High-performance end-to-end deep learning IM/DD link using optics-informed neural networks.

Area of Science:

• Optical communications
• Deep learning
• Photonic device physics

Background:

• End-to-end deep learning (DL) models are increasingly used for optical transmission systems.
• Current models often utilize standard activation functions like ReLU, which may not fully leverage optical physics.
• Neuromorphic photonics offers inspiration for novel DL components.

Purpose of the Study:

• To introduce and experimentally validate optics-informed neural networks (NNs) for intensity modulation/direct detection (IM/DD) optical links.
• To investigate the performance benefits of an optics-inspired activation function, the Photonic Sigmoid, in DL models.
• To compare the proposed model against state-of-the-art ReLU-based configurations.

Main Methods:

• Development of an optics-inspired NN architecture incorporating a Photonic Sigmoid activation function derived from semiconductor nonlinear optics.
• Implementation of End-to-End DL configurations for fiber optic communication links.
• Extensive simulations and experimental validation of the proposed model's performance.

Main Results:

• Optics-informed NNs with the Photonic Sigmoid demonstrate superior noise and chromatic dispersion compensation compared to ReLU-based models.
• The Photonic Sigmoid NNs achieve performance below the hard-decision forward error correction (HD FEC) limit.
• Successful operation demonstrated for fiber lengths up to 42 km at 48 Gb/s.

Conclusions:

• Optics-informed NNs represent a promising approach to enhance DL performance in optical communication systems.
• The Photonic Sigmoid activation function offers significant advantages for IM/DD links, improving robustness and data rates.
• This work bridges neuromorphic photonics and DL for practical advancements in fiber optic transmission.