SpecWave: Spectral-Enhanced Graph Learning for Crosstalk-Aware Waveform Prediction
Year: 2026
Type: conference-paper
Abstract: With technology nodes continue to scale, crosstalk effect has become indispensable to accurate static timing analysis (STA). Existing machine learning-based approaches struggle to achieve high-fidelity waveform prediction under crosstalk due to the intricate time-domain characteristics of signal waveforms. To address these challenges, we propose SpecWave, a spectral-enhanced graph learning framework for crosstalk-aware waveform prediction. SpecWave first extracts spectral representations of crosstalk waveforms in the frequency domain via the Fast Fourier Transform. Hierarchical graph learning is then employed to capture the input-specific transfer function. Subsequently, high-precision output waveforms are reconstructed via a frequency-domain fusion mechanism and a differentiable inverse transform. Experimental results on open-source designs demonstrate that SpecWave achieves an average waveform error of 3.04%, while delivering a 40.94× speedup over HSPICE and a 3.40× acceleration compared to commercial STA tool.
Authors: Yanlong Mao, Jiajie Xu, Bowen Shi, Yaning Jia, Hao Yan
Institutions: Southeast University, Southeast University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224913481