Physics-informed neural network-based magnetostriction model for grain-oriented electrical steels

The magnetostriction of grain-oriented electrical steels is important for evaluating the core vibrations of operational power transformers. However, magnetostriction is regarded as the square of magnetization in most studies, where influencing factors such as stress and magnetic saturation effects are disregarded. In this study, a two-step magnetostriction-prediction technique based on a physics-informed neural network (PINN) is proposed. First, the physical model of magnetostriction is deduced from the classical Jiles-Atherton model by introducing an elastic potential energy term. Second, a PINN-based model is used to obtain the resulting magnetostriction by adding the residual of the physical equation. The magnetic parameters and magnetostriction of the test specimens are measured experimentally under different excitation voltages and tensile stresses. Results show that magnetostriction is a highly stress-dependent phenomenon. As the tensile stress increases, the elastic potential energy gradually becomes the dominant factor affecting magnetostriction, thereby resulting in a significant amplitude shift in the frequency domain. The performance prediction of the PINN is significantly better than that of artificial neural networks, particularly for unknown samples with slight voltage variations.