A novel peak-searching method for multiple radioisotopes based on deep learning.

Gamma-ray spectroscopy is the cornerstone of nuclear security, arms control verification, and emergency response. However, traditional radionuclide identification methods struggle with the massive data streams and complex multi-nuclide environments characteristic of modern mobile detection platforms. This study proposes an automated deep learning-based peak-searching framework: it outputs the channel coordinates of photopeaks, which are subsequently matched against a radionuclide energy library for identification. A comparative analysis of three distinct architectures-convolutional neural networks (CNNs), residual networks (ResNets), and Transformers-was conducted. The results demonstrate that The CNN model provides the most balanced performance, achieving a precision of 75.41% and a recall of 92.53% (F1 = 0.8310) under the strict channel-matching criterion, and 89.00%/95.73% (F1 = 0.9224) under the ±FWHM tolerance criterion. The Transformer model exhibited poor localization precision under strict constraints, attributable to the mismatch between its global self-attention mechanism and the strictly local nature of photopeak centroids; its performance rivalled that of the CNN under the ±FWHM tolerance criterion. Meanwhile, the ResNet achieved the highest recall, albeit with a higher false-positive rate. This study provides a robust theoretical and engineering foundation for automated real-time radionuclide identification systems in complex radiological environments.