Unit Module-Based Convergence Acceleration for Topology Optimization Using the Spatiotemporal Deep Neural Network

Abstract

This study proposes a unit module-based acceleration method for 2-D topology optimization. For the purpose, the first-stage topology optimization is performed until the predefined iteration. After a whole design domain is divided into a set of unit modules, information on the spatiotemporal characteristics of intermediate designs and a filtering radius is used to separately predict a near-optimal design of each unit module through a trained long short-term memory (convLSTM) network. Then, in the second-stage topology optimization, a combined near-optimal design of a whole design domain is used as an initial design to determine the optimized design in a more efficient way. To train a convLSTM network, a history of intermediate designs is obtained under a randomly generated boundary condition of a unit module. The filtering radius is also used as the training data to reflect the geometric features affected by a filtering process. For four examples with different design domains and boundary conditions, the proposed method successfully provides the accelerated convergence up to 6.09 with a negligible loss of accuracy less than 1.12% error. These numerical results also demonstrate that the proposed unit module-based approach achieves a scalable convergence acceleration at a design domain of an arbitrary size (or resolution).

Keywords

• Convergence acceleration
• deep learning
• finite element method
• structural topology optimization