Study on Efficiency Map Prediction of Synchronous Motors Using Tandem Variational Autoencoder-Based Surrogate Model

Abstract

To reduce the computational cost of finite-element analysis for rapid optimal motor design, surrogate modeling techniques have been widely studied. While conventional surrogate models mainly predict scalar performance metrics, propulsion motors must be evaluated over a wide operating range, requiring the prediction of high-dimensional outputs, such as efficiency maps. Owing to the large dimensional gap between design variables and efficiency maps, conventional neural networks often suffer from limited accuracy and unstable convergence. This article proposes a Tandem variational autoencoder (TD-VAE) that directly predicts high-dimensional efficiency maps from motor design variables. The model employs a tandem architecture in which a variational autoencoder (AE) compresses and reconstructs efficiency maps, while an auxiliary AE reconstructs the design variables to enforce latent consistency. This structure enables stable training and allows previously collected simulation data to be used without modification. The TD-VAE is further extended to predict physically interpretable loss components, including copper and iron losses. The model performance is evaluated using mean absolute percentage error (MAPE) and the coefficient of determination (R²), and predictive uncertainty is analyzed through pixel-wise confidence and prediction intervals. When applied to an interior permanent magnet synchronous motor, the proposed TD-VAE achieves a maximum MAPE of 0.18% compared with finite element analysis, demonstrating accurate prediction, stable convergence, and significantly reduced design iteration time.