Deep learning approach to nuclear fuel transmutation in a fuel cycle simulator

Abstract

We trained a neural network model to predict Pressurized Water Reactor (PWR) Used Nuclear Fuel (UNF) composition given initial enrichment and burnup. This quick, flexible, medium-fidelity method to estimate depleted PWR fuel assembly compositions is used to model scenarios in which the PWR fuel burnup and enrichment vary over time. The Used Nuclear Fuel Storage, Transportation & Disposal Analysis Resource and Data System (UNF-ST&DARDS) Unified Database (UDB) provided a ground truth on which the model trained. We validated the model by comparing the U.S. UNF inventory profile predicted by the model with the UDB UNF inventory profile. The neural network yields less than 1% error for UNF inventory decay heat and activity and less than 2% error for major isotopic inventory. The neural network model takes 0.27 s for 100 predictions, compared to 118 s for 100 Oak Ridge Isotope GENeration (ORIGEN) calculations.We also implemented this model into Cyclus, an agent-based Nuclear Fuel Cycle (NFC) simulator, to perform rapid, medium-fidelity PWR depletion calculations. This model also allows discharge of batches with assemblies of varying burnup.Since the original private data cannot be retrieved from the model, this trained model can provide open-source depletion capabilities to NFC simulators. We show that training an artificial neural network with a dataset from a complex fuel depletion model can provide rapid, medium-fidelity depletion capabilities to large-scale fuel cycle simulations.