Enhancing generalization in neural network-based waveform-level channel modeling for optical fiber transmission...

Area of Science:

• Optical Communications
• Computational Photonics
• Machine Learning in Engineering

Background:

• Accurate optical fiber channel waveform modeling is critical for optical communication systems.
• Traditional methods like split-step Fourier method (SSFM) are computationally intensive.
• Neural network (NN) approaches offer reduced computational load with comparable accuracy.

Purpose of the Study:

• To improve the generalization capability of NNs for optical communication system modeling.
• To develop a novel parameter encoding structure for enhanced NN performance.
• To create a single NN capable of generalizing across diverse system parameters.

Main Methods:

• Introduction of a novel parameter encoding structure for NNs.
• Pre-encoding of system parameters to enhance NN generalization.
• Training and validation of a single NN across multiple optical system parameters.

Main Results:

• The parameter encoding structure significantly improves NN generalization.
• Waveform modeling accuracy increased by 49.9% and 69.7% in generalized scenarios.
• A single NN demonstrated generalization across modulation format, symbol rate, WDM channel space, laser parameters, dispersion, span length, and distance.

Conclusions:

• The proposed parameter encoding structure substantially enhances NN generalization for optical waveform modeling.
• A single, generalized NN was developed for the first time, covering multiple system parameters simultaneously.
• This enhanced NN offers significant potential for optimizing optical transmission system design.