Machine learning for digital twins of optical networking infrastructure

In core networks there is a huge quantity of metrology available ranging from SNR and BER measured by the transceivers through to the input and output powers of optical amplifiers and switches. The aim of this PhD project would be to use machine learning applied to metrology information, combined with the underpinning physics to create a digital twin of the physical optical network infrastructure.

Not only would the digital twin be updated in real-time, but it would also incorporate the uncertainty into the model allowing for real-time margin allocation to improve the operation of the optical network. Ultimately it would allow the optical network to operate at the limits of performance, based on measurements of the actual infrastructure, so as to allow the throughput as a function of availability to be optimised, to give the best user experience of a given installed infrastructure.

Key challenges the student will need to consider include:

1. The choice of the models for the digital twins (including whether it is a statistical model rather than purely a deterministic physics-based model) 2. The optimal means of incorporating measurement uncertainty into the models to create the digital twin

3. How to use the digital twin to give information regarding throughput, availability etc. so as to optimise the overall performance of the optical network 4. Ultimately, how to change current practice regarding the design and operation of the high availability

optical networks based on a data driven, probabilistic based approach, rather than the established margin-based approach.

The research strategy would start with a single element, such as an optical amplifier, transmitter or receiver and create a mathematical model that is refined based on measurements taken. Research currently carried out in Cambridge has demonstrated that using Gaussian Processes with priors based on a physical model allows for a physically informed machine learning approach.

While it is anticipated a Bayesian framework would be utilised, options of physical models and shades of grey, ranging from white box models to black box models would be explored. Having created the mathematical basis for a digital twin the statistical variation in the performance can be quantified and verified, with the complexity of the system increased until initially a digital twin is created for the 1200 km link in the Cambridge lab.

The application of the digital twin to optical infrastructure is a concept that is only just emerging.

Within the Cambridge University Engineering Department work is already underway on the National Digital Twin Programme https://www.cdbb.cam.ac.uk/what-we-do/national-digital-twin-programme, which we would hope to leverage as part of the same department as the PhD student and the PI.

While the project would create a methodology and software to enable a digital twin of the optical infrastructure to be formulated, ultimately the aim of the project would be to create a paradigm shift in the design and operation of the high-availability optical networks.