Accurate and fast prediction of radioactive pollution by kriging coupled with auto-associative models

Abstract. Uncertainty estimation is a key issue in nuclear crisis situations. Probabilistic methods for taking uncertainties into account in assessments are often costly in terms of the number of simulations and computation time. This is why emulation methods, which enable rapid estimation of numerical model outputs, represent a promising solution. However, in the context of radioactive dispersion modeling, existing emulators are mostly limited to scalar outputs. In a crisis context, decisions are often based on dose maps, which are mathematically represented by high-dimensional data. In this study, we use the auto-associative model method to reduce the dimension of dose results and then predict these reduced representations using kriging. We also compare this prediction method with others used by the French Nuclear Safety and Radiation Protection Authority (ASNR) to predict the consequences of a nuclear accident.