A numerical study of sub-grid scale stresses in a turbulent channel flow using deep learning

Multilayer perceptrons (MLPs), a type of fully connected neural network (FCNN), have been developed to model subgrid-scale (SGS) stress in a turbulent channel flow. The MLP-based SGS model is trained using filtered data with a top-hat filter of size applied at a Reynolds number of and a friction Reynolds number of. In this study, we perform an a priori analysis of a data-driven approach for SGS modeling in a three-dimensional turbulent channel flow. The FCNN model employs resolved flow statistics, including filtered velocity gradients and wall distance, to estimate SGS-stresses. Our data-driven closure model is based on localized learning and utilizes an FCNN architecture with a point-to-point mapping framework. In an a priori test, the model achieved a correlation coefficient exceeding ۹۵% for most predictions, except for and, which attained slightly lower values of ۸۶% and ۸۷%, respectively, compared to the true SGS stresses. Finally, we discussed potential strategies for improving prediction accuracy and overall performance.

کلیدواژه ها:

Turbulent channel flow ، FCNN ، SGS-modeling ، a priori analysis