Deep Learning Based OFDM Physical-Layer Receiver with Multidilated Convolutions

In this paper, we propose a novel machine learning (ML) based physical-layer receiver architecture for demodulating orthogonal frequency-division multiplexing (OFDM) signals of high modulation orders up to 4096-QAM that are subject to severe nonlinear distortion due to the transmitter power amplifier (PA). Specifically, we develop a convolutional neural network receiver system with parallelized dilated convolutions, that allow for a shallower neural network, while efficiently and reliably demodulating the bits despite the substantial PA distortion. Representative set of training procedures and numerical results are provided, in the context of 6G mobile communication network and corresponding base-station demodulation performance for uplink operating at 7 GHz. The obtained results show that the proposed receiver systems are able to clearly outperform classical LMMSE receivers even at high error vector magnitude (EVM) levels, considering modulation orders up to 4096-QAM. Additionally, the results show that the proposed deep learning receiver outperforms the prior-art machine learning receiver methods.