Abstract
Gamma-ray spectrometry stands as a traditional technique for identifying and quantifying γ-emitting radionuclides in a wide range of nuclear physics applications, including rapid detection of illicit nuclear material trafficking, decommissioning of nuclear facilities, and in situ environmental analysis following a radiological or nuclear incident. For these applications, there is a growing demand for automatic analysis tools that can be used by non-expert users and also enable robust decision-making under short-duration measurement conditions (i.e., low statistics). Furthermore, the measurements carried out in complex environments can lead to variability of the shape of γ-spectra due to physical phenomena such as attenuation, Compton scattering and fluorescence resulting from the interactions of γ-photons in the surroundings between the radioactive source and the detector.To address these challenges, this thesis introduces a hybrid approach that combines machine learning and statistical methods to more accurately reflect the physical properties of the measurement process. Central to this approach is the Interpolating Autoencoder (IAE), a machine learning model designed to capture spectral variability. The IAE is evaluated using Geant4 radiation-matter simulations with a geometry involving a point source located in a sphere for the NaI(Tl) detector. Two types of IAE models are considered: an individual model, which learns the spectral variability independently for each radionuclide, and a joint model, which captures correlations in spectral variability across all radionuclides. The IAE models the spectral signatures—representing the detector's response to γ-photon emissions—as a function of a one-dimensional latent variable λ. Building upon the IAE, a novel hybrid full-spectrum spectral unmixing algorithm, SEMSUN, is developed to jointly estimate the spectral signatures and counting of all radionuclides using the maximum likelihood estimation under a Poisson distribution of the measurements. The SEMSUN algorithm is then combined with a model-selection strategy based on the likelihood-ratio test to enable automatic radionuclide identification. Finally, to quantify the uncertainty of the hybrid estimator, two Bayesian inference techniques are investigated: the Laplace approximation (LA) and Markov Chain Monte Carlo (MCMC).The results demonstrate that the IAE models can effectively capture spectral deformation, even in complex scenarios involving up to 12 radionuclides. The proposed hybrid approach has been compared to end-to-end machine learning methods, representing a current trend in γ-ray spectrometry. The findings show that the hybrid method outperforms end-to-end ML approaches in both identification and quantification tasks. It maintains a false positive rate (or false alarm rate) close to the expected value and gives superior detection capability under low-statistics conditions. The performance of this approach using individual IAE models is slightly less precise than that of the joint IAE model, as it does not capture the correlations between the spectral signatures of different radionuclides. Regarding uncertainty quantification, while the LA method is computationally efficient, the MCMC method offers more robust results when constraints strongly influence the distribution, though at a higher computational cost.
Key words
Gamma-Ray spectrometry, Machine learning, Spectral unmixing, Hybrid algorithm, Spectral variability, Artificial intelligence
Full text
Full text: https://theses.fr/2025UPASG068