Prediction method of reactor core 3D neutron flux based on ARGA-3D CNN

<bold>Background</bold>The three-dimensional prediction of neutron flux is crucial for the design, optimization, and safety analysis of reactor cores. However, due to the compact spatial configuration of small lead-cooled fast reactors and the difficulty in placing detectors, existing methods mainly focus on two-dimensional analysis, with limited attention to three-dimensional flux prediction.<bold>Purpose</bold>This study aims to propose a model called the Genetic Algorithm-Enhanced 3D Convolutional Neural Network with Multi-Head Self-Attention and Residual Connections (ARGA-3D CNN), which integrates Residual Networks (ResNet) and Multi-Head Self-Attention (MSA) mechanisms, to efficiently capture the spatial distribution features of neutron flux in reactor cores and address the issue of spatial dependencies.<bold>Methods</bold>The 3D convolutional neural networks (3D CNN) within this ARGA-3D CNN model were employed to effectively capture the spatial distribution of neutron flux in the reactor core, and the ResNet was used to alleviate the vanishing and exploding gradient problems, improving training stability, whilst the MSA was applied to enhance the identification of key regions. In addition, the Genetic Algorithm (GA) was used to optimize hyperparameters for further improving the accuracy of neutron flux prediction. Finally, this model was trained and optimized using a dataset constructed from results of Monte Carlo simulation software SuperMC, and thereafter applied to predicting 3D neutron flux of a small lead-cooled fast reactor.<bold>Results</bold>The prediction results show that the predicted values of the ARGA-3D CNN model, compared to the SuperMC simulation results, achieve Mean Absolute Error (MAE), Mean Squared Error (MSE), and <italic>R</italic>² values of 3.19×10^-6, 2.14×10^-11, and 0.973 5, respectively. The computational efficiency is significantly improved, with each prediction taking only seconds. Compared to traditional models such as Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM) networks and Transformer, the prediction performance is superior.<bold>Conclusions</bold>The ARGA-3D CNN model proposed in this study provides a high-precision and computationally efficient method for predicting three-dimensional neutron flux in a reactor core. It offers a new approach for the rapid prediction of nuclear reactor core parameters, which has significant practical value and implications for reactor design, operation, and safety analysis.