Metamodeling for crashworthiness design: Comparative study of kriging and support vector regression
Maliki Moustapha, Bruno Sudret, Jean‐Marc Bourinet, Guillaume, Benoît · Repository for Publications and Research Data (ETH Zurich) · 2014
Abstract. The use of metamodels as surrogates of time-consuming functions has widely spread within the academia and the industry. In this paper, two metamodels are considered for crashworthiness design of an automotive body structure, namely Kriging and support vector regression. Variants of these two metamodels associated with various kernel or auto-correlation functions are first compared on analytical functions. The conclusions of this benchmark analysis are then considered to select the most appropriate ones for application on the so-called sidemember subsystem. This is a subsystem of an automotive front end under frontal impact. The outputs to emulate are highly non-linear and noisy. The SVR and Kriging models are shown to produce roughly the same level of accuracy for prediction when considered with isotropic kernels or auto-correlation functions, with a slight advantage to Kriging. Besides, the anisotropy in the auto-correlation functions clearly improves the Kriging surrogates. Some outputs of the crash simulation were however hard to surrogate.