This dissertation presents a novel neural-network (NN)-based subgrid-scale (SGS) modeling approach for large eddy simulation (LES) and its application to flows over complex geometries. The research is structured in two major parts: first, developing a...
This dissertation presents a novel neural-network (NN)-based subgrid-scale (SGS) modeling approach for large eddy simulation (LES) and its application to flows over complex geometries. The research is structured in two major parts: first, developing and validating an NN-based SGS model for flow over a circular cylinder, and second, extending this approach to more complex geometries.
In the first part, fully-connected neural networks are constructed to predict SGS stresses for flow over a circular cylinder at Reynolds number 3900. A new NN architecture is proposed that integrates both grid and test-filtered variables as inputs through a fusion process. The performance of this architecture is compared with traditional models and NN architectures without fusion process through both a priori and a posteriori tests. Results demonstrate that models utilizing grid- and test-filtered inputs with fusion (T-SR-FU and T-VG-FU) significantly outperform traditional models such as dynamic Smagorinsky model (DSM), particularly when applied to grid resolutions and Reynolds numbers (5000 and 10000) different from training conditions.
In the second part, the NN-based SGS model is enhanced for application to more complex and untrained geometries. The generalizablity of the NN-based model is improved with key design features including: 1) using more training datasets from diverse flows such as turbulent channel flow and flow over a circular cylinder, 2) using input features derived from traditional turbulence models and 3) excluding bias terms and batch normalization to ensure homogeneity while preserving non-linearity. The enhanced model is tested on both trained and untrained flows (backward-facing step flow at Reh = 5100 and flow over an SD7003 airfoil at Rec = 60000). Results show remarkable generalizability across different flow configurations, accurately predicting critical flow parameters such as recirculation length, pressure recovery, and transition processes.
This research demonstrates that NN-based approaches, when carefully designed with appropriate physical constraints and generalizable features, offer a promising direction for developing advanced SGS models for complex turbulent flows. The models show superior ability to capture complex flow phenomena, making them valuable tools for engineering applications involving complex geometries.