Fast Topology Optimization for PM Motors Using Variational Autoencoder and Neural Networks with Dropout

This study proposes a novel topology optimization (TO) method for permanent magnet (PM) motors based on a variational autoencoder (VAE) and a neural network (NN). The VAE is trained to embed various shapes generated from the TO into the latent space. The NN is trained to predict the characteristics of the PM motor from its latent representation derived using the VAE. After training, TO is performed in the latent space based on the prediction using the NN. We adopt the Monte Carlo dropout to maintain prediction reliability using the NN during optimization, where prediction deviation is evaluated and used to eliminate unreliable solutions. The proposed method yields Pareto solutions within 80 s in a single-thread CPU machine, which is considerably faster than numerical analysis-based optimization, such as finite element analysis.