Latent Space Bayesian Parameter Topology Optimization of Permanent Magnet Synchronous Motors via a Variational Autoencoder

This study introduces a novel approach that combines a variational autoencoder and Bayesian optimization to accelerate the simultaneous parameter and topology optimization of interior permanent magnet synchronous motors. Traditional parameter and topology optimization methods, typically relying on heuristic algorithms such as genetic algorithms, require a large number of finite element method evaluations due to the high dimensionality of the design space. To address this challenge, the proposed method uses a variational autoencoder to compress the design space into a low-dimensional latent representation. Bayesian optimization is then applied within this latent space, utilizing a probabilistic surrogate model to efficiently explore promising designs. When applied to an interior permanent magnet synchronous motors design problem, the proposed method achieved a 78.7% reduction in search time compared to conventional approaches, without compromising optimization performance.