Demonstration of a Low-Phase-Noise MIMO 2-D Convolutional Neural Network Nonlinear Equalizer in an MIMO DMT Long-Haul D-Band RoF System

The D band presents a possible solution for accommodating ultrahigh-capacity data transmission. Multiple-input–multiple-output (MIMO) technology can further double system transmission capacity. Our study explores the integration of photonics-aided millimeter wave with MIMO technologies. Based on phonics-aided technology, we realized the transmission of 10-Gbaud D-band polarization-division multiplexing (PDM) discrete multitone (DMT) 16QAM signals at 125 GHz over a 1.2-km wireless link in an Intensity-modulation (IM)/direct-detection (DD) radio-over-fiber (RoF) system. However, apart from severe nonlinear impairment and PDM crosstalk, the D-band DMT RoF system is subject to phase noise (PN). Traditional Volterra series and deep neural network (DNN) equalizers can only compensate for nonlinear impairments but are poor at correcting phase distortion. To realize the mitigation of $X/Y$ polarization crosstalk, nonlinear impairments, PN, and I/Q imbalance, we propose a novel MIMO 2-D convolutional neural network (2D-CNN) by addressing complex-valued two PDM signals and their phase information simultaneously. This model uniquely combines the ability to mitigate these four critical impairments within a single processing framework, enhancing the robustness of MIMO IM/DD RoF systems. Our MIMO 2D-CNN significantly improves the BER performance compared to traditional digital-signal-processing (DSP) models, providing an almost tenfold enhancement. The net bit rate can achieve 31.49 bit/s under the 7% hard decision forward error correction (HD-FEC) threshold of $3.8\times 10^{-3}$ . Meanwhile, due to its unique ability to extract information from 2-D space–time images, the advanced MIMO 2D-CNN can effectively compensate for the PN of the signals with minimal phase rotation. It surpasses the limitations of Volterra series and MIMO 2D-DNN approaches, which struggle to handle such PN. Moreover, regarding receiver sensitivity, our MIMO 2D-CNN outperforms the previously proposed MIMO 2D-DNN by over 1 dB at the BER of $3.8\times 10^{-3}$ . In the future, we expect that the low-PN MIMO 2D-CNN, with its accurate BER decision ability, will be a valuable tool for further investigations into MIMO IM/DD RoF networks.