Automated Metamodeling for Efficient Multi-Disciplinary Optimisation of Complex Automotive Structures

The use of multi-disciplinary optimisation methods (MDO) in the development process of complex automotive structures is often hindered by several problems. The required resources for very expensive simulations such as crash or 3D CFD analyses rapidly exceed the means available – especially whenever many input parameters, disciplines or load cases are involved. Furthermore, we have experienced that it can be difficult to assure stable runs of simulation processes over a longer period of time. As a result, ‚trivial’ problems such as missing licenses, an overload in network or hard disk resources can lead to a termination of the optimisation process. Not to mention that an optimisation run based on different disciplines can only start once all disciplines involved have set up their respective simulation models. Even a simple change in only one affected discipline would necessitate the optimisation run to start from scratch (with simulations for all load cases/disciplines to be redone). Here, metamodeling techniques can lead to a significant increase in efficiency since all information on the system behaviour gained from former analyses can be reused e.g. for optimisation runs or sensitivity analyses. In addition to this data storage functionality, the use of metamodels also decouples the occupation of computing resources from the actual use of the information. That means that idle CPU time can be used to collect more information on the product or system leading to reduced computation times in the actual optimisation method. Problems in particular simulation runs do not automatically result in a termination of the MDO method, but can easily be repeated. Consequently, it is also possible to evaluate the different disciplines independently even when other disciplines cannot provide a final simulation model yet. All these advantages together result in a much more efficient usage of computation resources. However, the complexity and diversity of metamodeling techniques often prevent the potential user from these benefits. Typically, the choice between the different metamodel formulations is not easy to make. In this paper, an approach is presented which allows for an automated model selection and fitting process. This approach enables the user to use metamodels rid of the complicated selection and fitting process. This task is undertaken by an optimisation algorithm which automatically generates a large variety of metamodels and accesses their respective applicability by means of statistics. As a result, the user gets the most suitable metamodel for each load case or discipline individually and in addition important information about the accuracy of the approximation. The approach will be illustrated by a typical example of a multi-disciplinary optimisation of automotive structures.