Improvement of Radiological Waste Characterization with Bayesian Machine Learning

We present a new methodology for the Improvement of RAdiological WAste Characterization with Bayesian MAchine Learning (IRAWAMA) using information from high-resolution gamma spectrometry with high-purity germanium (HPGe) detectors. As a real-life example, we consider the gamma spectrometry of a 200-L waste drum within the QUantitative ANalysis of TOxic and nontoxic Materials (QUANTOM®) device. To precisely represent the spatial material distribution of the waste matrix, the waste drum and the surrounding measuring arrangement of the QUANTOM device, and to correctly consider all relevant electromagnetic interactions (photoelectric absorption, Compton scattering, bremsstrahlung, pair production and annihilation), the Monte Carlo N-Particle Transport code MCNP 6.2 is used in combination with the ENDF/B VII.1 nuclear data library to compute the energy-dependent count rate measured by the HPGe detectors of the QUANTOM device. A large number of MCNP calculations are performed to calculate the HPGe detector signals for different combinations of Monte Carlo–sampled activities and spatial locations of the considered radionuclides and of the material density of the waste matrix. The resulting multivariate data set is used as training data for the applied Bayesian machine learning model. To quantify the performance of the trained model with respect to the prediction of activities and spatial locations of radionuclides within the waste matrix from the measured gamma spectra, the trained model is applied to a large number of simulated test cases. For each test case, the activities and spatial locations predicted from the gamma spectrometric data are compared to the true activities and spatial locations. Our test results show that the IRAWAMA methodology has the potential to significantly reduce the uncertainties of activities and spatial distributions of radionuclides within waste packages and to reliably quantify the related prediction errors. This is the case even if the density distribution within the waste matrix is unknown. The IRAWAMA methodology may, therefore, be of great benefit to producers of low- and intermediate-level waste intended for final disposal.