A novel physics-informed neural network for modeling electromagnetism of a permanent magnet synchronous motor

This study presents a novel physics-informed neural network (PINN) architecture designed to address the challenges of replicating an electric motor. The proposed architecture has three key features. First, it uses three partial differential equations with rotational coordinate transformation to supervise the neural network during training with limited data, which improves the accuracy of the solution. One of the differential equations is expressed in variational form to effectively compute the numerical integration. Second, separate networks are proposed for the rotor and stator domains due to their distinct characteristics during operation, namely, that the rotor rotates while the stator remains fixed. An interface loss is included in the entire loss function to compensate for the significant discontinuity and incompatibility between the separate networks when estimating the results of both domains. Third, a learning rate annealing method is introduced to update the adaptive weights of each loss term, thus improving the accuracy and robustness during the training of the neural network. The perfor-mance of the proposed PINN was validated using electromagnetic response datasets obtained from both mea-surements and finite element analyses. Systematic analysis demonstrated that the three features significantly improved the accuracy and robustness of the neural network when estimating the electromagnetic responses of an electric motor. Furthermore, the inference time of the PINN is ten times faster than that of a finite element analysis with a similar level of accuracy, making it suitable for control and design purposes in various real-world applications. Consequently, the versatility of the proposed PINN can accelerate the development of digital twins for intelligent systems by deploying an electric motor, and it could also be used for prognostics and health management because it can estimate electromagnetic responses under both normal and failure conditions.

Automated summary

This study proposes a novel physics-informed neural network (PINN) architecture to replicate an electric motor. The architecture uses three partial differential equations to supervise the neural network during training, separate networks for the rotor and stator domains, and a learning rate annealing method to update the weights. The PINN demonstrates improved accuracy and robustness in estimating electromagnetic responses, with significantly faster inference time compared to finite element analysis.