Interpretable model-driven end-to-end deep learning of geometric shaping for DML-based IM/DD systems.

Area of Science:

• Optical Communications
• Nonlinear Optics
• Machine Learning in Photonics

Background:

• Direct modulation lasers (DMLs) are cost-effective for short-reach optical systems.
• Chirp-dispersion interaction causes nonlinear distortion, limiting DML performance.
• Neural network (NN) autoencoder (AE)-based geometric shaping (GS) offers constellation optimization but requires extensive training data.

Purpose of the Study:

• To develop a low-complexity, model-driven framework for optimizing geometric constellation in DML systems.
• To mitigate nonlinear distortion caused by chirp-dispersion interaction.
• To combine physics-based interpretability with deep learning for adaptive channel modeling.

Main Methods:

• Proposed a model-driven framework using composite second-order (CSO) distortion theory to create a surrogate channel model.
• Employed AE-based geometric constellation optimization on the surrogate channel.
• Validated the approach through experimental transmission over 10-km standard single-mode fiber (SSMF).

Main Results:

• Achieved suppression of chirp-dispersion interaction-induced nonlinear distortions.
• Demonstrated a 1-dB receiver sensitivity improvement for 64-QAM signals.
• Confirmed the effectiveness of the physics-informed deep learning approach.

Conclusions:

• The proposed low-complexity framework effectively mitigates nonlinear distortions in DML systems.
• Combining model-driven insights with deep learning offers an efficient solution for constellation optimization.
• This approach enhances the performance of short-reach optical communication systems.