FLUX CHARACTERIZATION OF THE PURDUE UNIVERSITY SUBCRITICAL PILE USING NEUTRON ACTIVATION AND MACHINE LEARNING

Advancements in engineering have pushed research efforts for radiation damage and effects testing in extreme radiation environments, increasing the usage of subcritical piles to perform testing. Piles operate using an external neutron source and feature a flexible configuration. Generating an accurate flux map is essential to quantifying the radiation exposure for devices and materials placed in the pile. Conventional approaches for characterization often rely on simulated data due to challenges including spatial measurement limitations and time constraints. The presented thesis demonstrates the construction of a flux map for the Purdue University subcritical pile through the integration of physical measurements taken using neutron activation analysis and machine learning regression models. Experimental data were collected at multiple locations throughout the pile and were used to train and evaluate Random Forest, K-Nearest Neighbors, Support Vector Machine, and Gaussian Process models. Performance of each program was assessed using mean absolute error, mean squared error, root mean squared error, and R². Following a comparison, the Gaussian Process model achieved the highest accuracy, with minimized error and an R² of 0.725. To improve efficiency and mitigate the challenges mentioned, a Random Search optimization algorithm was implemented to minimize the number of required data points while preserving the accuracy of the model, using only 20% of the original data while preserving 90% prediction accuracy. Future efforts will explore more advanced learning algorithms with the goal of further increasing the accuracy of the predictive model and extend the methodology to the flux mapping of the Purdue University Reactor One (PUR-1) research reactor.