Search Results

Response Surface Models are often used as a surrogate for expensive black-box functions during optimization to reduce computational cost. Often, the CAE analysis models are highly nonlinear and multi-modal. A response surface approximation of such analysis as a result is highly multi-modal; i.e. it contains multiple local optima. A gradient-based optimizer working with such a response surface will often converge to the nearest local optimum. There does not exist any method to guarantee a global optima for non-convex multi-modal functions. For such problems, we propose an efficient algorithm to find as many distinct local optima as possible. The proposed method is specifically designed to work in large dimensions (about 100 ∼ 1000 design variables and similar number of constraints) and can identify most of the locally optimal solutions in a reasonable amount of time.

An interactive tool for predicting the performance of vehicle designs in the pedestrian leg impact test has been developed. This tool allows users to modify the design of a vehicle front structure through the use of a graphical interface, and then evaluates the performance of the design with a response surface. This performance is displayed in the graphical interface, providing the user with nearly instantaneous feedback to his design changes. An example is shown that demonstrates how the tool can be used to help guide the user towards vehicle designs that are likely to improve performance. As part of the development of this tool, a simplified, parametric finite element model of the front structure of the vehicle was created. This vehicle model included eleven parameters that could be adjusted to change the structural dimensions and structural behavior of the model.