Prediction of plutonium content of BWR/MOX fuel using a neural network

Abstract

The Nuclear Material Balance (NMB) code is a comprehensive nuclear fuel cycle simulator developed by the Institute of Science Tokyo and the Japan Atomic Energy Agency. By modeling a fleet of reactors, fuel cycle simulators like NMB can help to optimize uranium and plutonium utilization, develop a reprocessing plan, and manage nuclear waste. However, a key challenge in this code is to simulate the fabrication of Mixed Oxide fuel, where the plutonium content must be properly tuned to achieve the targeted multiplication factor after irradiation. Traditionally, NMB’s prediction relies on predefined one-group cross sections and computationally intensive depletion calculations that involve solving burnup equations. This study addresses the inaccuracy due to the one-group cross-section as well as the computational burden by developing and integrating a Multi-Layer Perceptron (MLP) model into NMB. A custom PyTorch MLP was trained on a dataset of 5,000 depletion calculations generated with the SERPENT code for an ATRIUM-10 Boiling Water Reactor geometry. The developed MLP, in conjunction with a dichotomy search algorithm, determined the plutonium content with an error below 1% and achieved calculation speed of approximately 60 times for determination of Pu enrichment and 4 times for overall calculation.