Interactive Aerodynamic Design via Reduced Order Modeling

Abstract Aerodynamics plays an essential role in the transition process towards more eco-friendly fleets. The development of an aerodynamic yet stylistically appealing vehicle requires close cooperation between aerodynamicists and stylists – a cooperation currently hindered by the enormous computational effort of aerodynamic simulations. Joint, interactive aerodynamic design sessions that exploit real-time simulations via Reduced Order Models (ROM) have the potential to improve efficiency. Major advances were made in three core building blocks of an interactive aerodynamic design process. First, to speed-up training data generation, simulations with the Lattice Boltzmann method (LBM) on Graphical Processing Units were investigated with the solver Altair ultraFluidXTM. The aerodynamic coefficients from ultraFluidX agreed excellently with measurements for a Volkswagen Golf 7. The numerically predicted flow fields, however, still showed deficiencies in the details. While the accuracy was comparable with that of Detached-Eddy simulations (DES), the turnaround times were reduced significantly. Because of that, the creation of typical training data sets is possible within one working week – a timescale well-suited for the systematic integration of Reduced Order Models in the development process. The main objective was to assess the prediction accuracy for aerodynamic fields and coefficients for real-life industrial test cases. Two test cases were created: 100 DES simulations for a Volkswagen Jetta with six geometrical parameters and 1000 LBM simulations for a DrivAer with 15 geometrical parameters (one Boolean). Based on this training data, the solution space was reduced via Proper Orthogonal Decomposition (POD), and the obtained low-dimensional representation was interpolated via Response Surface Models (RSM) to predict the flow field at unseen parameter combinations (so-called POD+I approach). For aerodynamic coefficient predictions, the accuracy was satisfactory for both test cases with mean average errors of about 1 and 2 drag counts for the Jetta and DrivAer, respectively. For the field predictions, POD+I yielded results that are qualitatively and quantitatively in good agreement with CFD results. The wall clock time for the combined prediction of fields and drag coefficient was measured to be approximately 0.6 s (Jetta), making it suitable for use within an interactive aerodynamic design process. Third, the feasibility of two ROMs based on nonlinear dimension reduction – Isomap+I and neural network Autoencoder+I – was evaluated. Isomap+I was outperformed by POD+I in the field predictions while also being considerably slower. Autoencoder+I, on the other hand, slightly surpassed POD+I in prediction accuracy for both the Jetta and DrivAer; however, the marginal gain came at the cost of much greater training times. Reduced Order Modeling via POD+I turned out to be the most appropriate method for an interactive aerodynamic design process; the achievable accuracy is considered as sufficiently accurate for adoption in the development process. The innovative design process developed in this thesis enables aerodynamicists and stylists to cooperate closely and develop vehicles with very good aerodynamics and compelling designs.