Sequential Optimal Design of Coaxial Magnetic Gears for EV Reducers Using Taguchi Method and Machine Learning Surrogate Model

Coaxial magnetic gears (CMGs) have attracted considerable attention as alternatives to mechanical gears, offering contactless torque transmission and reduced mechanical losses. However, further improvements in torque density are still required for practical applications, and optimization processes based on finite element analysis often entail high computational cost due to the large number of function evaluations. To address this, a sequential optimal design method is proposed that combines the Taguchi approach with machine learning–based surrogate modeling, enabling efficient exploration of the design space with reduced computational effort. Both surface permanent magnet (SPM) and Halbach-type CMGs are investigated. Candidate pole combinations are first selected considering torque ripple and unbalanced magnetic forces, and initial models are derived using Taguchi orthogonal arrays. Surrogate models are then constructed using support vector machines, with 1,000 samples for the SPM-type and 3,000 for the Halbach-type, and validated using MAPE, RMSE, and the R²-score, all indicating high reliability. Finally, a genetic algorithm was applied to obtain the optimal designs, and the proposed method successfully derived CMG configurations with improved torque density compared with the initial models, demonstrating its applicability and effectiveness for magnetic gear design optimization.