Abstract
Convolution neural networks (CNNs) are well-suited to model the nonlinear relationship between the microscale geometry of porous media and the corresponding flow distribution, thereby accurately and efficiently coupling the flow behavior at the micro- and macroscale levels. In this paper, we have identified the challenges involved in implementing CNNs for macroscale model closure in the turbulent flow regime, particularly in the prediction of the drag force components arising from the microscale level. We report that significant error is incurred in the crucial data preparation step when the Reynolds averaged pressure and velocity distributions are interpolated from unstructured stretched grids used for large eddy simulation (LES) to the structured uniform grids used by the CNN model. We show that the range of the microscale velocity values is 10 times larger than the range of the pressure values. This invalidates the use of the mean squared error loss function to train the CNN model for multivariate prediction. We have developed a CNN model framework that addresses these challenges by proposing a conservative interpolation method and a normalized mean squared error loss function. We simulated a model dataset to train the CNN for turbulent flow prediction in periodic porous media composed of cylindrical solid obstacles with square cross-section by varying the porosity in the range 0.3 to 0.88. We demonstrate that the resulting CNN model predicts the pressure and viscous drag forces with less than 10% mean absolute error when compared to LES while offering a speedup of O(10 6 ).