Convolutional Encoder Decoder Network for Topology Optimization

This paper presents a deep learning approach for topology optimization. Since topology optimization is widely used in mechanical engineering and material science,a deep neural network model is proposed to overcome excessive time and computational power problem that optimization requires.A deep neural network is trained to learn topology optimization process and give desired final structure with minimum error. Since it is an image segmentation problem, encoder-decoder network architecture is considered appropriate for the task. Our encoder-decoder network consists an architecture that takes second iteration image matrix and gradient matrix as input and generates desired image matrix as output. The evaluations have demonstrated that even the optimization is stopped in second iteration, our model can accurately predict the test datasetwith negligible computational time. With improving the proposed algorithm, this method can be adapted into topology optimization-based problems. Keywords - Deep Learning, Encoder-decoder network, Image segmentation, Topology optimization