Topology Optimization using Deep Learning

—Comparison of Simultaneous and Additional Learning—

Abstract

Deep learning is applied to IPM motors with different magnet shapes to realize fast topology optimization. In this method, the cross-sectional image of IPM motors with I-shaped and V-shaped magnets are input to a convolutional neural network to guess their average torque. It is shown that simultaneous learning, in which CNN is trained for both datasets, is superior over the additional learning where CNN is sequentially trained for the two datasets. Moreover, it is shown that the number of required finite element analysis can be reduced to about five percent using the trained CNN in the topology optimization.