Neural-network-based reconstruction of spent-fuel radioactive distribution in nuclear waste cask

Abstract

A passive, nondestructive measuring method based on a convolutional neural network (CNN) algorithm was proposed for the rapid identification of nuclides and the precise measurement of radionuclide distribution in solid nuclear waste. In this study, the distribution of long half-life activation products—namely, 58Co, 60Co, 51Cr, 137Cs, 152Eu, 59Fe, 54Mn, 124Sb, 65Zn, and 95Zr—in nuclear waste casks containing cement-cured spent fuel from pressurized water reactor (PWR) nuclear power plants was used as an example, and a Geant4-based “Radiation Distribution Detector Count” dataset was developed. Based on the features of distribution reconstruction, a CNN model was constructed, its parameters being modified and trained. We then evaluated the network model based on the reconstruction of the inhomogeneous distribution of radioactivity for each nuclide in the nuclear waste cask, the results proving to be acceptable for the evaluation parameters of the average relative error (ARE) and the maximum relative error (MRE). Further, a detector design optimization approach was provided for combining the reconstruction accuracy of the radioactivity distribution with the detector point arrangement.