난류 유동에서 다변수 인자에 대한 기계 학습 기법 비교 연구

Abstract

It is important to predict spatially varying parameters to model turbulent flows. In this study, the spatially varying parameters are modeled via machine learning techniques using experiment-based turbulent bubble flow data and DNS-based turbulent Prandtl number data. The prediction and generalization errors of machine learning models are evaluated, and the different techniques are compared. Among the artificial neural network (ANN) techniques, the regular ANN using the full-batch training and the stochastic gradient descent (SGD) ANN based on mini-batch training are compared with the random forest (RF) method. The prediction and generalization errors show different characteristics according to the data resolution. For the coarsest bubbly flow data set, SGD ANN shows stable training and prediction, which leads to the smallest prediction and generalization errors. For the data sets with a finer resolution, the generalization error of SGD ANN is smallest, whereas the average prediction errors of regular ANN and RF method are smaller compared to SGD ANN. When evaluated using the trained models, all machine learning techniques show similar spatial distribution to the original data.