AI-Assisted Complex Load Mismatch Prediction by Mm-Wave Reconfigurable Doherty Power Amplifier and Input/Output Power Sensors

This paper presents an AI-assisted approach for mm-Wave Doherty Power Amplifiers (PAs) to predict the PA’s unknown antenna load mismatches in large-scaled phased arrays using only input/output power sensors. In phased arrays, the antenna inter-element coupling creates a varying complex load (non-50 Ω), i.e. antenna load voltage standing wave ratio (VSWR), which often causes substantial performance degradation of the PAs. We address this challenge by proposing a machine-learning algorithm acting on an existing mm-Wave reconfigurable Doherty PA with only input/output power sensors. We use the Extreme Gradient Boosting (XGBoost) algorithm to predict the VSWR magnitudes and phases of unknown mismatched loads and thus the optimal settings for the PA’s main/auxiliary paths’ adaptive biasing voltages can be selected to restore its large-signal linearity and efficiency performance. For proof-of-concept, simulation-based experiments are performed on an existing 4-way mm-Wave PA design in a 45nm CMOS RFSOI process over VSWR magnitude from 1.5:1 to 2.5:1 and VSWR phase over 0^◦ to 360^◦. Two parallel ML models predict the VSWR magnitude and phase separately, achieving a VSWR magnitude prediction accuracy RMSE (normalized root mean squared error) as 0.003 and VSWR phase prediction RMSE as 5.22.