A Hybrid SRU-Transformer-Conv1d Architecture for High-Speed Signal Transient Prediction With Uncertainty Quantification

Time domain transient simulation is crucial in signal integrity analysis of high-speed serial link design. In this article, we propose a novel hybrid neural architecture combining simple recurrent unit (SRU), Transformer, and 1-D Convolution (Conv1d) layers for accurate and uncertainty-aware transient waveform prediction in high-speed serial links. In this hybrid neural architecture, SRU effectively models local temporal dependencies, the Transformer encoder captures global context, and the Conv1d module enhances fine-grained pattern extraction in the output stage. Compared to other time-series prediction models, this method performs better in waveform fitting accuracy and shape preservation. To estimate prediction uncertainty, Monte Carlo dropout is applied during inference, offering reliable confidence intervals of ±2σ. In addition, the experiments shows that it also has high practical value in eye diagram prediction.