Accurate Prediction of Design Rule Violations Targeting Different Routers By Transfer Learning

Recently, it has been attracting increasing attention for identifying and avoiding design rule violations (DRVs) during early design stages. Machine learning-based methods have been proposed for the DRV prediction before either the global or detailed routing stage. However, there are still challenges for predicting the number of DRVs (#DRVs) purely based on the placement results. On the one hand, the state-of-the-art machine learning methods do not scale well for different design sizes. On the other hand, the training data is typically obtained using a single routing flow, which makes the trained model infeasible for predicting the behavior of other routers. When the global or detailed router changes, it is necessary to regenerate the training data and retrain the model. To address the above issues, this paper proposes a new deep learning method, called NeuroDRV, which accurately predicts #DRVs with significantly enhanced scalability for different design sizes. Moreover, transfer learning is adopted in NeuroDRV such that the trained model on one routing flow could be easily transferred to new routers with notably reduced training data. Experimental results show that compared with the state-of-the-art method, NeuroDRV obtains an average reduction of 72.2% in prediction errors for #DRVs. And the transferred model using only 25% of the training data achieves acceptable prediction results.