Neural Network Meta-Model Method for Performance Prediction of Axial Flux Machines

This paper proposes a computationally efficient surrogate modeling methodology for axial flux traction motors (AFMs) utilizing a neural network-based meta-model constructed from unit linear motor (ULM) simulations. By decomposing the complex three-dimensional (3D) finite element analysis (FEA) into multiple radial segments represented by ULMs, the developed meta-model significantly reduces computational overhead compared to conventional multi-slice 2D and full 3D FEA methods. Trained on a dataset comprising 20,000 ULM simulation cases, the neural network achieves an exceptional predictive accuracy, demonstrated by an R-squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) value exceeding 0.99 and a mean normalized error below 1% for key performance metrics, including torque, line voltage, electromagnetic losses, and material cost. Furthermore, the proposed method offers substantial flexibility in evaluating diverse and complex magnet geometries without extensive remeshing efforts. Consequently, this surrogate modeling approach substantially accelerates the preliminary optimization and design phases for AFMs, balancing computational efficiency, modeling flexibility, and robust prediction accuracy.