Machine learning assisted modulation-format transparent and nonlinearity tolerant carrier recovery scheme for intelligent receiver

A modulation-format transparent and nonlinearity tolerant carrier recovery scheme, based on machine learning and kurtosis, is proposed and demonstrated experimentally. In our scheme, kurtosis-based blind phase search (BPS) and machine learning assisted maximum likelihood (ML) estimator are deployed together to achieve precise phase noise estimation. In the first stage, kurtosis-based BPS is used to estimate phase noise coarsely in a decision-free and modulation-format transparent manner. Then, mean-shift clustering, the well-known machine learning algorithm, is employed to find the centroid points of clusters automatically. Finally, the centroid points are taken as reference in ML estimator for precise residual phase noise estimation. Compared with BPS and universal carrier phase estimation scheme (UCPE), higher nonlinearity tolerance in our scheme is demonstrated via numerical simulation. Meanwhile, the linewidth tolerance of the proposed scheme is six times than that of UCPE for 64QAM. Our scheme is also verified experimentally for both 28GS/s QPSK and 16-QAM systems in the back-to-back (BTB) and transmission case. The results indicate that our scheme obtains comparable OSNR sensitivity compared with UCPE and BPS in the BTB case. Furthermore, the proposed scheme also achieves 0.4 dB and 0.2 dB Q-factor improvement than UCPE and BPS for 16QAM after 915 km fiber link transmission, respectively.