Bias-Corrected Eddy-Current Simulation Using a Recurrent-Neural-Net / Finite-Element Hybrid Model

This work combines recurrent neural networks (RNNs) with the finite element (FE) method into a hybrid model to correct time-dependent discrepancies in low-fidelity engineering simulations. The hybrid model is trained on sparse data from high- and low-fidelity simulations, employing techniques to prevent overfitting and balance accuracy with neural network generalization. It is successfully applied to an eddy-current simulation of a quadrupole magnet, demonstrating its accuracy in adjusting low-fidelity models. The results confirm the potential of this hybrid modeling approach for model-based predictions in dynamic multi-fidelity modeling contexts.