Deep-learning-enabled high-performance full-field direct detection with dispersion diversity

Data center interconnects require cost-effective photonic integrated optical transceivers to meet the ever-increasing capacity demands. Compared with a coherent transmission system, a complex-valued double-sideband (CV-DSB) direct detection (DD) system can minimize the cost of the photonic circuit, since it replaces two stable narrow-linewidth lasers with only a low-cost un-cooled laser in the transmitter while maintaining a similar spectral efficiency. In the carrier-assisted DD system, the carrier power accounts for a large proportion of the total optical signal power. Reducing the carrier to signal power ratio (CSPR) can improve the information-bearing signal power and thus the achievable system performance. To date, the minimum required CSPR is ∼7 dB for all the reported CV-DSB DD systems having electrical bandwidths of approximately half of baud rates. In this paper, we propose a deep-learning-enabled DD (DLEDD) scheme to recover the full optical field of the transmitted signal at a low CSPR of 2 dB in experiment. Our proposal is based on a dispersion-diversity receiver with an electrical bandwidth of ∼61.0% baud rate and a high tolerance to laser wavelength drift. A deep convolutional neural network enables accurate signal recovery in the presence of a strong signal-signal beat interference. Compared with the conventional method, the proposed DLEDD scheme can reduce the optimum CSPR by ∼8 dB, leading to a significant signal-to-noise ratio improvement of ∼5.8 dB according to simulation results. We experimentally demonstrate the optical field reconstruction for a 28-GBaud 16-ary quadrature amplitude modulation signal after 80-km single-mode fiber transmission based on the proposed DLEDD scheme with a 2-dB optimum CSPR. The results show that the proposed DLEDD scheme could offer a high-performance solution for cost-sensitive applications such as data center interconnects, metro networks, and mobile fronthaul systems.