Multi-Objective Rotor Shape Optimization of SynRM Using Data Augmentation via Autoencoder

This study proposes a data augmentation method based on latent-space sampling using an Autoencoder (AE) for multi-objective topology optimization of synchronous reluctance motors (SynRMs). The original design data are encoded into a latent space, where their statistical distribution is modeled by a Gaussian Mixture Model (GMM). New latent vectors are then probabilistically sampled and decoded to generate geometrically diverse shapes consistent with the characteristics of the original data. The generated shapes are labeled with estimated torque and torque ripple using a deep learning model, enabling efficient dataset expansion with greatly reduced computational cost compared to finite element analysis (FEA)-based optimization. Numerical experiments demonstrate that the proposed method achieves a better balance between data diversity and distribution consistency than random sampling and conventional FEA-based optimization, resulting in an improved Pareto front with higher torque and reduced torque ripple.