Fusing Physics to Fiber Nonlinearity Model for Optical Networks Based on Physics-Guided Neural Networks

Machine learning (ML) has been widely used for physical layer modeling in optical networks for its high accuracy and efficient calculation structure. However, traditional ML-based methods purely rely on the available data rather than physical laws, resulting in relatively low reliability and generalizability. To further improve the modeling performance and speed up the deployment of ML-based models in real systems, one possible solution is to combine ML-based methods and physical knowledge so that the accuracy and reliability of the model can be both improved. In this paper, we propose a fiber model based on the physics-guided neural networks (PGNN). First, the rough estimation of an analytical model is integrated into the input feature of a ML-based model as calibration to reduce the learning burden of ML engines. In addition, the known physical law is manually exploited and designed as a physical regularization based on a semi-supervised learning structure to improve physical consistency. To demonstrate the performance improvement in accuracy, physical consistency, and generalizability, a simulation validation under different link and signal configurations is conducted. Moreover, the PGNN-based model is compared with traditional neural networks in an experimental link to further illustrate its superior accuracy and generalizability.