Deep learning approach for defect identification and current distribution reconstruction in REBCO-coated conductors based on magnetic field mapping

High-temperature superconducting (HTS) rare-earth barium copper oxide (REBCO)-coated conductors (CCs) are widely used in large-scale applications owing to their high critical performance and excellent mechanical stability. However, defects can significantly degrade their performance by distorting current distributions. Conventional nondestructive evaluation techniques typically involve stepwise inversion from magnetic field measurements to current distributions, and then to defect morphology. However, these approaches suffer from intrinsic ill-posedness and limited ability to identify complex or multiple defects. To address these challenges, we propose an integrated inverse-forward framework that combines deep learning with physical modeling to directly map surface magnetic field measurements to internal defect morphology and current density distributions. The core of this method is a specially designed conditional generative adversarial network (cGAN) as the primary learner. It undergoes adversarial training on a dataset containing complex multi-defect configurations and their corresponding magnetic field distributions, enabling robust inverse prediction of defect morphology. The reconstructed defect geometry is then incorporated as a boundary constraint into the forward solution of Maxwell's equations coupled with superconducting nonlinear constitutive relations, enabling high-fidelity reconstruction of the current density distribution. Validation was performed using finite element simulations of REBCO-CCs with various defect types under current-carrying conditions. The adversarial training was performed using the generated finite element dataset. The trained model accurately identified multiple complex defect morphologies across different operating conditions and demonstrated strong generalization capability. Furthermore, the reconstructed current density accurately reproduced the original magnetic field distribution, confirming the reliability of the proposed method. This work integrates data-driven inference with physics-based modeling to establish a unified “magnetic field–defect morphology–current density” inversion paradigm, providing an efficient and reliable approach for nondestructive evaluation and performance assessment of REBCO-CCs.