Investigation into the use of machine learning assisted prediction of nodal parameters for reduced order neutronic simulation models

Deep neural networks (DNNs) were trained to predict nodal cross sections, chi values, and assembly discontinuity factors (ADFs) for pressurized water reactor (PWR) 2D pin cell models and 2D lattice models to assess the feasibility of using DNNs as nodal parameter generators. Separate DNNs were trained for each individual nodal parameter to improve prediction accuracy and to transfer learning employed to reduce dataset volume requirements. DNNs were found to train and predict well for pin cell and lattice models when provided with sufficient data, and the required number of data points required to develop accurate lattice DNNs could be significantly reduced through transfer learning using a previously trained pin cell DNN. (authors)