The operation of nuclear power plants requires regular fuel reshuffling. At the beginning of each cycle, elaborate computational modeling tools are employed to design a new core configuration. Surprisingly, loading patterns are still typically generated manually in practice, largely on the basis of heuristic rules and the experience of the core designer. The use of automated optimization algorithms to determine applicable core loading patterns has been the subject of research for decades. Earlier, researchers focused on various methods such as Simulated Annealing (SA), Genetic Algorithms (GA), Tabu Search (TS). Due to the large phase space of the optimization, these algorithms suffered significant performance drawbacks. Therefore, research interest in this field turned toward the application of artificial neural networks recently.

The aim of the proposed PhD project is to explore opportunities for improving loading pattern optimization procedures using artificial intelligence techniques. The discrete-system optimization problem has an inherently game-like structure, thus  Reinforcement Learning emerges as a natural candidate for its solution. The further goal will include the study of various models and formulations of a suitable optimization environment, including the definition of state, action, and reward representations consistent with the physical and operational constraints of the problem. The developed framework will then be trained and evaluated through repeated interaction with reactor-physics simulation tools, with particular emphasis on computational efficiency, robustness, and generalization capability.

To assess its practical applicability, the Paks Nuclear Power Plant is planned to be used as a case study, where the C-PORCA code system is available for highly accurate simulation of the plant’s reactor-physics processes, along with extensive historical measurement data and the expert knowledge required for safe and efficient plant operation.