Deep Learning-Based Reconstruction of Characteristic Mode Currents on Perforated PEC Patches

In this paper, we propose an efficient deep learning framework for predicting the characteristic mode currents of 2D perforated PEC patches. A U-Net-based neural network is developed and trained using high-fidelity datasets generated by the Method of Moments (MoM). Notably, the proposed model is designed to directly reconstruct both the normalized magnitude and direction angle of the surface current distributions for the first four characteristic modes from binary geometry masks in a single forward pass. To verify the performance of the proposed method, we apply it to numerical results obtained from a dataset of generalized quadrilateral geometries with random perforations. Numerical experiments validate that the model provides high prediction accuracy and strong generalization to previously unseen complex topologies. In other words, this method offers a more accurate and efficient neural network-based solution for rapid CM analysis in antenna design.