Font Size: 

Due to superior performance yet light weight, shell-infill structures have been preferable in many engineering disciplines. Topology optimization of shell-infill structures can directly generate an optimal material configuration with the shell-infill feature. The standard procedure requires repeatedly finite element analysis, which is difficult to realize real-time topology predictions. By employing deep learning approach, topology optimization of shell-infill structures can be extremely accelerated due to feature learning through the training process.

Compared to standard structures, topology optimization of shell-infill structures implements feature information into the material interpolation model. The complicate feature also brings challenges to standard deep-learning-based topology optimization. To guarantee the prediction accuracy, we introduce a novel and differentiable structural similarity (SSIM) loss function into conditional generative adversarial network (cGAN) to develop SS-cGAN model. A special initial condition coding strategy is developed to generate datasets of optimized shell-infill structure. SSIM loss function is the comprehensive value of the three functions of luminance, contrast, and structural contrast. This loss function can perceive the overall structure of the sample. SS-cGAN can generate shell-infill structures in real time after training with this efficient small-scale dataset. By evaluating the deep learning models on different examples under multiple metrics, such as mean absolute error, mean relative error of volume fraction, mean relative error of compliance, we demonstrate that SS-cGAN can predict shell-infill structures with lower error and more structural integrity than a baseline cGAN.

shell-infill structure; topology optimization; deep learning; conditional generative adversarial network; structural similarity loss function