Efficient Automatic Vehicle Shape Determination Using Neural Networks and Evolutionary Optimization

Anton Lundberg (Chalmers University)

CHALLENGE - Car manufacturers need to apply new techniques in order to further reduce emissions from vehicles. Improving the aerodynamic shape of a vehicle holds a large potential for cuts in emissions. This study attempts to optimize vehicle shape using neural networks and evolutionary optimization.

SOLUTION - The proposed method enables a study of several design parameters to be carried out in a short period of time, requiring the construction of morphing boxes as the only manual work, with everything else automated. A solver approximation was used instead of a real solver to cut computational time. A database is generated from simulations on a number of vehicle shape configurations. These are chosen based on a Latin hypercube sampling and the database is used to train a neural network to act as an approximation to the simulations. The optimal vehicle shape is determined from the neural network using particle swarm optimization. The method was incorporated in an optimization tool compatible with Volvo Car Group's CAE process. 








BENEFITS - The optimization tool modeFRONTIER was used on a simplified low-drag car model in a study of realistic changes of five design parameters. Results show that an improved shape with a 12.6% lower drag coefficient (CD) was achieved. The prediction error of DC was 0.3%.