首页|Multi-nuclide identification in HPGe gamma-ray spectra using physics-informed peak features and nuclide-wise neural networks
Multi-nuclide identification in HPGe gamma-ray spectra using physics-informed peak features and nuclide-wise neural networks
Zhu, Dr. Zuolong Zhuang, Prof. Sixuan Wen, Mr. Xiang-Lin Ma, Mr. Yinglin Jiang, Ms. Xiao-Xue Li, Mr. Lun Li, Mr. Ye-Mian Zhang, Ms. Ge Wang, Mr. Yu-Fei Zhang, Mr. Gang
Multi-nuclide identification in HPGe gamma-ray spectra using physics-informed peak features and nuclide-wise neural networks
Multi-nuclide identification in HPGe gamma-ray spectra using physics-informed peak features and nuclide-wise neural networks
摘要
Automatic radionuclide identification in high-purity germanium (HPGe) γ-ray spectra remains difficult under low-count, multi-nuclide, interference-rich, and variable detector response conditions. This study proposes a modular multi-label framework based on physics-informed peak features and nuclide-wise neural networks. Each network combines local spectral windows around selected primary characteristic peaks with physics-informed features describing local peak evidence, full-spectrum contribution, and multi-peak consistency. Training spectra were assembled at the event level from FLUKA-generated single-nuclide event pools and measured HPGe background events, with randomized count levels, interference compositions, channel shifts, and resolution scaling. Independent simulated test spectra and measured spectra were used to assess simulation-to-measurement transfer, transfer to an external HPGe system, and deployment-oriented application. Ablation analysis showed that the spectral and physics-informed branches were independently informative and complementary. Among the evaluated variants, the full fusion model achieved the highest library-level label and exact-match accuracies in both simulated and measured evaluations. Targeted augmentation of difficult samples had a smaller, condition-dependent effect but improved overall exact-match accuracy. In the four-nuclide benchmark, label accuracy was 99.863% for simulated mixtures and 99.895% for mixtures constructed from independently acquired single-nuclide source event pools; exact-match accuracy was 99.450% and 99.580%, respectively. For 11,000 simulated ten-nuclide mixtures, label accuracy reached 99.244%, whereas exact-match accuracy was 92.909%. This reduction was associated with extremely weak components rather than with mixture cardinality alone. On spectra sampled from ten-nuclide mixed-source event pools, the library achieved 99.999% label accuracy and 99.990% exact-match accuracy on the development HPGe system, with corresponding values of 99.986% and 99.860% when applied to an external HPGe system without retraining. The library was then expanded to 19 radionuclides and applied to two concrete samples. The principal outputs were consistent with characteristic-peak evidence from 40K and natural decay-series radionuclides, whereas marginal 109Cd scores exposed sensitivity to unmodeled local interference. Taken together, the four-, ten-, and nineteen-nuclide evaluations support the framework's modular extensibility, physical traceability, effective generalization from simulated training data to measured spectra, preliminary transfer across HPGe systems, and deployment-oriented application to measured samples with unknown radionuclide compositions.