Super-Resolution Based Topology Optimization for Rapid Generation of Low Mass Structural Designs

Mass is a primary design driver in NASA spacecraft applications, directly limiting the number of scientific instruments that can be flown within a mission architecture, especially in the case of interplanetary missions. Topology Optimization (TO) is an emerging physics-based computational design tool with the ability to address this need, generating designs often 15-20% lighter than conventional design approaches. In general, the computational cost and run time of TO are directly correlated with the desired features resolution of the generated design. Unfortunately, the thin geometries required for many aerospace structures require high resolution finite element meshes to yield quality TO designs, resulting in high computational power needs coupled with long solution times. The ability to solve a "quick" coarse-scale TO problem and, using Machine Learning (ML), map the design to a corresponding fine-scale design could reduce a design cycle time from weeks to days.In this paper, we developed a novel approach of mapping a 2D coarse-scale TO design (which runs fast with minimal compute resources) to a 2D fine-scale design (which would normally require significantly more time and compute resources using TO). Conventional TO methodologies (in particular, the Solid Isotopic Material Penalization – SIMP – methodology) generate designs in a format similar to grayscale images, providing a direct link with computer vision-based ML. This methodology is therefore similar to an image sharpening task that can build upon classical approaches such as the Single Image Super Resolution (SISR) task. We explore and propose an attention-based SRResU-Net that leverages rapid coarse-scale TO designs to seed robust high resolution TO designs. To ensure feature retention during the coarse-scale design generation, the optimization algorithm is de-tuned to generate "fuzzy" images. The results demonstrate that the proposed model enables 10x faster design generation than high resolution TO alone while maintaining comparable accuracy and performance.