Currently the development of suspension layouts in cars relies heavily on multibody simulation models. Those simulations test the suspension during a maneuver and can last from minutes to hours. Optimization algorithms need thousands of simulations, so this combination makes this a computationally intensive approach. All the above leads to the following question: Is there a way to replace the computationally expensive high-fidelity multibody simulation model with an accurate enough metamodel so designers can efficiently optimize suspension design? A high-fidelity multibody simulation model was used to produce all the necessary data. In total, 94 initial parameters with their corresponding boundaries were defined as input for both the multibody simulation and the metamodel. An ISO steering maneuver was programmed, and the roll angle of the vehicle was chosen to be the main output. The large number of input parameters led to an extra sensitivity analysis step to identify parameters that could be omitted. The following Design of Parameters (DoE), using the Latin hypercube sampling with the max-min criterion. After creating the design space, a metamodel was built. The relatively high number of data points made a neural network a good fit. To select the hyperparameters, an automated machine learning technique, based on Bayesian optimization, was employed. A particle swarm optimizer was used in the end to find an optimum design with respect to the L2-norm of the roll angle.
Session: VEHICLE DYNAMICS & SIMULATION | | 12:30 - 13:00