ELBAGNN: Topology‐Unrolled GNN for Pre‐Routing Timing Prediction with Edge‐ and Level‐Biased Attention

Pre‐routing timing prediction is important for accelerating physical optimization, yet existing learning‐based methods still face two key limitations: they often adopt level‐synchronous or soft‐attention message passing that may violate the causal propagation semantics of static timing analysis, and they usually rely on discretized or hand‐crafted cell‐delay representations that cannot faithfully model the continuous dependence on input slew and output load. To address these issues, we propose ELBAGNN, a novel graph neural network integrating three key innovations: a topology‐unrolled causal schedule to prevent information leakage, differentiable lookup table embeddings to capture slew/load dependencies, and a biased attention mechanism to guide information flow. Our model demonstrates superior performance, improving the unified timing accuracy metric from 0.933 to 0.943 and cutting per‐epoch training time by half across 21 real‐world circuits. These results indicate ELBAGNN's efficacy as a fast, practical solution for pre‐routing timing estimation, even for challenging large‐scale, long‐chain designs.