Study on the Applicability of a Machine-Learning Framework to Improve Modeling for the Stratification Phenomenon

Abstract

Abstract In chemical plants, heat storage systems, nuclear power plants, etc., when hot water is injected into a tank containing cold water, mixing of hot and cold fluid is suppressed by buoyancy effects and thermal stratification can occur. This leads to the deterioration of heat storage capability and the effect of thermal stratification can adversely affect the structure. Therefore, a measure for eliminating thermal stratification was constructed and its effectiveness was examined by CFD analysis. Since the fluctuations in the gravity direction are suppressed by buoyancy in the thermal stratification field, the prediction accuracy becomes low when RANS based on the isotropic eddy viscosity hypothesis like the conventional k-ε model is used. Although the above phenomena can be predicted by performing a high-precision analysis such as LES analysis, the analysis cost becomes enormous and it is unsuitable for design evaluation. In this study, turbulent heat flux and anisotropic Reynolds stress were modeled using a Gene Expression Programming (GEP) framework. In the GEP framework of this study, a non-linear Reynolds stress relationship is constructed by adding non-linear terms to the linear Boussinesq approximation and the turbulent Prandtl number is expressed as a function of velocity and temperature gradients. This modeling is based on an LES dataset. RANS analysis was carried out using these two models on a test case different from the training case, and calculation results are compared with experimental test results. As a result, the decay of the thermal stratification phenomenon was reproduced with high accuracy. This paper introduces this new turbulence modeling process.