Nuclear waste characterization using Bayesian methods

The relance of the nuclear industrial sector in France is an ambitious plan expected to meet the energy needs of the general public in the 21st century. Along with the construction of new nuclear power plants, advancements are taking place in the fields of the nuclear fuel cycle and nuclear waste management. In order to comply with the criticality limits, nuclear waste is stored in special drums and different types of nuclear measurements are employed in order to determine the amount of the nuclear matter (U, Pu) contained inside them. In this work, we present an experimental design for the measurement of waste drums using the SYMETRIC 3He detector array located at CEA Cadarache. An experimental design using simulated data has been constructed for different parametrizations of waste drums in order to establish a model that has the potential of effectively predicting the calibration coefficient CP9, which is directly related to the amount of fissile matter inside the drum. Up until now, the models trained by the simulated data of the experimental design were exported by multiple linear regression. In an attempt reduce the statistical uncertainties, we have performed the analysis of the experimental design by applying two Bayesian methods: the Gaussian process and a Bayesian neural network. Both methods have resulted in a considerable improvement on several statistical metrics, showing promising performance for future applications.