Inverse regular perturbation with ML-assisted phasor correction for fiber nonlinearity compensation

Area of Science:

• Optical communication systems
• Machine learning applications in signal processing
• Nonlinear fiber optics

Background:

• Digital back-propagation (DBP) is crucial for mitigating nonlinear impairments in optical fiber systems.
• Traditional DBP methods, often based on the split-step Fourier method (SSFM), require significant computational resources.
• Inverse regular perturbation (RP) offers an alternative model for managing nonlinear effects.

Purpose of the Study:

• To enhance the inverse regular perturbation (RP) model using machine learning (ML) for improved performance in optical fiber communications.
• To develop a learned RP (LRP) model capable of jointly optimizing key parameters like step-size, gain, and phase rotation.
• To compare the performance of the LRP model against established ML-based DBP methods.

Main Methods:

• Implementation of a machine learning technique to improve the inverse regular perturbation (RP) model.
• Joint optimization of step-size, gain, and phase rotation for individual RP branches within the learned RP (LRP) model.
• Utilizing fractional step-per-span (SPS) modeling for complexity reduction in the LRP approach.

Main Results:

• The proposed learned RP (LRP) model demonstrated superior performance compared to the learned digital back-propagation (DBP) method.
• Achieved up to 0.75 dB gain in an 800 km standard single mode fiber link.
• The LRP model maintained superior performance even with fractional step-per-span (SPS) modeling compared to 1-SPS SSFM-DBP.

Conclusions:

• The learned RP (LRP) model represents a significant advancement in mitigating nonlinear impairments in optical fiber systems.
• LRP offers a computationally efficient alternative to traditional DBP methods, providing enhanced performance.
• The LRP model's ability to optimize parameters and handle fractional SPS modeling opens new avenues for high-performance optical communication.