Development of Deep Convolutional Neural Network for Structural Topology Optimization

The paper presents a method to develop an accurate surrogate model, a deep-learning-based convolutional neural network (CNN) to optimize various types of structures in 2D and 3D using topology optimization. In general, structural topology optimization requires plenty of computations because of a large number of required finite element analyses (FEAs) to obtain optimal structural layouts to reduce the weight. Machine learning has been applied in many previous studies to increase computational efficiency. Researchers have proposed various methods to develop a surrogate model with a neural network to predict the material density configuration using the static analysis results obtained for the initial geometry without performing many iterative FEAs. In this research, we propose the use of a new framework that can improve the data utilization efficiency for training and predicting the optimal densities for the topological optimization of structures. To evaluate the proposed method, three case studies were conducted on the following: a 2D cantilever plate with a point load, a 2D simply-supported plate with a distributed load, and a 3D stiffened panel with a distributed load. In all cases, the developed surrogate models can predict the optimum structures with equivalent structural performance levels as those derived through conventional topology optimization. Also, when the optimal structures were derived using the proposed method, the total calculation time was reduced by 98% as compared to conventional topology optimization, once the CNN has been trained.