A neural network-based DPD coefficient determination for PA linearization in 5G and beyond-5G mmWave systems

Abstract

This work presents a neural network DPD for mmWave RF-PAs. Differently from existing neural network-based DPDs, the neural network in the proposed DPD does not reside in the forward data path. Instead, it estimates the polynomial coefficients from the complex Fourier amplitudes of harmonics during a calibration sweep. It can compensate for PA nonlinearity under various operating conditions with lower hardware complexity compared to conventional DPDs. The proposed design is validated on a 28GHz CMOS phased-array transceiver. In 256-QAM 5G-OFDMA-mode, the proposed neural network DPD achieved an improvement in EVM from -28.7dB to -32.0dB, while maintaining an ACLR of -33.4dBc.