Deep Reinforcement Learning-Based Ground-Via Placement Optimization for EMI Mitigation

Abstract

Placing ground vias plays a crucial role in mitigating electromagnetic radiation from printed circuit board (PCB) edges. Stitching dense ground vias along PCB edges is not always feasible due to limited layout areas. Finding a ground-via placement strategy to achieve the best electromagnetic interference (EMI) mitigation using a specified number of ground vias is desired in the industry. However, this process is usually tedious and labor-intensive because of the enormous search space. This article proposes an optimization algorithm based on deep reinforcement learning to adaptively seek the optimal ground-via placement strategy in complex packages. First, an evaluation module based on convolutional neural networks (CNNs) is trained to predict the EMI mitigation for any possible ground-via placement. Then, relying on the well-trained CNN, an optimization module based on dueling double deep Q network is developed to find the best ground-via placement strategy through exploration and training without prior electromagnetic knowledge. The final optimization strategy obtained by our proposed algorithm has more effective EMI mitigation than the commonly used “edge-wrapping” solution in industrial products. Our proposed algorithm also provides guidelines and insights about the design rule and the behind physics.