[Paper Review] Accurate Closed-Form Real-Time EGN Model Formula Leveraging Machine-Learning over 8500 Thoroughly Randomized Full C-Band Systems

Mahdi Ranjbar Zefreh, F. Forghieri+2 more|arXiv (Cornell University)|May 29, 2020

Optical Network TechnologiesEngineering24 references47 citations

TL;DR

This paper proposes CFM4, a highly accurate, real-time closed-form model for non-linear interference (NLI) in C-band WDM optical systems, derived by enhancing a base formula with machine learning over 8500 randomized system configurations. By integrating physical parameters and optimized coefficients, the model achieves near-perfect agreement with the numerically integrated EGN model and split-step simulations, with sub-0.2 dB SNR error across diverse scenarios.

ABSTRACT

We derived an approximate non-linear interference (NLI) closed-form model (CFM), capable of handling a very broad range of optical WDM system scenarios. We tested the CFM over 8500 randomized C-band WDM systems, of which 6250 were fully-loaded and 2250 were partially loaded. The systems had highly diversified channel formats, symbol rates, fibers, as well as other parameters. We improved the CFM accuracy by augmenting the formula with simple machine-learning factors, optimized by leveraging the system test-set. We further improvedthe CFM by adding a term which models special situations where NLI has high self-coherence. In the end, we obtained a very good match with the results found using the numerically-integrated Enhanced GN-model (or EGN-model). We also checked the CFM accuracy by comparing its predictions with full-C-Band split-step simulations of 300 randomized systems. The combined high accuracy and very fast computation time (milliseconds) of the CFM potentially make it an effective tool for real-time physical-layer-aware optical network management and control.

Motivation & Objective

• Address the lack of accurate, real-time NLI models for diverse, realistic optical WDM systems with complex parameters like dispersion slope and frequency-dependent loss.
• Improve upon existing closed-form models (CFM0/CFM1) that fail to capture key physical effects such as NLI coherence and system diversity.
• Develop a scalable, computationally efficient model suitable for real-time physical-layer-aware network control and optimization.
• Leverage a large, diverse dataset of 8500 randomized C-band systems to train machine-learning corrections that enhance accuracy without sacrificing speed.

Proposed method

• Derive CFM1 as an enhanced closed-form approximation of the incoherent GN-model, incorporating dispersion slope, frequency-dependent loss, and non-uniform spans.
• Train machine-learning correction factors using a big-data set of 8500 randomized C-band WDM systems, optimizing coefficients against the EGN-model benchmark.
• Introduce a coherence-correction term to address high-NLI-coherence outlier systems, improving peak error performance.
• Refine the model to include channel roll-off effects, resulting in CFM4, which accounts for spectral shaping in practical transmission systems.
• Validate the final model using full C-band split-step simulations on 300 randomly selected systems from the test set.
• Use statistical sensitivity analysis to confirm the necessity of each physical parameter (e.g., β2,acc, Φ, symbol rate) in the machine-learning correction factors.

Experimental results

Research questions

• RQ1 Can a closed-form NLI model be made both accurate and real-time across a wide range of realistic, diverse C-band WDM system configurations?
• RQ2 How effective is machine learning over a large dataset in correcting systematic biases and variance in a physics-based NLI approximation?
• RQ3 What role does NLI self-coherence play in model error, and can it be effectively modeled in a closed-form framework?
• RQ4 To what extent do physical parameters like accumulated dispersion, modulation format, and roll-off influence the accuracy of the NLI estimation?
• RQ5 Does the final model (CFM4) maintain high accuracy when compared to full split-step simulations, confirming its reliability for real-world deployment?

Key findings

• CFM4 achieves a mean SNR estimation error of -0.00 dB and a standard deviation of 0.04 dB when compared to the EGN-model benchmark across 8500 systems.
• The peak SNR error was reduced to 0.18 dB, and the peak-to-peak error to 0.30 dB, indicating high consistency and reliability.
• The addition of the NLI coherence correction term reduced the peak error significantly, especially in outlier systems with high self-coherence.
• Sensitivity analysis showed that removing either accumulated dispersion (β2,acc) or the format-dependent coefficient (Φ) increased the error variance by over 300% and peak error by up to 3x.
• The model maintained high accuracy when validated against 300 full C-band split-step simulations, confirming its fidelity to full numerical integration.
• The machine-learning correction factors were found to be non-redundant: removing any component degraded performance, confirming the necessity of each term in the analytical form.