Variational Autoencoder-Based Metamodeling for Multi-Objective Topology Optimization of Electrical Machines

Conventional magneto-static finite element (FE) analysis of electrical machine design is time-consuming and computationally expensive. Since each topology has a distinct set parameters, optimization commonly performed independently. This article presents novel method for predicting key performance indicators (KPIs) differently parameterized topologies at the same time by mapping high-dimensional integrated parameters in lower-dimensional latent space using variational autoencoder (VAE). After training, via space, decoder multi-layer neural network will function as meta-models sampling new designs associated KPIs, respectively. enables parameter-based concurrent multi-topology optimization.