서강대학교

초록/요약

For modeling of turbulent flow and heat transfer phenomena, this study aims to analyze and present efficient and stable modeling techniques based on artificial neural network (ANN) and parameter selection. Among existing approaches, we consider a few filter and wrapper methods using ANN and a few widely used methods (e.g. principal component analysis, genetic algorithm). Considering a few features desirable for physical modeling and ANN-based models, we focus on the wrapper parameter selection and intend to analyze and improve it. In order to examine performance of parameter reduction techniques, we apply to manufactured problems based on random parameters. For validation with fluid phenomena, we apply to the modeling of a turbulent bubbly flow and heat transfer over a wedged wall. We found that wrapper parameter selection methods are more effective than a correlation-based filter parameter selection in terms of removing duplicate or unnecessary parameters. As the model complexity increases, it is observed that the non-linearity between the parameters is detected well, which leads to the appropriate number of selected parameters. However, overfitting occurs when the model complexity increases or noisy data is included, which leads to selection of the parameters unnecessary for modeling. As a remedy, a gradient-based removal is proposed and shown to be more stable in the overfitting environment than the previous methods. For validation studies, RANS simulations based on data-based models are performed for turbulent bubbly flows and heat transfer over a wedged wall to examine the effects of data-driven models (such as two-way coupling effects), which shows improved results compared to previous empirical models.