AI-Based Prediction Module of Key Neutronic Characteristics to Optimize Loading Pattern for i-SMR with Flexible Operation

This paper proposes an AI-based module for a loading pattern (L/P) optimization algorithm applied to the i-SMR, designed for flexible operation. The AI module can be used as a surrogate model in the simulated annealing (SA) screening process, which allows for more efficient optimization. The convolution neural network (CNN) model was trained using reactor core L/Ps and corresponding core parameter values derived from a realistic core simulation code. For load-following operations, we selected core parameters such as control rod insertion depth, radial peaking factor, axial shape index, and effective multiplication factor. To calculate the objective function of an L/P during the SA process using core design codes, it takes approximately 3 s, while the AI-based module can predict the objective function within about 0.1 ms. During the prediction of selected parameters, we discovered two factors affecting prediction accuracy. First, the model exhibited a significant increase in error when trained on dataset containing negative values. Second, utilizing batch normalization (BN) layer and squeeze and excitation (SE) module, intended to improve accuracy, resulted in a decrease in performance of the model. Our study demonstrated that the CNN-based model achieves excellent prediction accuracy and has an ability to accelerate optimization algorithms by taking advantage of artificial intelligence’s inherent computational speed.