Fast Unsupervised Generative Design for Structural Topology Optimization

Ethan Herron, Anushrut Jignasu, Jaydeep Rade, Xian Yeow Lee, Aditya Balu, Adarsh Krishnamurthy, Soumik Sarkar

Keywords: Deep learning, topology optimization, generative modeling, generative design

TL;DR: A novel Generative Adversarial Network framework for structural topology optimization.

Abstract: Exploring the intersection of generative design and structural topology optimization has been a popular research area recently. Existing structural optimization methods have been shown to generate high-performance and aesthetically pleasing structures but at a tremendous computational cost. The rapidly advancing field of deep learning, particularly generative modeling, has substantial potential to tackle the structural generative design problem. Previous works have utilized deep generative models for highly specific cases, ranging from a small set of loading conditions to heavily relying on supervised loss functions for training. We propose a new method targeted at generating near-optimal structures over a wide variety of initial conditions in a completely unsupervised manner. We accomplish this by implementing a novel Generative Adversarial Network (GAN) framework to generate densities that match our given target distribution and encode extremely efficient latent representations of the initial physical conditions of the sample. The target distribution used in this work comes from data generated via the solid isotropic material condition with penalization (SIMP) topology optimization algorithm. Our results show that the proposed framework can generate similar structures to those found using the SIMP optimization algorithm, which consequently demonstrates the potential variability in solution spaces for arbitrary problems in generative design.