Aristotle University Of Thessaloniki (AUTH)
Addressing hardware limitations in silicon photonics for scalable optical interconnects and neural networks
Abstract
dc:descriptionThe rapid growth of data-center applications, including AI (Artificial Intelligence) and cloud computing, has placed unprecedented demands on communication and computing infrastructures, particularly in terms of bandwidth, energy efficiency, and latency. Silicon photonics has emerged as a promising technology to address these challenges, offering high bandwidth density, low energy per bit, and compatibility with mature CMOS (Complementary Metal-Oxide-Semiconductor) fabrication processes. However, the deployment of photonic systems at scale remains constrained by a range of system-level challenges, including thermal sensitivity of resonant devices, susceptibility to analog noise and variability, and the mismatch between photonic hardware characteristics and conventional digital signal processing and machine learning frameworks. This thesis addresses these challenges through a system-level approach that integrates photonic hardware, control mechanisms, and learning algorithms. The first contribution presents a dual-feedback stabilization framework for ring-based silicon photonic transmitter architectures, enabling robust operation of both passive and active resonant devices under temperature variations and wavelength drift. The proposed approach leverages ContactLess Integrated Photonic Probes (CLIPPs) and introduces a second derivative-based locking mechanism, achieving stable performance in wavelength-division multiplexed interconnect systems, which open the path for their implementation in data center applications. The second contribution focuses on improving the robustness of Photonic Neural Networks (PNNs) against hardware-induced noise and limited bit precision. A noise-resilient training and inference methodology is introduced, incorporating dual-phase quantization and activation-function slope adaptation to align neural network operation with the physical constraints of analog photonic hardware. Experimental validation using a photonic–electronic multiply–accumulate neuron (PEMAN) demonstrates improved classification accuracy and improved performance, compared to conventional neural network configurations, under increasing noise and bandwidth limitations. The third contribution introduces optics-informed neural networks for optical communication systems, in which activation functions are derived directly from the nonlinear response of optical devices. By embedding these physically informed nonlinearities into an end-to-end deep learning framework for intensity-modulated direct-detection (IM/DD) links, the proposed approach enables more effective compensation of chromatic dispersion, nonlinearities and noise. Both simulation and experimental results demonstrate significant improvements in bit-error-rate (BER) performance and transmission reach compared to conventional neural network implementations based on digital activation functions. Collectively, these contributions establish a framework for scalable photonic systems based on the co-design of hardware, control, and learning layers. This work bridges some of the gaps between photonic device innovation and practical deployment, enabling more robust, efficient, and adaptive photonic platforms for next-generation communication and computing applications.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Roumpos, Ioannis
- Ρούμπος, Ιωάννης
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62399