Binomial training algorithm for neuromorphic photonics applied to channel equalization

Area of Science:

• Photonics
• Optical Communications
• Artificial Intelligence

Background:

• Optical communication systems face signal degradation over long distances.
• Channel equalization is crucial for maintaining signal integrity.
• Existing equalization methods can be complex and computationally intensive.

Purpose of the Study:

• To propose and evaluate a novel photonic time-delayed neural network for channel equalization.
• To demonstrate the effectiveness of a binomial training algorithm for this optical neural network.
• To assess the performance and generalization capabilities of the developed system.

Main Methods:

• Development of a photonic time-delayed neural network architecture using delay lines, phase, and amplitude modulators.
• Numerical implementation and training using experimentally acquired data from a 10 Gb/s optical transmission system.
• Evaluation of bit error rate (BER) and eye diagram improvements after equalization.

Main Results:

• Significant improvements in BER and eye diagram for 10 Gb/s on-off keying (OOK) signals over 100 km of single-mode fiber.
• The photonic neural network successfully restored highly degraded signals, opening closed eye diagrams.
• 93.3% of transmitted bit sequences achieved performance below the forward error correction (FEC) limit, indicating strong generalization.

Conclusions:

• The proposed photonic time-delayed neural network with binomial training offers an efficient solution for optical channel equalization.
• The discrete, hardware-friendly optimization is suitable for photonic integrated circuits.
• The system demonstrates excellent performance and generalization for real-world optical transmission challenges.