Global Routing Under a Congestion-Aware Reinforcement Learning Model

As one of the challenging problems in VLSI physical design, global routing is facing increasing difficulties and more and more algorithms attempt to introduce machine learning-based solutions. While most of these solutions lack high enough routability and routing efficiency. In this paper, we propose a congestion-aware reinforcement learning model for global routing. The global routing problem is solved based on pattern routing and a reinforcement learning technique is introduced for better searching global optimal routing decisions. We convert the network relation into a graph structure and introduce a graph attention network to provide global congestion features. Moreover, a net segment-based feature extraction mode for pattern routing is proposed to enhance the routing efficiency. Experimental results on ISPD18 benchmark show that the proposed algorithm can route different unseen layouts with an average routability over 99.9% and the average number of the failed nets is around 6.