Machine Learning Assisted High-order Modulation For Short Reach Optical Interconnects

In recent years,the rapid development of information technology and industries such as cloud computing/storage,4G/5G wireless communications,artificial intelligence,etc.,has led to a continuous explosive growth of data traffic.bringing demand for large-capacity,highbandwidth,low-loss optical communications.In particular,short-range optical interconnects for data center(DC),high-performance computer(HPC),and wireless fronthaul applications.For short-range scenarios,cost,power consumption,and density are the key requirements.Direct detection of optical interconnects is the main solution,including traditional intensity modulation and direct detection(IM-DD)and recently emerging stokes vector direct detection(SV-DD).Based on these solutions,further adoption of advanced technologies including high-order modulation coding,error correction coding,and digital signal processing(DSP)can effectively improve the transmission rate.The high-order modulation of IM-DD system mainly includes N-order pulse amplitude modulation(PAM-N),carrier-free amplitude phase modulation(CAP),discrete multi-tone(DMT),and so on.Among them,the DMT with relatively low complexity and high spectral efficiency has attracted widespread attention in the IM-DD system.DMT loads QAM modulation with different number of bits on each carrier according to the signal-to-noise ratio(SNR)distribution of the channel.Due to the good adaptability of DMT,it is regarded by the IEEE 802.3bs working group as the most promising next-generation 400 Gbps Candidate format for standard transmission.In addition,Stokes Vector Direct Detection(SV-DD)has also received a lot of attention.SV-DD has high spectral efficiency,no localoscillator,fewer fast Fourier transform(FFT)operations,and no laser frequency offset and phase Advantages such as noise tracking are a low-cost direct detection technology and an important development direction for short-range optical interconnection.In this thesis,the research of machine learning assisted short-range optical interconnection high-order modulation technology is studied.It mainly aims at the two high-order modulation methods of DMT and SV-DD,and uses the machine learning-assisted DSP method to improve the highorder modulation performance.The main research contents include:(1)Propose a machine learning multi-classification decision method based on support vector machine(SVM)to solve the non-linear distortion problem of DMT modulation.Experiments have achieved a single lane 112-Gbps milestone high rate.Experimental research on two main optical interconnection systems: Mach-Zehnder modulator(MZM)-single-mode fiber(SMF),vertical cavity surface laser(VCSEL)-multimode fiber(MMF).Five different feature one verse rest(Ov R),symbol encoding(SE),binary encoding(BE),constellation row and column(RC),and in-phase and quadrature components(IQC)are proposed and compared for QAM Multi-classification of features.The experiments verify that the IQC classification method has the lowest complexity.(2)Propose a polarization estimation method based on end-to-end neural network(NN-end2end)to achieve SV-DD demodulation with lower bit error rate.Traditional SV-DD demodulation methods include indirect demodulation based on polarization rotation matrix estimation and direct demodulation based on blind equalization.In the neural network SV-DD demodulation proposed in this paper,the input layer is the received SV,and the output layer is the bit encoding information corresponding to the constellation point.The output result can be directly used to calculate the bit error rate.In the 2-D detection simulation of single-polarization modulation 4-QAM,the SNR of NN-end2 end is 2d B lower than the SNR based on the blind estimation(B-E)scheme at 7% FEC;In the 3-Ddetection simulation of dual-polarization modulation 4-QAM + PAM-2,at 7% FEC,the SNR of NN-end2 end is 2d B lower than the SNR based on the blind adaptive equalization(BA-E)scheme.It can be seen that NN-end2 end is significantly superior to traditional demodulation methods in recovering polarization rotation and mitigating linear and nonlinear distortion.