Topology optimization with advanced CNN using mapped physics-based data

This research proposes a new framework to develop an accurate machine-learning-based surrogate model to predict the optimum topological structures using an advanced encoder–decoder network, Unet, and Unet++. The trained surrogate model predicts the optimum structural layout as output by inputting the results from the initial static analysis without any iterative optimization calculations. Input and output data are generated using the commercial finite element analysis package, Abaqus/Standard, and an optimization package, Abaqus/Tosca. We applied the data augmentation technique to increase the amount of data without actual calculations. Primarily, this research focused on overcoming the weaknesses of previous studies that the trained network is only applicable to limited geometry variations and requires an organized grid rectangular mesh. Therefore, this study suggests a mapping process to convert the analysis data on any type of mesh element to a tensor form, which enables training and employing the network. Also, to increase the prediction accuracy, we trained the network with the labeled optimum material data using a binary segmented output, representing the structure and void regions in the domain. Finally, the trained networks are evaluated using the intersection over union (IoU) scores representing the classification accuracy. The best-performing network provides highly accurate results, and this model provided the IoU scores for average, maximum, and standard deviation as 90.0%, 99.8%, and 7.1%, respectively. Also, we apply it to solve local-global structural optimization problems, and the overall calculation time is reduced by 98%.