AI-Assisted Template-Seeded Pixelated Design for Multi-Metal-Layer High-Coupling EM Structures: A Ku-Band 6G FR3 PA in 22nm FDX+

Designing and optimizing passive matching networks is crucial for RFICs. Recent advances have utilized deep learning to implement random pixelated structures. However, these approaches remain constrained and cannot utilize multi-metal layers in modern silicon processes. This article introduces a template-seeded AI-assisted design approach for data generation and training AI models for multi-metal-layer EM structures in silicon technologies. A CNN-based model is trained to predict the S-parameters across 10-60 GHz with 0.1 GHz step, enhancing the training accuracy. A proof-of-concept design is demonstrated as dual-metal-layer tightly coupled yet pixelated transformer designs and their uses as the input, inter-stage, output matching networks of a 12-16GHz 6G FR3 PA in Globalfoundries 22nm FDX+ technology. This PA measures 19.84-21.63 dBm OP1dB, 25.75-33.39% PAEOP1dB, 21.17-22.20 dBm Psat, and of 27.74-34.11% PAEsat. When tested with a 2400 MHz 64-QAM signal at 13 GHz, this PA achieves 14.83 dBm Pavg and 16.79% PAEavg, EVMrms of -25 dB.