Task-Adaptive Dual-Graph Neural Network for Joint Congestion and Timing Prediction

Early physical-design decisions benefit from jointly estimating routing demand and timing risk, but a netlist edge and a geometric-neighbor edge encode different phenomena. We present TD-DGNN, a task-adaptive dual-graph predictor with separate spatial and netlist GraphSAGE branches, learned congestion- and timing-specific fusion gates, and a directed congestion-conditioned timing head. To isolate architectural effects, we construct 144 independently generated synthetic placed directed-acyclic circuits (18,311 cells), using a rectangular uniform wire-density estimator and a congestion-detour longest-path timing engine. Nine configurations, including a capacity-matched baseline, are trained with three seeds and assessed on 24 unseen circuits. TD-DGNN attains mean per-design timing correlation 0.855 and violation ROC-AUC 0.825 versus 0.814 and 0.810 for a parameter-matched single-graph reference; the paired timing gain is 0.041 (95% bootstrap interval 0.032–0.050). Without retraining, on 24 larger synthetic circuits it attains correlation 0.836 versus 0.797 for that reference. Directional predecessor aggregation adds 0.0063 timing correlation relative to undirected coupling. The independent effect of adaptive gates and congestion-conditioning is smaller on this benchmark. These results establish a controlled test of graph separation rather than an estimate of commercial sign-off accuracy.