Reinforcement Learning-Assisted Optimization of Power Plane and Placement of Decoupling Capacitors in Power Delivery Networks

This paper introduces a novel reinforcement learning (RL) optimization framework for designing power delivery networks (PDNs) on printed circuit boards (PCBs). The proposed framework aims to simultaneously minimize the number of decoupling capacitors (decaps) and optimize the power plane area while satisfying predefined target impedance (TI) requirements. To address complex design situations, such as irregularly shaped power planes, convolutional neural networks (CNNs) have been integrated into the RL framework, enabling efficient handling of power plane geometries. Experimental results demonstrate that this integration, coupled with a carefully structured action space, significantly accelerates the training process. Notably, the implementation of multiple parallel sub-actor networks within a hierarchical action space allows simultaneous optimization of various design parameters. This enhanced approach not only supports more realistic and complex design cases but also achieves a fourfold reduction in training time.