Fiche individuelle

This study presents a hybrid approach combining the Jiles-Atherton (JA) physical model with a feedforward neural network (FFNN) to enhance the modelling accuracy and computational efficiency of magnetic properties in electrical steels. Traditional JA models face limitations in handling dynamic and multi-physical conditions. By integrating a neural network to correct the residual error of the JA model, the proposed method achieves significant improvements in predicting both static and dynamic hysteresis loops. Hyperparameter optimization reduces the error compared to standalone JA simulations. Results demonstrate that the prediction of the hybrid model maintains physical interpretability while leveraging machine learning for rapid adaptation to complex scenarios, achieving errors below 1% in static cases and robust performance in dynamic cases.