Shape Optimisation of Vehicles for Crosswind Stability Using Neural Networks
Examensarbete för masterexamen
Recent advances in computational power and the availability of large labelled data sets have contributed to the progress and interest in deep learning. One product development area associated with vast amounts of data is aerodynamics and previous studies have successfully applied neural networks to predict aerodynamic coefficients. The aerodynamic development process for a passenger vehicle usually focuses on the aerodynamic drag force and the lift force, with air flow coming head on. Fewer efforts are spent on investigating how all aerodynamic forces and moments are affected by strong crosswinds. It is excepted that the aerodynamic lift and side force together with the yaw moment affect driving stability. Furthermore, driving stability issues are often discovered late in the development process and methods for assessing this issue virtually are desired. Thus, a method for investigating the optimal shape changes of a passenger vehicle to minimise crosswind sensitivity was developed. Simulated data was created by generating 120 CAD models, each with a unique configuration of five design parameters: amount of corner filler below the roof spoiler, length and height of roof spoiler and the width and length of the back-light. A CFD simulation was performed on each configuration with a flow-angle of 7.5 and a driving velocity of 100 km/h. Three neural networks were trained to predict the drag, rear lift force and the yaw moment, respectively. The effect of each design parameter was studied, as well as the optimal configuration for minimising each coefficient. A genetic algorithm was used to find the settings for minimising yawmoment and rear lift force, while retaining other coefficients within set requirements. The final design was estimated to decrease the yaw moment by 4.15 %, but due to the small variation of yaw moment in the generated data set, the results are uncertain. The optimal design to minimise CLR achieved a decrease of 0.0185 in rear lift.
neural networks , CFD , stability , yaw moment , rear lift force , genetic algorithm