Design Rule Violation Prediction with Clustering and Pin Features using Convolutional Neural Network

Abstract

Deep learning-based Design Rule Violation (DRV) prediction is increasingly applied in advanced integrated circuit design. In this paper, we present an optimized Convolutional Neural Network (CNN) model that incorporates two novel input features for improved prediction. Clustering degree feature and complex pin feature significantly enhance the model ability to extract layout information. Additionally, the optimized CNN model “RouteNet-AMK” improves processing capabilities for complex layout data by incorporating the attention mechanism and multi-scale feature fusion technique. Experimental results show that the F1 score of the proposed method on the CircuitNet N28 dataset is 3.51% higher than that of the RouteNet model using conventional features.