Classification of group structures for a multigroup collision probability model using machine learning

Abstract

Abstract Multigroup neutron transport models can be significantly faster than continuous energy methods. While minimizing the number of energy groups improves runtime, this causes the accuracy of a calculation to become strongly dependent on selected energy group boundaries. Machine learning can exceed human performance on a range of classification tasks. In this work, the potential for supervised machine learning to classify few-group energy structures is evaluated. Artificial neural network and random forest classifiers were trained to determine whether a given 20-group energy structure enables a multigroup collision probability model to calculate accurate neutron multiplication factors in a light water reactor lattice simulation. The training data consisted of 20,000 random 20-group structures and their associated multiplication factors. Five-fold cross-validation was used for optimizing the hyperparameters of both machine learning algorithms. The trained neural network and random forest could classify input 20-group structures with up to 95.3% and 95.5% accuracy respectively.