Scientific Machine Learning Methods for Parametric Modeling of Electric Machines

High-dimensional parametric studies to analyze and design electric machines become intractable as the number of input parameters grows. Current computational techniques are not viable due to the computational limitations caused by this curse of dimensionality. The objective of this work is to set forth computationally efficient models for conducting high-dimensional parametric studies in the context of electric machines. Our approaches use scientific machine learning methods (i.e., Gaussian processes, deep neural networks) and the first-principles of electromagnetic physics to build a response surface representing magnetic-field-related quantities of interest as a function of space and parameters. The input space includes geometric features, material properties, and operating point conditions. The proposed models are well-suited for parametric studies such as uncertainty propagation, sensitivity analysis, and optimization design. We solve examples of these forward problems to demonstrate the application of our methods. The accuracy and computational cost of the proposed methods are assessed by comparing them to finite element predictions.