Critical heat flux prediction for narrow rectangular channels: a machine learning model based on the Kalman filter

High-flux research reactors, operating as non-power reactors with neutron fluxes not less than 10^14 /cm^2/s, play a vital role in scientific research, material property testing, and the production of radionuclides for medical and industrial applications. High-flux reactors predominantly employ plate-type fuel assemblies, featuring high core power density and more complex coolant flow conditions. Consequently, accurate CHF prediction becomes a critical element in core safety design and operation. Traditional CHF prediction methods exhibit significant prediction errors and limited applicability when addressing real-world scenarios involving complex geometries and non-uniform thermal loads. Therefore, to enhance the prediction accuracy of CHF in narrow rectangular channels of plate-type fuel elements, this paper proposes a machine learning model combining Kalman filter with BP neural network. This model utilizes a CHF look-up table (LUT) database as its foundation. It employs an Extended Kalman Filter (EKF) to dynamically correct and fuse CHF-LUT data, then trains a BP neural network to achieve high-precision prediction of CHF values in narrow rectangular channels. To validate the effectiveness of EKF-ML model, the prediction results are compared with those obtained directly from the CHF-LUT and the Sudo correlation. Results demonstrate that the EKF-ML model significantly enhances the prediction accuracy of CHF in narrow rectangular channels based on the CHF-LUT. The mean relative error reaches 0.78%, and the relative root mean square error (rRMSE) decreases to 12.48% under specified operating conditions. This study provides new insights and methodologies for predicting CHF values in narrow flow channels of high-flux reactor cores.