A decomposition method for surrogate models of large scale structures

Analysis is computationally the most expensive part of optimization. Surrogate models, which are approximate but faster statistical models, can be used in place of more precise but more computer-intensive methods like finite element method to improve efficiency. Unfortunately, the surrogate models are limited by the number of model parameters. So large-scale problems cannot be fully defined by a single surrogate model. Furthermore, current domain decomposition methods cannot be used with black-box models. This study presents a novel approach to design thin-walled structures using surrogate models that overcome the curse of dimensionality by a special decomposition method. A parametric panel structure is defined as a building block. An interface is developed to maintain compatibility across the blocks. Finally, an iterative algorithm finds the displaced state using only local information. Three test structures are used to show the convergence of the algorithm for static analysis. In these sample cases, number of steps required for convergence of the error did not change with the number of panels. This approach offers many benefits including automatic design creation and optimization, effective usage of stream processors and model reuse.