Machine learning for nuclear fission systems : preliminary investigation of an autonomous control system for the MGEP

Abstract

Commercial nuclear technology today is facing challenges due to both economic viability and concerns over safety. Next-generation reactors could potentially improve with respect to both concerns through recent advancements in computation and machine learning, through autonomous control systems which minimize human error. The MIT Graphite Exponential Pile (MGEP) has been selected as the basis of a realworld demonstration of such a system, because of its simple properties and inherent safety. This study evaluated the preliminary feasibility of an autonomous control system for the MGEP through two parallel avenues; a practical investigation of various machine learning algorithms applied to fission systems, as well as the design and fabrication of a control rod for the pile. It was found that Convolutional Neural Networks (CNNs) outperform Support Vector Regression (SVR) in predicting the MITR power-shape. Additionally, acceptable results were achieved when applying the CNN algorithm to the MGEP to predict the flux distribution of its fuel elements. Finally, it was verified that neutron detectors in the pile respond predictably to control rod insertions. Taken together, the groundwork for the further development of an autonomous control system has been laid, and the path forward is promising.