Aerodynamics plays a key role in the engineering design of ground vehicles, with every vehicle undergoing numerous design iterations to optimize its performance. While physical testing is the ultimate arbiter of performance, its high costs in time and money have caused the industry to turn to Computational Fluid Dynamics (CFD) as an alternative for design development. CFD enables engineers to rapidly iterate through multiple geometry designs and extract comprehensive insights directly from the flow field at a fraction of the time and monetary cost of a single physical test. Ubiquitous adoption of CFD has led to a highly structured engineering pipeline consisting of three main components: code development, methodology development, and flow analysis-informed geometry design. Despite this widespread use, there is a severe lack of feedback tools to enhance each stage of the process, forcing engineers to rely on ad-hoc rules and generational knowledge. To address this critical gap, this dissertation develops a suite of accessible, Machine Learning-driven (ML) feedback frameworks designed to enhance user analysis and overall process efficiency across all three primary stages of the CFD pipeline. First, the code development phase is addressed by applying explainable ML to interpret the complex input-output relationships of the Reynolds-Averaged Navier-Stokes (RANS) SST $k-\omega$ turbulence model. Using the Ahmed body as a test case, Shapley Values and the SHAP method were used to quantify the influence of individual closure coefficients on the lift and drag predictions. This explainability framework provided direct, interpretable feedback on model behavior, enabling the targeted tuning of closure coefficients that reduced force prediction errors to less than 7\% and 0.5\% for the 25-degree and 40-degree slant configurations, respectively. Proceeding to the methodology stage, an ML-based framework was developed to aid in spatial domain discretization assessment. An unsupervised ML clustering algorithm and a novel similarity scoring approach were developed to allow users to visually identify flow regions that have achieved mesh independence. To develop and validate the tool, three canonical flows were utilized: a zero-pressure-gradient flat plate, a bump in a channel, and an axisymmetric free jet. Across all three cases, the calculated similarity scores were found to agree with traditional flow field probes regarding the level of mesh refinement required for convergence. The approach also successfully recommended spatial refinement in flow regions that align with established fluid dynamic intuition. Finally, to address the design stage of the CFD pipeline, the similarity score framework was adapted for geometry design. This tool seeks to help address the underutilization of data in large-scale geometric sweeps. To test the similarity score on design studies, a full-factorial design space was generated for two 2D airfoils: one in free air and one in ground effect. The algorithm's spatial regionalization was modified to utilize radial clustering, and the similarity score was recalibrated to incorporate both spatial distance and shape similarity. The framework's utility was then demonstrated through two example cases. The first case paired the refined scoring method with an ML outlier-detection model, demonstrating how anomalous flow fields can be automatically isolated. The second case demonstrated how the method can simultaneously compare geometries across multiple operating conditions, quickly highlighting flow regions of interest for further engineering investigation. Ultimately, this dissertation demonstrates how ML techniques can be seamlessly integrated into existing engineering workflows. This work shows that by leveraging ML, the paradigm of CFD development can shift from a reliance on ad-hoc rules to the utilization of intelligent feedback tools, empowering engineers to make data-driven decisions at every stage of the pipeline.
