Reinforcement learning for BWR fuel management

Boiling water reactor (BWR) core design is a highly constrained optimization problem that directly affects fuel costs in operating reactors. The problem consists of both fuel bundle design and core loading pattern design, each of which presents significant computational and combinatorial challenges. Historically, stochastic optimization methods such as genetic algorithms and simulated annealing have been the most widely used approaches for automated bundle and core design. More recently, reinforcement learning (RL) has emerged as a promising class of algorithms capable of solving complex design problems while exploiting underlying physics-based patterns. This work presents a reinforcement learning framework for both fuel bundle and loading pattern optimization and demonstrates a sequential approach to BWR core design. First, the performance of the loading pattern optimizer is evaluated independently by reproducing a historic fuel cycle from an operating reactor using industry-designed assemblies. The optimizer consistently meets all cycle constraints and statistically outperforms the reference loading pattern, as demonstrated by a Wilcoxon signed-rank test. The full sequential optimization framework is then applied to design new fuel bundles and corresponding loading patterns for the same cycle. These results demonstrate that reinforcement learning can effectively address the sequential bundle and core design problem and highlight its potential to reduce design time and fuel costs in BWR core optimization.