(A) study of CFD subcooled boiling constitutive relations improvement using machine learning technique

Abstract

Safety assessments of nuclear power plants depend greatly on the understanding of various accidents. However, it is very difficult to conduct a real-scale nuclear power plant accident experiment, thus one relies often on computational analysis to analyze various accident scenarios. Existing one-dimensional (1D) safety analysis codes have difficulty in simulating the flow including three-dimensional (3D) phenomena. Therefore, an approach to model complex two-phase flows through Computational Fluid Dynamics (CFD) with 3D capabilities is very important for analyzing accident scenarios. This is expected to lead to development of better prevention and mitigation strategies for nuclear power plant accident situations. A subcooled boiling phenomenon is one of the important phenomena for the nuclear power plant safety analysis. To analyze the phenomenon accurately, constitutive relations of the wall boiling model used in commercial CFD codes were developed based on experimental data under high pressure conditions. Therefore, it has limitations in simulating low pressure subcooled boiling conditions. In order to accurately predict the accident situation, it is necessary to improve the constitutive relations. Therefore, in this thesis, the existing constitutive relation applied in CFD is improved by using machine learning technique. The bubble departure diameter, a sub model of the wall boiling model used in ANSYS-CFX, was improved to fit data better under the low-pressure condition by applying machine learning technique. The exiting model and the improved model by machine learning are compared using ANSYS-CFX. From the comparison, it is evaluated how much the prediction accuracy can be improved by using the suggested method. The potential of the machine learning technique to improve CFD is further discussed.