102.3 Tbit/s C+L-band transmission over 1512 km SSMF enabled by an uncertainty-aware Bayesian CNN-BiLSTM hybrid...

Area of Science:

• Optical Communications
• Machine Learning in Telecommunications

Background:

• Dense Wavelength Division Multiplexing (DWDM) systems face significant transmission impairments over long distances, including fiber nonlinearity and inter-symbol interference.
• Wideband optical transmission systems require sophisticated equalization techniques to maintain signal integrity and high data rates.

Purpose of the Study:

• To propose and experimentally validate a novel hybrid equalizer combining Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and a Bayesian head.
• To mitigate transmission impairments in high-capacity DWDM systems over standard single-mode fiber (SSMF).

Main Methods:

• Implementation of a hybrid equalizer integrating CNN for local feature extraction and BiLSTM for temporal modeling.
• Utilizing a Bayesian head for uncertainty-aware residual prediction to improve equalization accuracy.
• Experimental demonstration over a 1512 km SSMF link with a 12.2 THz bandwidth DWDM system.

Main Results:

• Significant improvement in Generalized Mutual Information (GMI)-estimated throughput from 97.1 Tbit/s to 102.3 Tbit/s.
• Reduction in training epochs by 80% compared to conventional neural network methods.
• Achievement of a spectral efficiency (SE) of 8.4 bit/s/Hz and an average single-carrier capacity of 838.5 Gbit/s.

Conclusions:

• The proposed Bayesian CNN-BiLSTM hybrid equalizer effectively mitigates impairments in wideband long-haul optical transmission.
• The hybrid equalizer enhances system throughput and spectral efficiency while reducing training overhead.
• This approach validates the effectiveness of advanced machine learning techniques for future high-capacity optical communication systems.