Authors: Vikas Raturi, M. Zahid Alam
Abstract: Optimizing the physical routing of electronic circuits is one of the hardest problems in semiconductor design, and it has a direct impact on circuit performance, power dissipation and manufacturability. Billion-transistor designs run up against the fundamen-tal computational limits of traditional routing algorithms. We present a new hybrid explainable artificial intelligence (XAI) framework combining graph neural networks (GNNs), attention mechanisms, and reinforcement learning to jointly optimize physical routing while providing interpretability. We conducted our empirical study by evaluating the framework on 47 bench-mark circuits used in both industrial and academic settings, ranging from 10K to 500K transistors. The framework can accom-plish a routing completion rate improvement of 18.7% over the state-of-the-art, with an average reduction of 23.4% in the time utilized for routing. The explanation module integrated into the design of LIME helps designers obtain actionable insights on routing decisions with 94.2% fidelity to used optimization objectives. Our extensive experiments show that this hybrid ap-proach gives a good trade-off between solution quality and computational efficiency whilst making transparency essential for adoption in the industry. The framework demonstrated the ability to perform well despite changes in circuit complexity, differ-ences in design rules, and routing constraints. Statistical analysis shows the approach is beneficial everywhere, from congestion metrics to wire length optimization to design rule compliance. This work connects black-box AI tools to industries that demand explainable and trustworthy optimization tools in VLSI design automation.
