Data-Driven Geometric Representation and Flow-Field Prediction in Industrial Fluid Dynamics

High-fidelity computational fluid dynamics (CFD) simulations require significant resources, especially in the early stages of vehicle aerodynamic design when many shape variations are assessed. Data-driven surrogate models can lower these computational costs, but their performance often depends on the geometric representation used during training, such as fixed design parameters, a specific computational mesh, or a particular spatial discretization. As a result, it is difficult to explore a wide range of designs or use the same surrogate models across different vehicle types and design processes. A two-stage aerodynamic surrogate framework is developed based on an explicit signed distance field (SDF) representation of geometry. In the first stage, geometry is reconstructed using a compact latent representation. For AhmedML, a parameter-driven branch maps design variables to the latent space, while a volumetric encoder offers a geometry-driven alternative based directly on the SDF. Both branches utilize a shared geometric decoder. To accommodate more realistic vehicle configurations, DrivAerML and Siemens geometries are processed through the volumetric SDF approach. This methodology enables both public and industrial vehicle geometries to be represented through a unified spatial interface without the need for a shared low-dimensional design parameterization. Stage~2 maps the reconstructed or reference signed distance function (SDF) to three-dimensional aerodynamic fields using a volumetric U-Net. The direct configuration is trained with reference SDFs, while the end-to-end configuration is trained with geometries reconstructed by Stage~1. Stage~1 is trained initially and then kept fixed, ensuring that the aerodynamic model is exposed during training to the same type of geometric input as during end-to-end inference. Consequently, the two stages are optimized sequentially rather than through joint backpropagation. Drag is modeled separately using a scalar regression approach when appropriate supervision is available. The framework is evaluated using geometries of increasing complexity, including analytical shapes, AhmedML, DrivAerML, and industrial Siemens vehicle configurations. Results indicate that the geometric stage reconstructs vehicle representations with high volumetric and surface agreement prior to aerodynamic prediction. The subsequent field-prediction stage maintains the principal flow structures across all datasets. Comparison between direct and end-to-end configurations enables identification of the impact of geometric reconstruction on downstream aerodynamic accuracy. Transfer experiments further demonstrate that the shared SDF interface allows adaptation of the staged methodology from benchmark geometries to realistic and industrial vehicle families without dependence on a common design-variable space. The primary contribution of this thesis is methodological. The proposed framework distinguishes geometry reconstruction from aerodynamic prediction by introducing an explicit, independently testable geometric interface while maintaining a consistent volumetric representation across diverse vehicle families. This distinction allows for independent and joint analysis of geometric and aerodynamic errors throughout the entire design-to-geometry-to-flow process. By integrating both parameter-driven and geometry-driven entry paths with a unified signed distance function (SDF) representation, the framework supports rapid aerodynamic evaluation and facilitates broader exploration of vehicle design spaces where a single common parameterization is not feasible.