A Multi-Objective Topology Optimization Methodology Using Deep Learning and Its Application to Electromagnetic Devices

In this article, a multi-objective topology optimization (MOTO) methodology using Non-dominated Sorting Genetic Algorithm II (NSGAII) and convolutional neural network (CNN) is proposed. The original NSGAII is improved to achieve better global search ability and uniform distribution of Pareto solutions. And CNN is applied as a surrogate model for finite element analysis (FEA). The framework of the proposed methodology is elaborated. To validate the proposed methodology, it is applied to the TO of an electromagnetic actuator. Numerical results validate the proposed methodology and demonstrate that computational cost of TO can be reduced.