56-Gbit/s PAM-4 Optical Signal Transmission Over 100-km SMF Enabled by TCNN Regression Model

 Authors: Jing Zhang; Lianshan Yan; Lin Jiang; Anlin Yi; Yan Pan; Wei Pan; Bin Luo
 Presented by: Muhammad Asif
 Date: 11/10/2023
 Abstract
Summary:
This paper presents a novel approach to enhance intensity modulation direct-detection (IM/DD) systems
using a temporal convolutional neural network (TCNN) regression model. The focus is on overcoming
challenges such as chromatic dispersion and nonlinear impairments in high-speed optical networks.
 Introduction

Background: high-speed optical networks are increasingly in demand, but face challenges like chromatic
dispersion and power fading.
Problem statement: traditional techniques struggle to address these challenges effectively.
Purpose of the study: introducing a low-complexity TCNN-based nonlinear equalizer (NLE) to improve the
performance of 56-gbit/s pam-4 optical signal transmission over a 100-km single-mode fiber (SMF).
 Operation Principle

TCNN structure: the TCNN model consists of multiple layers, including dilated convolutions, which allow it to
capture long-range dependencies in the signal.

Functioning: the TCNN processes the optical signal by effectively compensating for both linear and nonlinear
signal distortions, enhancing signal quality.
 Experimental Setup
Configuration: the experimental setup includes a PAM-4 transmitter, an optical fiber of 100-km, and a direct-
detection receiver.
Equipment: the experimental setup for the 56-gbit/s pam-4 transmission system includes the following specific
models and types of equipment.
Transmitter side:
• Laser: A laser operating at approximately 1549.32 nm with an approximately 100-khz linewidth.
• Mach-Zehnder modulator (MZM): a 40-ghz mach-zehnder modulator driven by a 64-gsa/s digital-to-analog
converter (DAC) with a 25-ghz analog bandwidth.
• Erbium Doped Fiber Amplifier (EDFA): used to amplify the optical signal after the MZM modulator.
• Variable Optical Attenuator (VOA): used to adjust the optical signal's power.
• Dispersion-uncompensated transmission link: consists of multiple fiber spans for transmitting the optical signal.
 Experimental Setup Cont…
Receiver side:
• Variable optical attenuator (VOA): adjusts the received optical power.
• Second erbium doped fiber amplifier (edfa): used to further amplify the received optical signal.
• Photodetector (PD): a 17-ghz photodetector that converts the received optical signal to an electrical signal.
• Digital storage oscilloscope (dsoz634a): an 80-gsa/s digital storage oscilloscope with a 33-ghz analog bandwidth. It is used to digitize the
received electrical signal.

Data processing:
• Training and testing data sets: random sequences generated by mixing multiple random sequences, with 100,000 samples used for training and another
100,000 samples used for testing.
• Digital signal processing (DSP): offline DSP includes resampling the digital waveform to 4 samples per symbol (4-sps) and passing it through a matched
filter to filter out out-of-band noise.
 Experimental Setup Cont…
Equalization models:
• One-dimensional convolutional neural network (1D-CNN): a neural network model designed for equalization
with optimal parameters.
• Two-dimensional convolutional neural network (2D-CNN): a neural network model designed for equalization
with optimal parameters.
• Fully connected neural network (FC): a neural network model designed for equalization with optimal
parameters.
• Volterra: a model introduced to measure equalization performance, designed according to the optimal
parameters.
Experimental Results and Discussion
Findings:
Effect of received power on BER (bit error rate):
• The BER results under different received powers ranging from -13 dbm to -8
dbm for back-to-back cases are shown in figure 5.
• All neural network-based (NN-BASED) equalizers exhibit decreasing BER
performance as the received power decreases. This decrease is due to
interference from single-end photodetector (PD) noise, including thermal
noise and shot noise, which cannot be eliminated by NN-based equalizers.
• The proposed TCNN equalizer performs similarly to other equalizers in terms
of BER performance.
 Experimental Results and Discussion Cont…
BER performance in 56-gbit/s PAM-470-km SSMF
transmission:
• Figure 7 displays the BER curves at different received powers in a 56-
gbit/s PAM transmission over a 470-km standard single-mode fiber
(SSMF) link.
• The TCNN equalizer, 2D-CNN equalizer, and 1D-CNN equalizer achieve
receiver sensitivities of approximately -9.5 dbm, -8 dbm, and -6 dbm,
respectively, at a 7% forward error correction (FEC) threshold (BER = 3.8
x 10^-3).
• The FCNN equalizer does not reach the 7% FEC threshold even with the
highest received optical power over a 70-km transmission link.
• TCNN outperforms 2d-cnn and 1d-cnn equalizers by improving the
receiver sensitivity by about 1.5 db and 3.5 db, respectively.
 Experimental Results and Discussion Cont…
Performance under different transmission distances:
• To evaluate the equalizer performance under varying transmission distances,
experiments were conducted.
• Figure 8 illustrates that the TCNN equalizer excels in the presence of significant
dispersion compared to the other four equalizers.
• Using a 7% FEC threshold (BER = 3.8 x 10^-3), the maximum transmission
distances achieved by 2D-CNN, 1D-CNN, and FCNN equalizers are 80 km, 72
km, and 38 km, respectively.
• The TCNN equalizer extends the maximum transmission distance to 100 km,
demonstrating its superior performance in mitigating dispersion.
 Experimental Results and Discussion Cont…
Complexity comparison:
• The complexity of the equalizer models is evaluated based on the number of multiplications required to execute them.
• The complexity of the proposed TCNN model is compared with other models, including fully connected neural networks (FCNN) and a three-order volterra model.
• The complexity ranking, from high to low, is as follows: 1D-CNN > 2D-CNN > TCNN > FCNN > volterra.
• Despite having a higher complexity than the volterra model, the TCNN equalizer's superior equalization performance justifies its adoption, striking a balance between
complexity and performance.

In summary, the TCNN equalizer demonstrates improved BER performance, especially in the presence of dispersion, and offers a favorable trade-off between complexity and
performance compared to other equalization methods. These findings make the TCNN equalizer a promising choice for optical communication systems.
 Experimental Results and Discussion Cont…
Analysis

The TCNN (temporal convolutional neural network) model exhibits significant performance advantages over traditional methods, particularly in optical communication
scenarios. The following points highlight the key reasons for the superior performance of the TCNN model
• Dispersion handling
• Neural network adaptability
• Improved receiver sensitivity
• Learning noise patterns

In conclusion, the TCNN model stands out in optical communication scenarios by effectively handling dispersion, adapting to diverse conditions, achieving better receiver
sensitivity, and striking a favorable tradeoff between complexity and performance. Its neural network architecture allows it to learn and adapt, making it a promising solution
for addressing the challenges posed by varying optical communication conditions.
 Conclusion
Summary of findings: the TCNN model effectively improves the performance of 56-gbit/s PAM-4
transmissions over 100-km SMF.
Implications: this study demonstrates the potential of using advanced neural network models in
enhancing high-speed optical communications.
 References
Citations:
Citation1: W. Yan et al., “80 km IM-DD transmission for 100 gb/s per lane enabled by DMT and nonlinearity management,” in proc. OFC, san francisco, CA, USA, 2014, pp. 1–3
Citation2: X. Tang, S. Liu, Z. Sun, H. Cui, and Y. Qiao, “c-band 56-gb/s PAM4 transmission over 80-km SSMF with electrical equalization at receiver,” opt. Exp., Vol. 27, pp. 25708–
25717, 2019.
Citation3: h. Wang et al., “Adaptive channel-matched detection for c-band 64- gbit/s optical OOK system over 100-km dispersion-uncompensated link,” IEEE/OSA J. Lightw. Technol.,
Vol. 38, no. 18, pp. 5048–5055, sep. 2020
Citation4: L. Sun, J. Du, C. Wang, Z. Li, K. Xu, and Z. He, “frequency-resolved adaptive probabilistic shaping for dmt-modulated IM-DD optical interconnects,” opt. Exp., Vol. 27, pp.
12241–12254, 2019.
Citation5: l. Tao, Y. Wang, J. Xiao, and N. Chi, “enhanced performance of 400 gb/s dml-based CAP systems using optical filtering technique for short reach communication,” opt. Exp.,
Vol. 22, pp. 29331–29339, 2014.
Citation6: d. Sadot, G. Dorman, A. Gorshtein, E. Sonkin, and O. Vidal, “single channel 112 gbit/sec PAM4 at 56 gbaud with digital signal processing for data centers applications,” opt.
Exp., Vol. 23, pp. 991–997, 2015.
Citation7: z. Li et al., “Signal-signal beat interference cancellation in spectrallyefficientwdm direct-detection nyquist-pulse-shaped 16-QAM subcarrier modulation,” opt. Exp., Vol. 23, pp.
23694–23709, 2015.
Citation8: a. Reza and J. Rhee, “nonlinear equalizer based on neural networks for PAM-4 signal transmission using DML,” IEEE photon. Technol. Lett., Vol. 30, no. 15, pp. 1416–1419,
aug. 2018.
Citation9: z. Xu, C. Sun, T. Ji, J. Manton, and W. Shieh, “computational complexity comparison of feedforward/radial basis function/recurrent neural networkbased equalizer for a 50-gb/s
PAM4 direct-detection optical link,” opt. Exp., Vol. 27, pp. 36953–36964, 2019.
Citation10: z. Xu, C. Sun, T. Ji, J. Manton, and W. Shieh, “cascade recurrent neural network-assisted nonlinear equalization for a 100 gb/s PAM4 short-reach direct detection system,” opt. Lett., Vol. 45, pp. 4216–4219, 2020.
Thank You