Response Surface Simulation - Patchwork Application

The repeated performance of structural analysis, e.g., within a reliability based optimization, for realworld problems require a high computational effort. Due to the complexity of the real-world investigations, which are based on sophisticated Finite Element analyses of large nonlinear systems, a particularly efficient form of an approximation scheme is required. In this paper the improvement of the numerical efficiency utilizing neural network based approximation schemes are discussed. Therefore the theoretical basics of neural networks are introduced. A further improvement of the approximation quality is obtained with network based approximation schemes. Beside committee machines and network composites, a section-wise application of neural networks is presented. The developments are demonstrated by means of numerical examples to emphasize their features and by a practical, industrial-sized example to underline their applicability.

application/pdf K-I-02.pdf — 2.3 MB