A Novel Surrogate Modeling Framework Based on Fast Time-Domain Simulation and Transfer Learning for Signal Integrity Analysis

In high-speed electronic systems, signal integrity (SI) issues become increasingly prominent as the operating frequencies rise and integration levels deepen, making accurate SI modeling a key requirement. This paper presents a modeling framework that combines fast time-domain simulation with transfer learning to reduce the computational cost of transient and electromagnetic simulations while building high-accuracy surrogate models. In the source domain, circuit simulation along with the multiple edge response method is used to generate large-scale transient waveforms and worst-case eye diagrams for pretraining neural networks. The target domain uses S-parameters extracted from electromagnetic simulation to construct interconnect models, and transient simulation is performed to obtain high-fidelity output responses. Subsequently, deep neural network and temporal convolutional network are used as surrogate models for geometric parameter-to-eye diagram metrics mapping and input-output waveform mapping tasks, respectively, where the eye diagram metrics include eye height, eye width, and jitter peak-to-peak. The knowledge learned from the source domain is transferred to transient simulation data in the target domain using either fine-tuning or partial-freezing strategies, which further enhance the performance under limited data scenarios, thereby validating the framework’s adaptability and robustness across diverse tasks and system configurations.