Machine Learning Based Digital Signal Processing For High Speed Optical Interconnects

In recent years,the explosively increasing of data traffic due to applications such as 4K/8K display,cloud computing,AI,4G/5G,AR/VR,etc.has been boosting the demand of optical interconnection inside or between data centers and high performance computers.Optical communication networks with high bandwidth,low loss and large capacity are indispensable to meet the demands of the increasing data traffic,especially for short-haul optical interconncets.Currently,in the applications of short-haul optical interconncetion,intensity modulation and direct detection(IMDD)system is the main scheme.However,the limtation of the systems such as noise,dispersion and nonlinearity makes the digital signal processing(DSP)techniques become important in high-speed transmission.At present,based on the comprehensive consideration of cost and performance,the traditional binary modulation is the main modulation for high-speed optical interconnection.For the next generation,since the bandwidth of the direct modulator and external modulator is hard to meet the demands,using advanced high-order modulation to improve the spectrum utilization efficiency is imminent.Advanced modulation formats like PAM: pulse amplitude modulation,OFDM: orthogonal frequency division multiplexing,DMT: discrete multi-tone modulation and CAP: carrier-less amplitude phase modulation have attracted a lot of interest.Among them,PAM,such as PAM-4 and PAM-8,can offer very good balance between performance and simplicity,so as to satisfy the practical application requirements.On the other hand,besides advance modulation formats,DSP methods such as maximum likelihood sequence estimator(MLSE),decision feedback equalizer(DFE)and feed-forward equalizer(FFE)have been utilized to improve the optical transmission performance.Another DSP techniques based on support vector machine(SVM)algorithm has shown excellent performance in mitigating modulation nonlinearity in IMDD system.The main research contents and innovations of this topic include:1.Machine learning based DSP technology for VCSEL multimode optical interconnection: Faced with the key topic of optical communication IMDD systems,nonlinearity problem,this paper focused on the characteristics of PAM signals,considered the performance and complexity of the algorithm,proposed a complete binary tree structure of SVM multi-classification(CBT-SVMs)demodulator.The performance of the algorithm has been verified by further experiments.In the vertical cavity surface emitting lasermulti-mode fiber(VCSEL-MMF)system,compared with traditional decision methods,CBT-SVMs could improve the receiver sensitivity by 1 dB,meanwhile,the relationship between eye linearity(EL)and algorithm performance has been quantitatively analyzed.With the deterioration of EL,the algorithm performance has been gradually improved.When EL reached 1.72,CBT-SVMs brought 2.5-dB receiver sensitivity optimization.2.Machine learning based adaptive optical receiving technology for Si-MRM system: Silicon micring modulator(Si-MRM)have the attractive features of compact footprint,high modulation speed,and low energy consumption.However,the high Q factor of the Si-MRM makes it very sensitive to resonance drift,which means it will cause serious damage to the signal.In this paper,CBT-SVMs has been used to improve the adaptive performance of the system.Experiments shown that the algorithm could effectively stabilize the system performance.This study quantitatively analyzed the relationship among resonance wavelength shift,PAM-4 level deviation(LD)and CBT-SVMs performance.Resonance wavelength shift at 0.06 nm would bring up to 36% LD,and the performance of the algorithm would also be improved.3.Machine learning based optical interconnection sequence dependence damage optimization technology: In addition to the problem of nonlinearity,we also find that when modulating PAM-4 signal with a modulator,the eye diagram will be skewed.This problem is prominent in VCSEL and will be affected by PAM-4 sequence.Therefore,it is necessary to propose a DSP method to solve this situation.Therefore,this paper innovatively proposed the recurrent neural network(RNN)based DSP method to solve this problem and experimental verified the feasibility of the algorithm.