A Deep-Learning Based BSIM-CMG FinFET Modeling and Parameter Extraction Approach

As a basic module for integrated circuit simulation, whether the compact model is accurate and efficient greatly affects the circuit simulation results. Parameter extraction is a key step in compact modeling. Consequently, we introduce a new deep learning (DL)-based parameter extraction method in this work, which generates 50,000 training data by Monte Carlo simulation of the parameters in the Berkeley short-channel IGFET model (BSIM)-common multigate (CMG). A multilayer perceptron model is constructed for training, where input are C
gg
- V
g
and I
d
-V
g
, output are BSIM-CMG parameters. Fin Field-Effect Transistor (FinFET) is simulated using TCAD with the same geometry to validate the DL model. The parameter extraction results based on the DL model are in perfect match with the TCAD data, and the Root Mean Square Error (RMSE) values are only 8.28% and 1.49%. In addition, the trained model performs well for the prediction of devices with a 10% variation in gate length and effective oxide thickness, demonstrating the great promise of deep learning in the field of parameter extraction.