Image Transfer Applied in Electric Machine Optimization

Abstract

Researches have been conducted on the surrogate-modeling for better trade-off between solution accuracy and solving effort in design space exploration. In this paper, a robust method combining the deep-learning technique, image-transfer, with finite-element-modeling (FEM) in the electric machine optimization to accelerate the convergence is proposed. Specifically, a conditional generative-adversarial network is built to learn from the FEM simulated data about the relationship between the geometric drawing input and magnetic field plot output. The learned model can obtain the result 24x faster than finite-element modeling while maintaining the accuracy. This approximation model is then applied as the sample filter prior to the FEM in the genetic-algorithm powered optimization framework. The test done on a V-shape magnet motor optimization shows that closely matched Pareto-frontier can be found by this approach while the computing time is reduced by >50% at beginning stage for acceleration.