Gaussian Process Surrogate Models for Vibroacoustic Simulations

In vehicle Noise Vibration Harshness (NVH) development, vibroacoustic simulations with Finite Element (FE) Models are a common technique. The computational costs for these calculations are steadily rising due to more detailed modelling and higher frequency ranges. At the same time the need for multiple evaluations of the same model with different input parameters - e.g., for uncertainty quantification, optimization, or robustness investigation - is also increasing. Therefore, it is crucial to reduce the computational costs dramatically in these cases. A common technique is to use surrogate models that replace the computationally intensive FE model to perform repeated evaluations with varying parameters. Several different methods in this area are well established, but with the continuous advancements in the field of machine learning, interesting new methods like the Gaussian Process (GP) regression arises as a promising approach. In Gaussian Process regression there are important parameters that strongly influence the prediction accuracy of the GP Model, namely length-scale, variance, and mostly the kernel function. In this contribution these parameters and their influence on the results are evaluated, with a focus on vibroacoustic simulations. For the kernel function, four different types - stationary, nonstationary, spectral and deep learning kernel, respectively - are under investigation. As a result, it can be shown that their performance corelate with the data complexity. Further investigations focus on the frequency as input parameters and the influence of the number of training samples. In these evaluations there is an interesting difference between a simple academic model and a body in white model. The underlying effects, such as damping, system complexity, uncertainty and load case are discussed in detail. Finally, a recommendation using GP as a surrogate model for vibroacoustic simulations is given.