Prediction of electrical characteristics using spectrum and metrological data based on deep learning model

The prediction of electrical characteristics (ET) is very important to prevent wafer loss by taking preemptive action against performance drop. The monitoring of threshold voltage of defective device, one of the various reliability values, is important since threshold voltage of defective device and gate-source/drain contact distance are highly correlated, and it is directly relative to gate-source/drain short defect. The existing simple linear regression modeling is less predictive to the ET values since the relationships between parameters related to ET have become more diverse and complex due to the finer and complicated structure. In this respect, deep learning is used to find latent features in the spectral data and clarify the correlation. We propose an improved predictive model using spectrum and Metrological data based deep learning modeling. The deep learning modeling using spectrum and Metrological data has been applied to directly predict the threshold voltage of defective device and it results in higher consistent results, correlation R-square value 0.43 from dataset consisting of spectrum data (Muller matrix components) of contact layer of source drain. The proposed deep learning based prediction model using spectrum and Metrological data will contribute to prevent wafer loss and to achieve the minimal yield loss in the manufacturing process by improving the ET prediction capability.