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 (PMNs) is critical for high-performance RFICs. Traditional methods rely on empirical models and computationally intensive EM simulations. Recent deep learning methods using random pixelated structures have expanded the design space but remain primarily limited to single-metal-layer structures, thus failing to fully exploit the advantages of multi-metal-layer high-coupling EM structures in modern silicon nodes. This article introduces a template-seeded AI-assisted design methodology for efficiently generating data and training AI models for multi-metal-layer EM structures in silicon technologies. A convolutional neural network (CNN)-based model is trained to predict the S-parameters from 10–60 GHz in 0.1 GHz steps. A proof-of-design is demonstrated as dual-metal-layer tightly coupled pixelated PMNs used as the input, inter-stage, and output matching networks in a 12–16 GHz wideband 6G FR3 PA, implemented in GlobalFoundries 22nm FDX+ technology. The demonstrated PA measures OP1dB of 19.84–21.63 dBm, PAEOP1dB of 25.75–33.39%, Psat of 21.17–22.20 dBm, and PAEsat of 27.74–34.11%. With a 2400 MHz 64-QAM signal at 13 GHz, the PA delivers 14.83 dBm Pavg and 16.79% PAEavg, with an rms EVMrms of -25 dB.