A Highly Tunable RF Self-Interference Canceler with Near Real-Time Machine-Learning Augmented Adaptation for Full-Duplex Radio Applications

This paper presents a highly tunable self-interference (SI) canceler with a machine-learning augmented adaptation loop for deep cancellation, enhanced linearity, and accelerated convergence time. A prototype 2.4 GHz canceler chip, fabricated using a 40 nm CMOS process, featuring extensive tuning capabilities in terms of filter coefficients, delay spread, and linearity, works in concert with a neural network (NN) algorithm implemented on a Xilinx RFSoC evaluation kit to enhance the canceler’s adaptation process. The canceler chip, implemented as a 6-complex tap finite impulse response filter, achieves a measured maximum SI cancellation of 32 dB over a 40 MHz bandwidth and an input-referred third-order intercept point $\left(\mathrm{IIP}_{3}\right)$ of +38 dBm. The proposed NN computes the initial solution for the canceler’s settings, followed by a local optimization, achieving a minimum convergence time of 0.15 seconds in real-world wireless environments. Additionally, the SI canceler achieves a maximum group delay of 2 ns, an output noise of $-184 \mathrm{dBV} / \sqrt{H z}$, and an $S_{11}$ of -10 dB across a bandwidth spanning from 2.3 GHz to 2.8 GHz.