Probabilistic modeling of heterogeneous radioactive waste for uranium radioactivity quantification using an AI-based surrogate model and Bayesian inference

In this study, we propose a modeling method applicable in situations where information regarding the physical geometry, chemical composition, and source distribution of the measured object is limited. In gamma spectrometry, reference materials or Monte Carlo simulations can be used for detection efficiency calibration. In the case of radioactive waste, using reference materials is challenging, making Monte Carlo simulations generally preferred. However, simulation accuracy diminishes for heterogeneous waste with scant detailed information. To address this challenge, we introduce a probabilistic waste matrix model for estimating the radioactivity of heterogeneous waste. Model parameters are determined using Bayesian inference, and an AI-based surrogate model is employed to generate spectra for likelihood evaluation. Our approach simplifies the complex geometry of radioactive waste into a unified structure with void regions and approximates its diverse chemical composition using three representative elements chosen based on mass attenuation coefficient ratios. Tests using synthetic datasets and experiments indicate that the proposed method enhances uranium radioactivity estimates by three-to six-fold over conventional deterministic variable-based nondestructive gamma spectrometry.