Predicting Drag Coefficient from Automotive Design Parameters Using Machine Learning

Abstract: This study explores the application of machine learning (ML) to predict the drag coefficient (Cd) of vehicles using design parameters, offering an alternative to conventional wind tunnel and computational fluid dynamics (CFD) methods. Using the publicly available DrivAerNet dataset, which contains over 4,000 vehicle configurations, the research focused on the 15 most influential geometric parameters identified through feature importance analysis. Multiple ML models were implemented, beginning with linear and polynomial regression to establish interpretable baselines, and extending to neural networks for capturing nonlinear aerodynamic relationships. Results show that while linear regression provided quick and transparent estimates, polynomial regression achieved moderate improvements, and tuned neural networks consistently delivered the highest predictive accuracy. The findings underscore the potential of ML to serve as a scalable, cost-effective, and efficient tool for aerodynamic assessment during the early stages of vehicle design. By reducing reliance on expensive simulations and physical testing, ML models enable engineers to

accelerate design workflows, optimize aerodynamics, and advance the development of energy-efficient vehicles. Future work may involve integrating 3D car mesh data and advanced deep learning methods such as graph neural networks or transformers to further enhance prediction accuracy and applicability.