Priority-Aware Routing Optimization with Full-Path Delay Prediction Using Machine Learning

Recent research has investigated the use of machine learning (ML) to predict timing characteristics at the pre-routing stage, effectively bridging the gap between early-stage estimations and final design outcomes. This enables more efficient and informed timing optimization throughout the design flow. However, prior work on timing prediction has primarily focused on estimating endpoint slack for only the most critical or longest timing paths, which limits its utility for comprehensive timing analysis. Moreover, the integration of timing prediction models into downstream design optimization remains unexplored. In this paper, we propose an ML-based framework that consists of a pre-route path delay prediction model and a subsequent routing framework that fully exploits the prediction outcomes. More specifically, the prediction model constructs timing paths by incorporating all intermediate cells between the start and end flip-flops, using these complete paths as input to the training dataset. It also considers multiple possible paths between the same pair of flip-flops, which enhances the robustness and accuracy of the delay prediction. Based on this prediction model, we classify nets as timing-critical and non-critical, and guide the commercial routing engine to prioritize the routing of timing-critical nets, which significantly improves overall timing quality. Experimental results demonstrate that our delay prediction model improves path delay prediction accuracy by 17% over commercial tools. Additionally, our routing framework achieves 6% and 13% improvements in total negative slack (TNS) and worst negative slack (WNS), respectively.