Reconstruction of turbulent channel flow from wall data using generative adversarial networks
This thesis explores the use of deep generative models to reconstruct three- dimensional turbulent channel flows from sparse wall-based information. The proposed model will be tested on a canonical turbulent channel flow at fric- tion Reynolds number Reτ = 180 simulated with the RHEA solver. This co...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/444129 |
| Acceso en línea: | https://hdl.handle.net/2117/444129 |
| Access Level: | acceso abierto |
| Palabra clave: | Aerodynamics Deep learning (Machine learning) Turbulence--Mathematical models Aerodinàmica Aprenentatge profund (Aprenentatge automàtic) Turbulència--Models matemàtics Àrees temàtiques de la UPC::Enginyeria mecànica |
| Sumario: | This thesis explores the use of deep generative models to reconstruct three- dimensional turbulent channel flows from sparse wall-based information. The proposed model will be tested on a canonical turbulent channel flow at fric- tion Reynolds number Reτ = 180 simulated with the RHEA solver. This configuration provides well-established reference results that can be used to assess and validate the reconstruction performance. The framework is based on a Generative Adversarial Network (GAN) architecture, where a U-Net generator captures multiscale structures through skip connections, while a PatchGAN discriminator with batch normalisation enforces local physical realism and stabilises the adversarial learning process. The proposed approach demonstrates that data-driven models can recover coherent near-wall streaks as well as outer-layer motions, providing recon- structions that are both physically consistent and visually faithful to the ref- erence flow. The methodology highlights the capability of adversarial train- ing to infer two-dimensional velocity fields from reduced sets of inputs, which can be stacked across wall-normal locations to form a volumetric reconstruc- tion, thereby bridging the gap between high-fidelity simulations and practical sensing limitations. Beyond canonical channel flows, this research emphasises the relevance of such tools for applied aerodynamics. In automotive engineering and motor- sport, the ability to reconstruct turbulence from sparse measurements opens new possibilities for drag reduction, flow control, and efficiency improve- ments, with direct implications for sustainability and performance optimisa- tion in vehicle design. |
|---|