HPGe-Compton Net: a physics-guided CNN for fast gamma spectra analysis via Compton region learning

High-purity germanium (HPGe) detectors have been golden standard for gamma spectrometry in low-level radioactive waste (LLW) analysis; however, their notable shortcoming is prolonged measurement durations for weak radioactive waste materials. The present study aimed to develop the HPGe-Compton Net, a 1D physics-guided convolutional neural network to accelerate LLW analysis by taking advantage of the entire response function of a HPGe detector for each radionuclide of interest, in contrast to the traditional methods that analyze only peak regions of the response. This acceleration is supported by two core innovative strategies: (a) channel-prompt method, a feature enhancement incorporating additional physical information to guide the model to locate the designated radionuclide; (b) the specially designed database to achieve effective targeted feature learning. The performance evaluation carried out for test data set showed a five times reduction in measurement time compared to a conventional spectral analysis method while maintaining comparable precision. Compton perturbation tests confirmed the model’s ‘smart’ adaptive utilization of the Compton regions. The generalization testing of four LLW samples as the external validation set proved its superior performance in low-count data with an average accuracy of 90% over 83% of the traditional method. Future work will focus on upgrading the HPGe-Compton Net for practical applications.