Research on the prediction method of reactor core 3D neutron flux based on ARGA-3DCNN

The three-dimensional prediction of neutron flux is crucial for the design, optimization, and safety analysis of reactor cores. However, due to the compact space and difficult detector layout of miniature lead cooled fast reactors, existing methods mostly focus on the two-dimensional level and pay less attention to the prediction of three-dimensional flux. This paper proposes a 3D Convolutional Neural Network (ARGA-3D CNN) model that integrates Residual Network (ResNet) and Multi head Self Attention (MSA) mechanism. The model effectively captures the spatial distribution characteristics of neutron flux in the core and solves the problem of spatial dependence. Relieve gradient vanishing and explosion through ResNet, enhance training stability, and strengthen key region recognition with MSA. In addition, genetic algorithms are used to optimize hyperparameters, further improving the accuracy of neutron flux prediction in the core. The experiment is based on the calculation results of the Monte Carlo particle transport simulation software SuperMC to construct a dataset, and uses this dataset to train and optimize the ARGA-3D CNN model for prediction. The results showed that compared with the SuperMC calculation results, the predicted values of the model reached 3.19 × 10-6, 2.14 × 10-11, and 0.9735 in terms of Mean Absolute Error (MAE), Mean Squared Error (MSE), and Coefficient of Determination (R ²), respectively. The computational efficiency was significantly improved, and a single prediction only took seconds. Compared with models such as Convolutional Neural Network (CNN), Artificial Neural Network (ANN), Long Short Term Memory (LSTM), and Transformer, the prediction performance was better.