AI-driven Electromagnetic Prediction with Pixelated Matching Networks for Broadband and Miniaturized Rectifier Design

Among wireless power transfer and harvesting, broadband and miniaturized rectifiers are key nonlinear modules for efficient RF-to-DC conversion, but their design relies heavily on experience and slow EM-schematic optimization, limiting timely meeting of bandwidth and dimension demands. This research proposes an AI-driven EM prediction approach with pixelated matching networks, enabling theoretical mapping of such networks to extend impedance regions, reducing reliance on iterations and experience. A CNN-Transformer hybrid architecture with an independent via encoding layer is used as a surrogate model, replacing full-wave EM calculations to rapidly predict S-parameters, achieving an average mean absolute error (MAE) < 0.03 and correlation coefficient of R = 98.7%. Using an improved particle swarm optimization (PSO) algorithm and the trained model, rectifiers are designed, fabricated, and tested. A 10 mm × 10 mm pixelated matching network is applied with HSMS-286C diodes on Rogers RO4350B substrate. Results show a good agreement: a 1.5-3.5 GHz broadband rectifier achieves > 67% efficiency @13 dBm; dual-frequency and single-frequency rectifiers reach maximum efficiency > 76%.