A Bayesian approach of the Source Term Estimate coupling retro-dispersion computations with a Lagrangian Particle Dispersion Model and the Adaptive Multiple Importance Sampling

This paper presents an enhanced version of a STE adaptive algorithm based on probabilistic Bayesian inference (AMIS) that estimates the parameters of an atmospheric pollution source term. After introducing the problem and presenting the initial results obtained with the first version of the algorithm, we describe an efficient way to reduce the computational load in its procedure by shifting the most time-consuming step outside the iterative loop. This is made possible by applying the duality relationship between forward and adjoint advection-diffusion equations, and is proven to work well on a synthetic example using Retro-SPRAY, the backward implementation of SPRAY, the Lagrangian particle dispersion model of the PMSS suite.