A Hybrid Jiles–Atherton–Sablik Physics-Informed Convolutional Recurrent Residual Network for Magneto-Mechanical Hysteresis Prediction in Non-Oriented Electrical Steel Fe3%Si

Accurate prediction of magnetic hysteresis under coupled mechanical and electromagnetic excitations remains a challenging problem in electrical steel modelling. Conventional phenomenological models, although physically interpretable, often suffer from limited adaptability when several effects such as stress dependence, frequency variation and non-linear domain mechanisms occur simultaneously. This work proposes a hybrid modelling framework combining a Jiles–Atherton–Sablik physical model with a deep residual correction based on a convolutional recurrent residual neural network (CRRN). The physical model provides the main magnetization mechanism, while the neural network learns the remaining discrepancy between measured and physics-based predictions. The experimental database comprises non-oriented Fe–3%Si electrical steel sheets tested under three orientations, nine stress levels (tensile/compression), three excitation frequencies, and multiple peak induction levels, totaling over 2000 BH loop files. The implemented framework optimizes the four Jiles–Atherton (JA) parameters (<!-- MathType@Translator@5@5@MathML2 (no namespace).tdl@MathML 2.0 (no namespace)@ -->