.png){kind=icon}
.png){kind=icon}
Return
Photonic Spiking Graph Neural Network for Energy-Efficient Structured Data Processing
.png){kind=icon}
PRE
AI
W
S
X
S
H
X
Y
H
Y
DOI：10.1109/jlt.2026.3703120.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Photonic computing exhibits broad application potential in signal processing and artificial intelligence (AI) acceleration owing to its high computational speed, low energy consumption, and intrinsic parallelism. Existing photonic computing research has predominantly concentrated on convolutional neural networks (CNNs) and fully connected neural networks (FCNNs), which are typically applied to tasks such as image classification and object detection. However, these traditional network architectures face inherent limitations in effectively processing graph-structured data. Graph neural networks (GNNs), specifically designed for modeling and analyzing graph-structured data, effectively capture the complex relationships among data points. We propose an computational architecture termed the photonic spiking graph neural network (PSGNN). The proposed architecture integrates the efficient structural modeling capability of GNNs, the temporal dynamics of spiking neurons, and the ultra-high speed and low power consumption advantages of photonic chips for parallel computation. Through hardware–software co-optimization, a bias-term simulation mechanism tailed for photonic computing chip is implemented via a feature-dimension expansion strategy, thereby facilitating effective model training. Experimental results on the KarateClub (PubMed) datasets demonstrate a training accuracy of 100% (92 ± 2%) and a test accuracy of 97% (90 ± 1%). A hardware experimental platform based on a silicon-photonics-based 4 × 4 Mach–Zehnder interferometer (MZI) array was constructed to perform classification experiments on the KarateClub datasets, achieving a test accuracy of 93%, thereby validating the effectiveness of the proposed architecture in structured-data processing tasks. The proposed system achieved an inference latency of 97 ps, demonstrating remarkable acceleration within the optical domain. The photonic linear computation achieved an energy efficiency of 280 GOPS/W and a computational density of 52 GOPS/mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>. These results underscore the potential of PSGNN for deployment across multiple domains, including dynamic topology analysis of social networks, intelligent traffic scheduling and early-warning systems, and fault diagnosis in industrial IoT.
Keywords:
Energy efficiency
Mach–Zehnder interferometer
photonic spiking graph neural network
Journal
IF:
4.8
Papers:
1.7W
Citations:
3.8W
Organization
Cited Papers