Transient Modeling of High-Speed Links Based on GAT-DNN-LSTM Hybrid Neural Network

Abstract

With the rapid advancement of high-speed digital communication technology, signal integrity (SI) issues in high-speed links have become increasingly prominent. To address the high computational complexity and low simulation efficiency of traditional electromagnetic simulation methods in multi-topology and multi-parameter scenarios, this paper proposes a surrogate modeling approach based on a hybrid model integrating Graph Attention Networks (GAT), Deep Neural Networks (DNN), and Long Short-Term Memory (LSTM) networks. First, this method utilizes GAT to extract global features from the topological structures and physical parameters of key link components, including drivers, packages, and channels. Subsequently, the DNN module performs nonlinear fusion of global features from the three components, forming a unified global representation that characterizes the electrical properties of the entire high-speed link. Finally, the global representation of the link is input into the LSTM network alongside the differential excitation waveform, enabling high-precision prediction of the transient differential response for high-speed links. Experimental results demonstrate that the proposed model achieves prediction accuracy comparable to traditional circuit simulation methods while significantly reducing computational cost. Compared with existing machine learning-based surrogate models, this approach effectively alleviates the limitations of fixed topologies and parameters, providing a highly accurate, efficient, and scalable modeling pathway for SI analysis, design space exploration, and simulation acceleration in high-speed links.