The development of high-throughput reactors is of great significance for supporting the development of nuclear science and technology, improving the efficiency of nuclear energy utilization, meeting the needs of radioactive isotope production, and carrying out irradiation tests and performance tests of new nuclear fuels and structural materials in reactors. Due to the high power density of the core fuel and the large demand for thermal cooling, the nuclear-thermal coupling phenomenon in the high-throughput lead-bismuth reactor (HT-LBR) is more significant than that in conventional lead-bismuth reactor (LBR). When the design optimization of high flux LBR is carried out, it is necessary to carry out collaborative optimization of multiple core parameters, improve the neutron flux density, and meet the physical / thermal constraints such as core refueling period, fuel cladding temperature and coolant flow rate. Therefore, the design optimization of high flux lead-bismuth cooled reactor is a complex problem of multi-physics, multi-variable and multi-constraint coupling.

This study aims to improve the neutron flux level of LBR and solve the optimization design problem of HT-LBR.

Firstly, a HT-LBR training database was constructed to contain different core design parameter combinations and corresponding objective function response values and constraint condition response values. Based on the reactor Monte Carlo code RMC and sub-channel Code Subchanflow, a Back-Propagation (BP) neural network prediction model was established as a proxy model for reactor physical calculation and analysis to achieve rapid prediction of core neutron flux density and effective multiplication factor using aforementioned training database. Secondly, an updated iterative optimization method based on BP neural network Dynamic Surrogate Model (DSM) was proposed to improve the optimization efficiency and global optimization ability, and search for the optimal HT-LBR core design parameter combination within the design range. Thirdly, based on the open-source machine learning platform TensorFlow, coupled with the reactor physical and thermal calculation and analysis program, an iterative optimization method based on BP neural network prediction model was proposed. Combined with the sensitivity analysis method of core design parameters based on Sobol index method, a HT-LBR optimization design platform was developed to cover five functional modules: training database generation, physical and thermal parameters calculation and analysis, BP neural network model construction, core parameters sensitivity analysis, and core parameters optimization analysis. Finally, a multi-functional ultra-high-throughput reactor was used as a prototype to establish a model to be optimized, collaborative optimization verification of multiple core parameters including core grid diameter ratio, fuel pellet diameter, active zone height, and radial reflector thickness, was conducted.

Verification results show that the prediction accuracy errors for core neutron flux density and effective multiplication factor are maintained within 0.1%. The optimized neutron flux density is 15.41% higher than the original design. The influence degree of the four groups of core design variables on the maximum neutron flux is arranged in the order of reflector thickness < gate diameter ratio < active zone height < fuel pellet diameter. At the same time, the maximum temperature of the fuel pellet and the maximum temperature of the cladding are reduced by 23.57 ℃ and 8.20 ℃, respectively. The optimized core design scheme has a larger steady-state thermal safety margin.

The HT-LBR optimization design platform developed in this paper is effective and reliable.