Physics-Based Deep Learning for Optical Data Transmission and Distributed Sensing

In order to cope with ever-increasing data traffic demands, recent years have witnessed an explosion of research interest in replacing handcrafted communication algorithms with machine-learning-based artificial intelligence (AI).

However, prior work has largely resorted to massively overparameterized neural networks, resulting in unrealistic hardware requirements and unsatisfactory black-box solutions.

This project will instead explore a new approach, specifically for optical fiber systems, that will not only lead to improved transmission performance, but also transform already-deployed communication fibers into intelligent sensors providing real-time information about various physical effects along the propagation path, e.g., mechanical stresses or vibrations.

Our approach leverages well-established physical models and principles as a foundation, in particular the nonlinear differential equations governing the propagation dynamics.

The proposed 4-year project is divided into four objectives/tasks: (i) receiver-side machine-learning models, (ii) transmitter-side signal shapers, (iii) data-driven optimization, and (iv) trade-offs between communication and sensing.

The data used for training the developed solutions will originate from simulations supplemented by lab transmission experiments.

Our target is to develop new AI-based solutions that combine, for the first time, high-speed optical data transmission and distributed sensing capabilities into a single DSP platform.