Comparative analysis of automatic radionuclide identification in γ-ray spectrometry: Machine learning versus statistical approaches

Gamma-ray spectrometry is a widely used technique for identifying and quantifying γ-emitting radionuclides in many nuclear applications. Currently, there is a growing trend to address the problem of automatic identification of radionuclides by implementing machine learning (ML) approaches such as multilayer perceptrons (MLP) or convolutional neural networks (CNN). Alongside these ML methods, the statistical method based on full-spectrum analysis with Poisson likelihood yields reliable results. However, due to the lack of a common benchmark, a comparison of these methods has not yet been conducted. In this work, we introduce a benchmark to evaluate and compare these methods under three scenarios: (1) known spectral signatures, (2) spectra deformed by physical phenomena and (3) gain shift. A large dataset of 200000 simulated spectra was generated for each scenario, spanning a broad range of radionuclide combinations and mixing ratios. This work utilizes a dictionary of nine radionuclide spectral signatures generated with the Monte Carlo code Geant4 for a 3"×3" NaI(Tl) detector, combined with an experimental natural background. For the ML method, we trained a dedicated CNN model for each scenario, carefully optimized its hyperparameters, and adjusted its classification threshold. The results show that the statistical method consistently outperforms the ML approach across all scenarios, particularly under low-statistics conditions. Furthermore, the statistical method maintains a false positive rate close to the predefined level.