Variational Autoencoder-Based Multiobjective Topology Optimization of Electrical Machines Using Vector Graphics

Abstract

The increasing automation in the design process of electrical machines for vehicles generates huge amounts of data, leading to a growing interest in using machine learning for faster predictions and optimization. This paper presents an innovative approach to simultaneously model and optimize different geometries of electrical machines. Unlike previous methods that rely on specific design variables or pixel-based representations, this approach is based on a flexible vector graphic representation that can easily be applied to any electrical machine configuration. This also sidesteps the problem of scalar-based design representation and, unlike pixel-based representation, can be directly imported into finite element simulations for validation. To speed up the optimization process, we transform the corresponding problem into a lower-dimensional latent space learned by a variational autoencoder. This is trained on a total of 6839 different 2D geometries of electrical machines, which were exported from the finite element simulation into the standardized scalable vector graphics format. The optimization results show that a sharp boundary is formed in the combined Pareto front of different rotor topologies. The validation results demonstrate that this novel method delivers results comparable to those of current techniques.

Keywords

• Deep neural network
• multiobjective optimization
• permanent magnet synchronous machine
• scalable vector graphics
• variational autoencoder