Multi-Head Self-Attention GANs for Multiphysics Topology Optimization

Machine learning surrogates for topology optimization must generalize well to a large variety of boundary conditions and volume fractions to serve as a stand-alone model. However, when analyzing design performance using physics-based analysis, many of the recently published methods suffer from low reliability, with a high percentage of the generated structures performing poorly. Disconnected regions of solid material between boundary conditions lead to unstable designs with significant outliers skewing the performance on test data. In this work, we introduce multi-head self-attention generative adversarial networks (MHSAGAN) as a novel architecture for multiphysics topology optimization. Our network contains multi-head attention mechanisms in high-dimensional feature spaces to learn the global dependencies of our data, i.e. connectivity between boundary conditions. We demonstrate our model on design of coupled thermoelastic structures and evaluate its performance with respect to the physics-based objective function used to generate training data. Our proposed network achieves over a 4X reduction in mean objective function error and an 8X reduction in volume fraction error compared to a baseline approach without attention mechanisms.