FusionCell: Cross-Attentive Fusion of Layout Geometry and Netlist Topology for Standard-Cell Performance Prediction

In three linesFusionCell predicts standard-cell performance by fusing routed layout geometry (DeiT encoder) and netlist topology (graph transformer). Trained on 19.5k 7nm cells (ASAP7), the model achieves 0.92% MAPE on delay/power metrics and accelerates characterization by orders of magnitude versus circuit simulation.

Summary generated by Claude — human-verified