Early Stage DRC Hotspot Prediction for Mixed-Size Designs Through an Efficient Graph-Based Deep Learning

Predicting hotspot locations in the early stage of Design Rule Check (DRC) is crucial for designers to proactively prevent design rule violations. However, obtaining an accurate and efficient predictor faces significant challenges due to the influence of available information and severe data imbalance. In this study, we investigate the potential of utilizing Graph Neural networks (GNN) to address this challenge. Our focus is specifically on accurately predicting DRC hotspot locations without relying on global routing techniques. We consider the presence of macros in mixed-size designs. We propose an adaptive adjacency matrix that demonstrates superior application effectiveness compared with traditional adjacency matrices. Furthermore, experimental results on benchmark circuits show significant improvements in the true positive rate (22.38% for the RouteNet model and 26.90% for the GNN model) and accuracy (6.97% and 6.76%, respectively) compared with these models. Our proposed model also maintains a low false positive rate and outperforms other Convolutional Neural Network and GNN models. Additionally, its efficient learning capability and lower computational time contribute to its outstanding training performance, with training time being approximately 10% of that required by other models.