Design and Analysis of Electric Machines Assisted by Scientific Machine Learning

Electric machines have been widely used in industrial motor drives, aerospace, and automotive applications. Its design and optimization rely heavily on finite element analysis (FEA) to solve a subset of Maxwell’s equations for performance evaluation and validation. Simulation tasks such as the topology exploration, performance characterization, noise, vibration and harshness (NVH) prediction, electromagnetic compliance (EMC), etc. can be time-consuming even with high performance computing (HPC). The recent success of scientific machine learning (ML) in processing 3-dimensional (3D) non-Euclidean geometries and combining pure data-driven ML techniques with physics insights makes it possible to accelerating those computationally expensive tasks by tens to hundreds of times. This paper provides the first overview of the key enabling techniques – geometric deep learning (DL) and physic-informed ML in the context of electric machine design and analysis. A general framework is then presented to apply these techniques to the design and optimization of electric machines efficiently. The effectiveness of the presented framework is verified by a traction motor simulation case. The open challenges and next steps are summarized, calling for more concerted effort into this fast-growing research area.