ICDaIR: Distribution-aware Static IR Drop Prediction Flow Based on Image Classification

During the integrated circuit design process, the maximum IR drop value is often given more attention. The frequency of the maximum IR drop in the actual circuits presents an uneven dispersion, i.e., long-tail distribution. To address this problem, this paper introduces ICDaIR, a distribution-aware static IR drop prediction flow based on image classification. ICDaIR first utilizes a random forest classifier to categorize sub-regions, obtained from segmentation, into three groups. There are small differences in sub-regional samples within each category. Then, a U-Net-based prediction model is used to obtain the IR drop map for each class of sub-region. The overall IR drop map is derived by combining the values from these sub-regions. On the dataset with a maximum layout area of 0.69μm2, compared with SOTA and the model without classification, the MAE of our model decreases by 50.5% and 14.4% respectively, with the prediction error controlled within 0.2029mV.