Optimization of Vehicle Structures under Uncertainties

Advancements in simulation tools and computer power have made it possible to incorporate simulation-based structural optimization in the automotive product development process. However, deterministic optimization without considering uncertainties such as variations in material properties, geometry or loading conditions might result in unreliable optimum designs. In this thesis, the capability of some established approaches to perform design optimization under uncertainties is assessed, and new improved methods are developed. In particular, vehicle structural problems which involve computationally expensive Finite Element (FE) simulations, are addressed. The first paper focuses on the evaluation of robustness, given some variation in input parameters, the capabilities of three well-known metamodels are evaluated. In the second paper, a comparative study of deterministic, reliability-based and robust design optimization approaches is performed. It is found that the overall accuracy of the single-stage (global) metamodels, which are used in the above study, is acceptable for deterministic optimization. However, the accuracy of performance variation prediction (local sensitivity) must be improved. In the third paper, a decoupled reliability-based design optimization (RBDO) approach is presented. In this approach, metamodels are employed for the deterministic optimization only while the uncertainty analysis is performed using FE simulations in order to ensure its accuracy. In the fifth paper, two new sequential sampling strategies are introduced that aim to improve the accuracy of the metamodels efficiently in critical regions. The capabilities of the methods presented are illustrated using analytical examples and a vehicle structural application. It is important to accurately represent physical variations in material properties since these might exert a major influence on the results. In previous work these variations have been treated in a simplified manner and the consequences of these simplifications have been poorly understood. In the fourth paper, the accuracy of several simple methods in representing the real material variation has been studied. It is shown that a scaling of the nominal stress-strain curve based on the Rm scatter is the best choice of the evaluated choices, when limited material data is available. In this thesis work, new pragmatic methods for non-deterministic optimization of large scale vehicle structural problems have been developed. The RBDO methods developed are shown to be flexible, more efficient and reasonably accurate, which enables their implementation in the current automotive product development process