Inverse estimation scheme of radioactive source distributions inside building rooms based on monitoring air dose rates using LASSO; Theory and demonstration

Shi, W.*; Machida, Masahiko  ; Yamada, Susumu  ; Yoshida, Toru*; Hasegawa, Yukihiro*; Okamoto, Koji*

Predicting radioactive source distributions inside reactor building rooms based on monitoring air dose rates is one of the most essential steps towards decommissioning of nuclear power plants. However, the attempt is rather a difficult task, because it can be generally mapped onto mathematically ill-posed problem. Then, in order to successfully perform the inverse estimations on radioactive source distributions even in such ill-posed conditions, we suggest that a machine learning method, least absolute shrinkage and selection operator (LASSO) minimizing the loss function, is a promising scheme. For the purpose of its feasibility demonstrations in real building rooms, we employ PHITS code to make LASSO input as the above matrix C connecting the radioactive source vector P defined on surface meshes of structural materials with the air dose rate vector Q measured at internal positions inside the rooms. We develop a mathematical criterion on the number of monitoring points to correctly predict source distributions based on the theory of Candes and Tao. Then, we confirm that LASSO actually shows extremely high possibility for source distribution reconstructions as far as the number of detection points satisfies our criterion. Moreover, we verify that radioactive hot spots can be truly reconstructed in an experiment setup. At last, we examine an influence factor like detector-source distance to enhance the predicting possibility in the inverse estimation. From the above demonstrations, we propose that LASSO scheme is a quite useful way to explore hot spots as seen in damaged nuclear power plants like Fukushima Daiichi nuclear power plants.