Crashworthiness Design Using Meta-models for Aprroximating the Response of Structural Members

Vehicle crashworthiness is an important yet difficult design attribute. It can be argued, that an assisting task for the design process would be to pre-compile library databases for the crash behavior of elementary structural members. Such databases would guide designers to quickly select the appropriate dimensions of structural members once estimates of the load requirements are available. Unfortunately, typical structural member cross-sections in the automotive body have complex geometrical shapes and no standard sizes. Thus, generating crash behavior databases for all possible dimensions of even a single shape is an extremely expensive task. A remedy is to generate databases containing relatively few sizes of the structural members, then employ meta-models to approximate the behavior of the structural members over the complete design domain of the dimensions. This paper investigates various meta-modeling techniques for estimating the behavior of structural members subjected to crash conditions. The examined meta-modeling techniques are: i) polynomial response surface fits, ii) cascaded feed forward neural networks and iii) radial basis neural networks. Results of the study show the superiority of the radial basis networks. Furthermore, a design example of a vehicle B-pillar subjected to roof crush is presented to demonstrate the advantage of precomplied database and meta-model component selection verses direct optimization of a full structure.