Deep learning-based I-V Global Parameter Extraction for BSIM-CMG
Fredo Chavez, Chien-Ting Tung, Ming-Yen Kao, Chenming Hu, Jen-Hao Chen, Sourabh Khandelwal
Abstract
A new deep-learning-based parameter extraction for a global (multiple gate lengths) BSIM-CMG drain-current model is presented in this paper. The approach starts with generating 300K training dataset, consisting of 778 million data points to train the deep learning engine. Training data is generated by Monte Carlo simulation. The I-V data and the device geometry information from multiple devices serve to train a deep-learning (DL) model to predict BSIM-CMG parameters. The performance of DL-based extraction is verified by using the trained DL model to extract parameters of 10 nm FinFET technology simulated with TCAD. The DL-extracted BSIM-CMG model shows a good accuracy for eight different gate-lengths. The created BSIM-CMG global model was also able to reproduce the scalability in key electrical performance parameters such as off current Ioff, saturation current Isat, linear current Ilin and the threshold voltage in linear Vth,lin and saturation Vth,sat conditions. The developed solution significantly reduces the model extraction time for a global BSIM-CMG model. This new technique can expedite the development of process design kits (PDK).