INTERPRETATION AND PREDICTION FOR PRANDTL NUMBER EFFECT IN TURBULENT HEAT TRANSFER USING GENERATIVE ADVERSARIAL NETWORKS

We propose a interpretable deep learning model that embedded effect of physical parameter in turbulence. Turbulence is a very complex flow, and the analysis of relationships between turbulent variables remains a fundamental challenge. Recently, studies applying deep learning are being conducted in attempts to analyze turbulence. Deep learning can extract physics features in data, which is turbulent analysis model to understand the physical relationship between variables in turbulent flows. In this study, we consider turbulent heat transfer to extract and analyze the effect of Prandtl number (Pr) in the data. The deep learning model uses conditional generative adversarial networks (cGAN) with decomposition algorithm. Our model predicts surface heat flux with various Pr from wall shear stresses in channel flow. The predicted surface heat flux reflected the characteristics with respect to Pr well, and also was statistically very similar to DNS. We analyzed the spatial nonlinear relationship between wall shear stresses and surface heat flux for Pr through gradient maps of trained our model. Furthermore, for analysis of effect of Prandtl number, we observed decomposed field into universal and Pr-dependent features based on turbulent data sets. Through interpretation of the deep learning model, it is possible to understand the physical interaction between variables, which can help to develop a turbulence model considering physics.