Large-scale design optimisation of boiling water reactor bundles with neuroevolution

We combine advances in deep reinforcement learning (RL) with evolutionary computation to perform large-scale optimisation of boiling water reactor (BWR) bundles using CASMO4/SIMULATE3 codes; capturing fine details, radial/axial fuel heterogeneity, and real-world constraints. RL constructs neural networks that learn how to assign fuel and poison enrichment by narrowing the search space into the areas where human/physics knowledge demonstrate merit. Evolution strategies diversify the search in these areas, through obtaining guidance from RL candidates. With very efficient/parallel implementation, our optimisation approach is able to solve a coupled multi-zone BWR bundle optimisation with similar to 40 constraints. The methodology is applied to a GE14-10x10 bundle, showing the ability of neuroevolution to find similar to 100 feasible designs. The optimal bundle has 7 axial zones with non-uniform enrichment radially and axially. The results of this work also demonstrate that our neuroevolution methodology is sufficiently generic to adapt to other assembly and reactor designs with minor adjustments. Published by Elsevier Ltd.