Physics-Enhanced Module for Small-Sample Radionuclide Identification

Accurate radionuclide identification is critical for nuclear safety and environmental radiation monitoring. To integrate the guiding effect of peak morphology features and the nonlinear fitting capability of data-driven models, this paper proposes a Physics-Enhanced Module (PEM). This module adopts the Laplacian of Gaussian (LoG) operator to decompose input gamma spectra and generate auxiliary features representing peak responses, which are concatenated with the original spectrum as an additional channel to inject peak position information into neural networks. Small-sample comparative experiments are conducted on a multi-label nuclide identification dataset covering 16 radionuclides, where PEM-ResNet and PEM-DenseNet are compared against vanilla ResNet and DenseNet baseline models. Experimental results demonstrate that all physics-enhanced networks outperform their corresponding baselines across all evaluation metrics. Consistent performance improvements observed on two mainstream convolutional architectures verify that the PEM effectively mitigates performance degradation induced by limited training samples. Further analysis reveals that the PEM provides the model with clearly identifiable peak response features for discrimination tasks. The integrated framework synergizes the fitting strengths of data-driven learning and the regularization merits of the fixed preprocessing module, thereby substantially improving the reliability and robustness of nuclide identification for nuclear energy applications.