Adaptive nonsymmetrical demodulation based on machine learning to mitigate time-varying impairments

Abstract

We proposed and experimentally demonstrated a machine learning-based nonsymmetrical demodulation technique for a DSP-enabled receiver, with the aim of enabling time-varying nonlinear mitigation. Experimental results showed that nonsymmetrical demodulation can reduce the SER by up to 0.7 decades, when assuming time frames consisting of 10 k symbols and fiber transmission of 250 km. The proposed technique is transparent to the specific source of nonlinearity, which makes it simple yet robust. This machine learning method may also allow simplification of the standard demodulation blocks in particular the equalizer. Employing short time windows for demodulation further enables inline optical monitoring, which is a valuable diagnostic tool for future terabit optical communication systems.