Self-controlling photonic-on-chip networks with deep reinforcement learning

Area of Science:

• Photonics
• Optical Networking
• Integrated Circuits

Background:

• Current bidirectional photonic switches limit Photonic-on-Chip Network (PCN) scalability to linear chains.
• There is a need for more flexible and scalable optical switching solutions for modern data centers and intra-chip communication.

Purpose of the Study:

• To introduce a novel four-degree photonic chip design for high-bandwidth optical switching.
• To enable the construction of scalable, multi-dimensional PCNs.
• To develop a self-controlling optimization model for PCNs.

Main Methods:

• Design of novel four-degree photonic chips with low insertion loss, low crosstalk, and low power consumption.
• Integration of these chips into a full-grid PCN architecture.
• Implementation of a Multi-Sample Discovery model (deep reinforcement learning based on Proximal Policy Optimization) for network self-control.

Main Results:

• The proposed four-degree photonic chips offer superior flexibility and scalability compared to bidirectional switches.
• The developed PCN architecture supports high-dimensional switching with low insertion loss and short switching times.
• The Multi-Sample Discovery model effectively optimizes PCN criteria like transmission loss, power consumption, and routing time.

Conclusions:

• The novel four-degree photonic chip design facilitates the creation of advanced, scalable, and self-controlling PCNs.
• This architectural innovation and optimization approach are crucial for future intra-chip communication and photonic data centers.
• The proposed self-controlling PCNs demonstrate effectiveness and scalability for dynamic network environments.