Abstract
Deep learning-based topology optimization predictors have been shown to be effective in generating optimal designs. However, these predictors are prone to topological errors, particularly for high-resolution domains. Although various methods have been developed to enhance the accuracy of predicted structures, such as using large training datasets, complex networks, and physics-based loss functions, they do not include topological metrics in the deep learning models. Similar issues arise in other applications, such as blood vessels, neurons, or road segmentation from images, and several modifications to typical loss functions have been proposed to improve the topological validity of the predictions. In this study, we evaluate and compare four distinct topological loss functions to explore their influence on the performance of deep learning-based topology optimization predictors. Our findings offer insights into the advantages and limitations of these modified loss functions and provide a basis for future research and development aimed at improving the accuracy and efficiency of deep learning predictors in topology optimization.
