Optimal design and failure prediction using neural networks

Abstract

Abstract—Topology optimization is the tool of choice in obtaining the initial design of structural components. The resulting optimal design from topology optimization will be the input for subsequent structural optimizations with regard to shape, size, and layout. In reality, however, iterative solvers used in conventional SIMP (Simplified Isotropic Material with Penalization) based topology optimization schemes consume a very high computational power and therefore act as a bottleneck in the manufacturing process. In this work, an accelerated topology optimization technique based on deep learning is presented. Conditional Generative Adversarial Network (cGAN) architecture is used to predict the optimal topology of a given structure subject to a set of input parameters. Next, stress contours are mapped onto the optimal structure to give accurate stress distribution over the structural domain. The predicted maximum Von-mises stress can directly be compared to the yield strength of the material for failure analysis. This technique is proven to arrive at the optimal design of a structure within a negligible amount of time. Also, this method is capable of predicting the stresses in the optimal design and hence plays a decisive role in the integrity of the optimal structure.