Hybrid Physics-Informed Learning for Multi-Objective Surrogate Models for Nonlinear Magnetic Bearing Actuator

Abstract

Learning-based surrogate modeling techniques have been extensively investigated and provide practical alternative in electromechanical systems. Traditional implementations rely on lookup tables, which, while accurate, depend on an interpolated discrete simulation dataset generated from the computationally expensive high-fidelity Finite Element Method (FEM) simulations. Moreover, these approaches scale poorly with increasing dimensionality. This paper proposes a multi-task hybrid physics-guided surrogate model for an axial active magnetic bearing system using a data-driven Neural Network (NN) and a physics-informed neural network (PINN). The main novelty lies in integrating algebraic electromagnetic force relations and dynamic voltage-current constraints within a unified multi-output learning framework for simultaneous estimation of force and inductance from coil current and rotor displacement inputs. Both models are trained using FEM-generated data, while PINN incorporates governing physics constraints within the loss function. Additional synthetic time-domain data are used in inductance physics loss. Offline validation using experimental axial AMB data demonstrates that the proposed surrogate models accurately reproduce the FEM behavior, while PINN provides a better alternative to a traditional lookup table due to physics constraints in its loss function.