PhD defence by Sergio Hernandez Fernandez

Abstract

As our world becomes more connected, the data centers that power cloud computing and AI face constant pressure to handle more traffic, more efficiently.

A key component in this infrastructure is the directly modulated laser (DML), a device prized for being affordable and energy-efficient. However, when pushed to very high speeds, DMLs can distort the signals they send, limiting performance. Traditionally, engineers have tried to fix this by digitally cleaning up the signal only at the receiving end, an approach with limited success as it can lead to sub-optimal solutions.

This research introduces a method called end-to-end (E2E) optimization, which treats the transmitter and receiver as a single, intelligent system that learns to cooperate. The core of the technique is to first create a highly accurate “digital twin” of the laser and fiber link using a specialized AI model.

This differentiable surrogate model enables an optimization algorithm to simultaneously fine-tune every part of the communication link—from shaping the signal at the start to adjusting the laser's power and configuring the receiver's filters. In essence, the system learns to intelligently shape the signal so that it arrives at its destination in a much cleaner, more decipherable form.

Supervisors

• Main Supervisor: Professor Darko Zibar, DTU Electro
• Co-supervisor: Associate Professor Francesco Da Ros, DTU Electro.
• Co-supervisor: Christophe Peucheret, Univ Rennes, CNRS, UMR6082 - FOTON, 22305 Lannion, France.

Assessment committee

• Professor, Søren Forchhammer, DTU Electro, Denmark (chair).
• Professor Roberto Gaudino, Politecnico Di Torino, Italy
• Associate Professor Oskars Ozoliņš, Tenured Full Professor at Riga Technical University, Latvia

Master of the Ceremony

• Associate Professor, Michael Galili, DTU Electro, Denmark.