Application of Machine Learning to MICADO Passive and Active Neutron Measurement System for the Characterization of Radioactive Waste Drums

A passive and active neutron measurement system has been developed within the Measurement and Instrumentation for Cleaning and Decommissioning Operation (MICADO) H2020 project to estimate the nuclear material mass inside legacy waste drums of low and intermediate radioactivity levels. Monte–Carlo simulations were performed to design a transportable neutron system allowing both passive neutron coincidence counting and active interrogation with the differential die-away technique (DDT). However, the calibration coefficients (CCs) representing the signal of interest (due to nuclear material) in these two measurement modes may vary by a large amount depending on the properties of the matrix of the nuclear waste drum. Therefore, this article investigates matrix effects based on 104 Monte–Carlo calculations with different waste drums, based on Taguchi experimental design with a range of densities, material compositions, filling levels, and nuclear material masses. A matrix correction method is studied using machine learning algorithms. The matrix effect on the neutron signal is deduced from the signal of internal neutron monitors located inside the measurement cavity and from a transmission measurement with an AmBe neutron source. Those quantities can be assessed experimentally and are used as explanatory variables for the definition of a predictive model of the simulated CC, either in passive or in active mode. A multilinear regression model of the CC based on ordinary least square (OLS) is built and compared to the random forest (RF) machine-learning algorithm and to the multilayer perceptron (MLP) artificial neural network. In passive neutron coincidence counting, the residual error of the regression is lower for the MLP and RF than for OLS. The agreement between the predicted CCs of four mockup drums used as test is better than 17% and 3%, respectively, with the MLP and RF methods, while three predictions are out of the 95% confidence level range with OLS. In active neutron interrogation, similar conclusions are drawn. The prediction of the CC for the four mockup drums is better than 12%, 35%, and 72% for the respective MLP, RF, and OLS methods. In conclusion, the MLP and RF regression model demonstrates more accurate results of the quantities of interest than the traditional OLS method. The future steps will focus on matrix heterogeneities, experimental validation, improving our models and testing new regression approaches.