Free-Form Optimization of Electric Machine Rotors Using Deep Learning

Abstract

Design of electric motors is often constrained by traditional finite element solvers that rely on fixed parameterizations. This thesis investigates whether machine learning can be used to move beyond parameterized design spaces while reducing reliance on simulations, with the objective of removing rotor material while retaining performance. An image-based design space is introduced to enable free-form shapes and placement of the rotor air ducts. Neural networks are trained on simulation data to predict key performance metrics and to learn a compact representation of feasible designs. This compact space is then explored using gradient-based optimization to identify improved rotor configurations. The models achieve good predictive accuracy overall, although mechanical stress is more difficult to estimate due to its sensitivity to fine geometric details. Several optimized designs outperform the best designs in the training data when comparing rotors of equal mass. In addition, the results indicate that approximately 20% of the rotor material could be removed with only a minor reduction in performance. Overall, the study demonstrates that machine learning can support the exploration of free-form rotor designs beyond conventional parameterizations. Moreover, the final optimized designs depend strongly on the choice of weights in the objective function, highlighting the importance of carefully formulating the optimization problem.