Multi-input neural channel waveform model for optical fiber WDM transmission based on Volterra series transfer...

Area of Science:

• Optical Communications
• Signal Processing
• Machine Learning

Background:

• Accurate optical fiber channel modeling is crucial for wavelength-division multiplexing (WDM) systems.
• Traditional split-step Fourier method (SSFM) is computationally inefficient.
• Existing deep learning models lack flexibility due to retraining requirements for varied system parameters.

Purpose of the Study:

• To develop a flexible and accurate deep learning model for WDM channel modeling.
• To overcome the limitations of current deep learning approaches in handling diverse system parameters.
• To improve the generalization capability of optical channel models.

Main Methods:

• Physics-based Volterra series transfer function algorithm.
• Neural network parameterization in the frequency domain (NN-VS).
• Evaluation through simulations of 40-channel and 5-channel WDM systems.

Main Results:

• NN-VS achieved high accuracy with an average Q-factor error below 0.15 dB.
• Demonstrated robust generalization across varying baud rates, dispersion, and nonlinearity coefficients.
• Showcased superior computational efficiency compared to SSFM, using <2% of real multiplications and achieving millisecond-scale GPU runtime.

Conclusions:

• NN-VS offers a highly accurate and flexible solution for optical fiber channel modeling in WDM systems.
• The model effectively handles multi-parameter inputs, reducing the need for retraining.
• NN-VS presents a computationally efficient alternative to SSFM for advanced optical communication network design.