Efficient Bayesian source reconstruction and uncertainty quantification of atmospheric radionuclide releases by...

Bayesian reconstruction is important for understanding the source of atmospheric radionuclide emissions, especially given the renewed development of nuclear energy. However, it is computationally challenging to reconstruct source location and time-varying release rates. We propose a novel Bayesian framework that retrieves such information while bypassing the time-consuming computation and avoiding the unrealistic constant-release assumption in current Bayesian methods. It replaces the exhaustive sampling of release rates at each time interval with Maximum-A-Posteriori (MAP) estimation of the entire time-varying release rates, thereby simplifying the computation. The MAP-estimated release rates are further refined using PAMILT to remove oscillations. The probabilities of these MAP estimates approximate the a posteriori probability of time-varying release rates at different time intervals, which is missing in current Bayesian reconstruction. Validation against multiscale real cases demonstrated the consistently superior accuracy and lower uncertainty of the proposed method compared with Bayesian reconstruction adopting the constant-release assumption, reducing source location errors by 44.29 %-85.43 %, and avoiding unrealistic releases of the latter. The reconstructed release profiles matched the reported time windows and total amounts. Sensitivity analyses demonstrated its efficiency in automatic parameter selection and its robustness to priors for MAP estimation of release rates and likelihood functions.