High-Speed Link System Surrogate Modeling and Co-Optimization Based on Multimodal Machine Learning and GraphVAE

With the continuous advancement of very-large-scale integra tion technology, high-speed interconnect systems are increasingly challenged by signal integrity (SI) issues arising from feature size scaling and higher data transmission rates. Traditional simulation-based methods incur high computational costs and suffer from low efficiency in SI analysis, making them inadequate for meeting the demands of modern design processes. This article proposes a transient surrogate modeling approach for high-speed link systems based on multimodal machine learning (MML). Long short-term memory (LSTM) networks, graph convolutional networks (GCNs), and deep neural networks (DNNs) are employed to extract modal features from the transmitter, channel, and receiver, respectively. A cross-channel attention mechanism is introduced to achieve adaptive fusion of multimodal features, thereby enhancing prediction accuracy. Furthermore, a collaborative optimization framework is developed by integrating a graph variational autoencoder (GraphVAE) with Bayesian optimization (BO). With the support of the surrogate model, the framework enables efficient performance evaluation and co-optimization of link topology and physical parameters. The proposed method significantly improves design efficiency and shortens the design cycle. It is demonstrated that the approach achieves high modeling accuracy while substantially reducing simulation costs, indicating its strong potential for application in high-speed link system design.