Machine Learning Metamodel of a Computationally Intense LOCA Code

Abstract

The nuclear power industry is increasingly identifying applications of machine learning to reduce design, engineering, manufacturing, and operational costs. In some cases, applications have been deployed and are providing value, particularly in data rich manufacturing areas. In this paper, we use machine learning to develop metamodel approximations of a computationally intense safety analysis code used to simulate a postulated loss-of-coolant accident (LOCA). Accurate metamodels run at a fraction of the computational cost (milliseconds) compared to the LOCA analysis code. Metamodels can therefore support applications requiring a high volume of runs such as optimization, uncertainty analysis, and probabilistic decision analysis, which would otherwise not be possible using the computationally intense code. In this study, training data is first generated by running the safety analysis code over a design of experiment. Exploratory data analysis is then performed followed by an initial fitting of several model forms, including neighbor-based models, tree-based models, support vector machines, and artificial neural networks. A neural network is selected as the most promising candidate and hyperparameter optimization using a genetic algorithm is performed. Finally, the resulting model, its potential applications, and areas for further research are discussed.