22 September 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
In advanced technology nodes, the back-end-of-line (BEOL) process is increasingly critical for determining chip timing and power consumption. This article presents CTRL-B, a transferable reinforcement learning (RL) framework that automates and accelerates this process. CTRL-B integrates two key innovations: 1) a graph attention network (GAT)-based encoder, pretrained on out-of-domain datasets to learn generalizable circuit representations and 2) a cross-domain transfer layer that reuses knowledge from previously optimized source domains. This architecture provides the RL agent with an effective initial policy, which significantly reduces redundant exploration and accelerates convergence to high-quality solutions. We demonstrate the effectiveness of CTRL-B against conventional methods, achieving a 0.4% power, 57.6% reduction in worst negative slack (WNS), and a 41.7% reduction in total negative slack (TNS), alongside 1.58% area reductions. The practical viability of the optimized BEOL configurations is confirmed through stable static IR drop and successful design rule (DR) violation checks. A series of ablation and sensitivity analyses further validates the robustness of our architecture across various designs. These results establish CTRL-B as a robust and scalable solution for automated BEOL optimization, highlighting its potential to shorten modern chip development cycles.