Multi-scale and multi-model methods for efficient crash simulation

摘要

A new way of crash simulation using explicit codes is presented, in which two distinct finite element models run simultaneously in their own Distributed Memory Processing (DMP) environment and interact in order to compute a solution for the union structure. This Multi-Model approach is very advantageous for analyzing large-scale vehicle crash models used in the automotive industry, which benefits from partitioning the whole finite element structure into a 'global' model and a 'local' model, where the global model contains the bulk of the vehicle structure, and the local model may represent significant parts of the structure for which a detailed deep analysis is desired. The global model is inevitably highly complex and heterogeneous, limiting the scalability. In contrast, the local model can be kept simple and homogeneous, providing the condition for excellent scalability. Model coupling is achieved using direct node-to-node interfaces between the models, accompanied by inter-code contact treatment. As a crucial aspect, the global and local models can have different mesh size scales, and hence different time steps, where the time step ratio of both models can take arbitrary integer values. This Multi-Scale option takes advantage of the subcycling technique, which allows significant savings in computation time. The developed Multi-Model and Multi-Scale methods have been implemented in a commercial explicit crash code. The features of the methodology are presented and applied to solve a typical large vehicle front crash simulation. Different grades of Multi-Scaling are investigated, with special emphasis on the efficiency analysis of different network configurations, including the gigabit Ethernet and the Myrinet 2000 interconnection standard.