Multi-labeled image data-based generative topology optimization of primary mirror with conditional designable generative adversarial network and reinforcement learning

In this study, topology optimization based on multi-labeled image data was conducted for a multi-objective primary mirror to produce novel designs with varying design variables. The primary mirror utilized in this application possessed a delicate structure where both the wavefront error and weight were subject to optimization. However, it was observed that the wavefront error and weight did not exhibit an inverse relationship, necessitating the development of a new multi-objective optimization approach for the primary mirror. Initially, 93 data points were gathered using the finite element method and categorized based on their wavefront error and weight. Topology optimization and the generation of novel designs were accomplished through iterative utilization of both the conditional designable generative adversarial network (CDGAN) and reinforcement learning (RL). To address the limitations inherent in each model and to ensure the effective implementation of the primary mirror, the structure of the CDGAN + RL model underwent modifications and optimizations employing multi-labeled images and considerations of boundary conditions. The application of CDGAN + RL successfully yielded multiple design solutions for unseen-optimized primary mirrors contingent upon the wavefront error and weight. Three-dimensional design variables (rib thickness, face-sheet thickness, cutting depth, and double arch) were optimized and validated based on the labels of the image data and the corresponding generated designs, revealing a minimum error rate of 1.73% and a maximum error rate of 9.37%. Comparative analysis against the initial design demonstrated performance enhancements of 41.84% and 5.41% for the wavefront error and weight, respectively.