Photonic spiking DQN-GAT architecture for intelligent routing

Intelligent routing plays a crucial role in modern communication infrastructures, including data centers, Internet of Things (IoT), and satellite communication networks. Although reinforcement learning (RL) has shown significant promise for intelligent routing, its practical deployment remains limited by high energy consumption and decision-making latency. Here, we propose a photonic spiking architecture that integrates a deep Q-network (DQN) and a graph attention network (photonic spiking-DQN-GAT architecture) to enable low-latency and energy-efficient routing decisions. A systematic performance evaluation was conducted on the Waxman network topology. The results show that the routing strategy based on the photonic spiking-DQN-GAT architecture, with a routing accuracy of 100%, significantly outperforms traditional open shortest path first (OSPF) and equal-cost multi-path (ECMP) algorithms in key metrics. Compared to OSPF/ECMP, the proposed architecture achieves 82.34%/81.86% lower latency and 66.64%/24.96% higher bandwidth. Furthermore, we implemented a hardware-software co-design architecture for the activation layer using a distributed-feedback laser with integrated saturable absorber (DFB-SA). Experiment results show that the power consumption per inference is 1.685 μJ/inf, with a latency of 191.20 ps/inf, demonstrating significant improvements in both energy efficiency and inference speed. The proposed photonic spiking-DQN-GAT architecture provides a pathway toward low-latency and energy-efficient intelligent routing.