NEURAL NETWORK-BASED SURROGATE MODELING FOR POST-PROCESSING OF TOPOLOGY OPTIMIZED STRUCTURES

Abstract

This thesis proposes a method of creating an accurate neural network-based surrogate model for post-processing a topologically optimized structure. Topology optimization using a cell-based method creates disagreements on mechanical measures (e.g., deformation, stress) when converted to computer aided design (CAD) files, which are defined by smooth boundaries. The conversion process is necessary to make a manufacturable geometry, but the process introduces some performance losses and disagreement of mechanical measures. Post-processing, a method of fine tuning the CAD geometry, is needed to recover the original measures from topology optimization. In this thesis, deep artificial neural network (DANN) is presented to create regression models that relate the CAD geometry inputs to multiple stress outcomes. The regression models by DANN are used as a surrogate model to fine tune the CAD model stress using a limited number of FE computations and regenerate the smooth-boundary CAD model that satisfies all the stress design requirements. The performance of DANN based regression model is compared with a second order polynomial fitting with response surface methodology (RSM). The usefulness of this method is verified both in 2D and 3D test studies.