Increasing automobile crash response metamodel accuracy through adjusted cross validation error based on outlier analysis

Abstract

Automakers spread on effort to maintain the crashworthiness of vehicle structures while aiming to reduce their weight. Substantial weight savings can be obtained by vehicle redesign through optimisation. Finite element based crashworthiness simulation models have contributed greatly to the optimisation of vehicle structures. These high-fidelity crash simulations may be performed many times during optimisation, thereby making optimisation studies computationally intractable. Metamodels (surrogate models) that can mimic the behaviour of the crash simulation models emerge as a solution to the computational burden. Prediction capability in metamodelling can be improved by combining many different metamodels in the form of an ensemble model. In this paper, approaches based on outlier analysis of cross validation errors are proposed to increase the accuracy of ensemble models constructed for crash response predictions. Full frontal and offset frontal crash response predictions of a c-class passenger car is used for demonstration, and it is found that the proposed approach reduces the metamodelling errors up to 12% and on average by about 4.5%.