Efficient nonlinear equalization across S+C+L bands using online knowledge distillation in high-capacity DWDM systems

Area of Science:

• Optical Communications
• Signal Processing
• Machine Learning

Background:

• High-symbol-rate dense wavelength division multiplexing (DWDM) systems face nonlinear equalization challenges.
• Existing methods struggle with the complexity of triple-band (S+C+L) equalization.

Purpose of the Study:

• To propose an online knowledge distillation method for efficient nonlinear equalization in DWDM systems.
• To improve net data rate and reduce computational complexity across S+C+L bands.

Main Methods:

• Utilized a teacher-student model for knowledge distillation.
• Trained teacher models on representative C-band channels.
• Distilled knowledge to lightweight student models for S+C+L band equalization.
• Employed 98 GBaud PDM-PCS 256-QAM signals over 150 km SSMF.

Main Results:

• Achieved a net data-rate improvement from 201.5 Tb/s to 231.6 Tb/s.
• Reduced computational complexity by up to 97.1% in the C band and 80% in S and L bands.
• Demonstrated comparable performance between student and teacher models.

Conclusions:

• The proposed online knowledge distillation method is effective for nonlinear equalization in high-symbol-rate DWDM systems.
• This approach offers significant improvements in data rate and computational efficiency.
• Lightweight student models can achieve high performance with reduced complexity.