Fast and Simple Pre-Design of Decoupling Capacitors Using Reinforcement Learning

In this paper, we propose a reinforcement learning (RL)-based methodology for the pre-design of decoupling capacitors (decaps) in power distribution networks (PDNs). The method aims to optimize the quantity and size of decaps across the PCB, package (PKG), and chip to meet target specifications of ripple voltage and power supply-induced jitter (PSIJ). By utilizing RL, the approach rapidly estimates the decap configuration without needing retraining for different RLC profiles, enhancing design efficiency. Compared to traditional genetic algorithms (GA), the proposed RL-based method significantly improves both performance and computational efficiency, achieving optimal results with over 10 times faster inference time. This approach is particularly useful in the early stages of design, where estimating the required decap quantity is more critical than determining precise placement. The results validate the effectiveness of RL in optimizing decap design for diverse product scenarios.