Learning to Tune Optical WANs: A Field Deployment of Noise Models in Optical Networks

Accurately modeling optical signal transmission is critical for optimizing network performance, particularly in large-scale fiber optic networks operated by Internet Service Providers. In this work, we develop a Gaussian Noise model for a New York state ISP's optical backbone. Our model accounts for all major network components, including amplifiers, fiber spans, reconfigurable optical add-drop multiplexers, and transceivers. By accurately predicting end-to-end signal-to-noise ratio, our model provides a foundation for network performance analysis and optimization. Then, we leverage hyperparameter search techniques-commonly used in machine learning-to identify amplifier gain settings that improve signal quality. By treating the model as an opaque box, we systematically search for amplifier configurations that maximize the predicted end-toend SNR while maintaining practical network constraints. We validate our approach through a field deployment by applying optimized amplifier gain settings in a live ISP network. Our results show a significant improvement in optical signal quality, achieving a 2 dB increase in SNR on a single wavelength 1 .

complex physical phenomena -like amplifier-induced noise on optical fiber -by representing the collective impact of these phenomenon as additive Gaussian-distributed noise. However, creating accurate GN models is challenging since it requires detailed knowledge of the network layout, the optical equipment used, and their configurations. Many of these important details are proprietary, making it challenging to build models that accurately reflect real-world network behavior. Model of production New York ISP. In this paper, we tackle this challenge by developing a high-fidelity GN model of NYSERnet. We gather comprehensive network topology information, operational parameters of optical equipment, and real-world signal quality measurements. To overcome gaps due to missing proprietary information, we approximate equipment using information in openly available vendor datasheets ( §3). Finally, we collect end-to-end signal quality measurements from the live network to compare the model's predicted signal quality against measured values and show that the model can accurately predict the end-to-end signal quality ( §5).

Field deployment of the model. Using the model as an opaque box, we then apply machine learning-inspired hyperparameter search techniques to optimize the configuration of amplifiers in the network. By systematically exploring the parameter space, we identify configurations that maximize the predicted end-to-end signal-to-noise ratio (SNR) while adhering to practical constraints ( §4). We deployed the optimized amplifier configuration in the production network and show that it leads to a measured signal quality increase of approximately 2 dB on targeted wavelengths. Artifact release. We are releasing the GN model we have developed and our comprehensive dataset -including network topology, equipment configuration and measured signal quality -to the research community. Code and data are available at https://github.com/artifact-release/ optical-model. By sharing these resources, we aim to support further research in the management and optimization of optical networked systems, ultimately improving reliability and efficiency across large-scale infrastructures.