Multi-Objective Motor Design Optimization with Physics-Assisted Neural Network Model

Electric machine design optimization tasks typically require a large number of time-consuming simulations using finite-element analysis (FEA) to iteratively evaluate the design candidates. Various surrogate modeling techniques have been investigated in order to speed up the design optimization process. In recent years, machine learning based surrogate models are explored, due to their advantages including extraordinary capability in learning highly nonlinear functions. However, typical neural network based machine learning models require a large amount of training data and long training time. In this paper, we propose a multi-objective optimization (MOO) scheme for electric machine design, using a physics-assisted neural network (PANN) as surrogate model. In the PANN method, a semi-analytical subdomain physics model is used to estimate the performance of the electric machine, and this calculated result is used as the input of a neural network, in addition to other design parameters. We show that PANN can achieve the same accuracy with significantly less training data, as compared with neural networks relying solely on data. The hybrid model also shows improved accuracy with the subdomain based physics model alone. We apply the PANN surrogate model to speed up the electric machine MOO by replacing the iterative FEA based optimization process. The Pareto front solutions obtained by the proposed method are further validated with FEA with good accuracy.