Metaheuristic Optimization of Control Rod Positioning in a Fast Breeder Reactor Core with Machine Learning-Based Surrogate Modelling for Fitness Evaluation

This study explores the feasibility of applying computational intelligence to optimize core design in fast breeder reactors. Optimization of the arrangement of twelve control rods within the active core of Prototype Fast Breeder Reactor (PFBR), satisfying operational and safety requirements is addressed as a case study. This constitutes a multi-objective, multi-constraint combinatorial optimization problem, encompassing a large search space comprising around 5 × 1014 potential configurations. Classical optimization techniques or manual approaches are inadequate to efficiently explore this space due to simulation-based fitness functions, discrete decision variables and multiple constraints in reactor physics and engineering domains. An integer coded Genetic Algorithm (GA) with customized crossover and mutation operators is developed to identify the best possible solutions for the problem. The most time-consuming and computationally intensive aspect of the GA workflow is the evaluation of core configurations, requiring multiple executions of the fast reactor core simulation code (FARCOB). To accelerate the GA optimization process, we propose a two staged Machine Learning (ML) based surrogate model to substitute FARCOB executions. First stage constitutes a classification ML model for screening sub-optimal configurations unlikely to satisfy operational or safety constraints. In the second stage, a regression model is employed to predict fitness values of configurations classified as potentially feasible, requiring substantially lesser number of FARCOB executions. Four ML algorithms—Long Short Term Memory networks (LSTM), Dense Neural Networks (DNN), Random Forests (RF), and Extreme Gradient Boosting (XGB)—are compared to identify the best performing one for the surrogate model. XGB is selected, as it demonstrated superior performance for both classification and regression. Comparative analysis demonstrates that incorporating surrogate model reduced computational time of GA optimization process by approximately 73%, while maintaining solution quality and convergence behaviour.