Deep Learning-Assisted RFIC Design With Dual-Metal-Layer Passive Matching Networks: A 15-22 GHz CMOS PA for 6G in 22nm FDX+

Radio Frequency Integrated Circuit (RFIC) design requires human experts’ experience and intuition. Tuning and optimizing passive matching networks (PMNs) using electromagnetic (EM) simulators are labor-intensive and time-consuming. Deep learning is a promising tool to expedite this process. However, data generation is usually computationally expensive, and difficult to fully explore the design space. In this article, we propose a template-guided approach for the data generation of dual-metal-layer PMNs, which is more practical than single-layer structures for CMOS RF/mm-Wave circuit implementations. A convolutional neural network (CNN)-based EM emulator is trained to expedite the evaluation process for rapid RFIC design. To validate the proposed design approach, a 15-22 GHz power amplifier (PA) is presented for sixth-generation (6G) frequency range 3 (FR3) applications, using the pixelated PMN from the proposed approach in the GlobalFoundries 22nm FDX+ process. This PA measures an OP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1dB</inf> of 16.7-20.1 dBm, a PAE<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">OP1dB</inf> of 24.4-30.6%, a P<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">sat</inf> of 14.3-15.1 dBm, and a PAE<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">sat</inf> of 26.3-35.8%. When tested with 5G FR2 OFDM 100-MHz 64-QAM signals, this PA achieves 6.61-7.83 dBm average output power and 9.89-12.40% average PAE, with an rms error vector magnitude (EVM<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rms</inf>) of -25 dB from 14.5-23.5 GHz.