitoring, radioactive waste management, and other fields where traditional methods face significant limitations.