Topology optimization using Generative Models

Abstract

Manufacturing has always sought to decrease production costs and increase efficiency. Moreover, it also needs to consider the environmental impact. Topology Optimization is a way to achieve both by reducing material used during production, leading to less waste and better long-term environmental benefits. The process involves using iterative algorithms to achieve optimal structural layout while considering different constraints, conditions, and loads. The aim is to maintain the strength of the final product while removing unnecessary material loads. There are several approaches to achieving topology optimization, including density-based methods such as SIMP and BESO. Although effective, these methods can be computationally costly and time-consuming. Deep learning models, such as CNN and GAN networks, have shown promise in decreasing the time and cost required for the optimization process. Research in this area is ongoing, and new techniques are continually being developed. In this research, the potential of different Generative Adversarial Networks in Topology Optimization has been analyzed. The training process of the GAN models has been tried to improve while keeping them highly accurate. This was achieved by the proposal of a new hybrid generator architecture. To compare the results of proposed network, [1] model and results has been referred as a benchmark. The proposed model could achieve nearly 1.6 times improved training time. A slight decrease of 0.02 MAE and MSE scores have been observed.