Abstract
This work addresses the NP-hard combinatorial optimization problem of determining the optimal fuel loading configuration for the IRT-T research nuclear reactor, a task critical for enhancing both economic efficiency and operational safety. Traditional methods are hindered by the core’s asymmetric beryllium reflector, heterogeneous fuel burnup distribution, and the vast combinatorial space of possible assembly arrangements (20! configurations), making exhaustive search or conventional optimization impractical. We propose a hybrid artificial intelligence system combining machine learning predictors and a genetic algorithm to efficiently navigate this high-dimensional solution space. Gradient Boosting models with L2-regularization were developed to accurately predict key neutronic parameters—power density distribution ( \(K_{V}\) ) and reactivity margin ( \(\rho\) )—accounting for localized effects from control rod movements. These ML models achieved high predictive accuracy ( \(R^{2}\) up to 0.98 for \(K_{V}\) and 0.97 for \(\rho\) ) and were integrated into a genetic algorithm optimization framework. The genetic algorithm employs adaptive modification operators–including priority-based crossover and power-law decay mutation–guided by a multiobjective fitness function balancing \(K_{V}\) minimization and \(\rho\) maximization. To ensure practical applicability, a novel symmetry-based regularization strategy was introduced, incorporating pairing, quadrant, and side symmetry coefficients to reflect real-world refueling constraints. Testing on actual IRT-T 2022–2023 fuel cycles demonstrated that the algorithm consistently identified improved loading patterns, yielding an average increase in fuel cycle length of 14 \(\%\) (up to 18 \(\%\) in best cases) compared to manual designs. The approach reduces optimization time from several hours to minutes while maintaining physical realism, offering a scalable and efficient solution for in-core fuel management in research reactors.
