Computation efficient sparse DNN nonlinear equalization for IM/DD 112  Gbps PAM4 inter-data center optical...

Area of Science:

• Optical communication systems engineering
• Signal processing
• Machine learning applications

Background:

• High-speed optical communication systems face nonlinear distortions.
• Deep neural networks (DNNs) show promise for equalization but can be computationally intensive.
• Existing methods like Volterra equalizers (VE) have limitations in complexity and performance.

Purpose of the Study:

• To propose and demonstrate a novel sparse deep neural network-based nonlinear equalizer (SDNN-NLE).
• To reduce computational complexity while maintaining high transmission accuracy.
• To mitigate nonlinear distortions in high-speed optical links.

Main Methods:

• Developed a two-phase pruning strategy for DNN weights: adaptive L2-regularization for significance identification and pre-defined sparsity for pruning.
• Implemented and experimentally validated the SDNN-NLE on a 112 Gbps PAM4 optical link.
• Compared performance against conventional fully connected Volterra equalizers and DNN-NLEs.

Main Results:

• The SDNN-NLE achieved significant complexity reductions: 71% vs. VE, 63% vs. fully connected DNN-NLE, and 41% vs. sparse VE.
• Demonstrated effective mitigation of nonlinear distortions on a 112 Gbps PAM4 signal over 40 km standard single-mode fiber.
• Maintained system performance without degradation compared to unpruned counterparts.

Conclusions:

• The proposed SDNN-NLE offers a computationally efficient and effective solution for mitigating nonlinear distortions in high-speed optical communication.
• This sparse DNN approach presents a promising advancement over traditional equalization techniques.
• The method successfully balances complexity reduction and transmission accuracy in demanding optical links.