Neural-network-based prediction of cryogenic BSIM4 model parameters from small datasets

Abstract

Neural-network-based prediction of BSIM4 model parameters for current-voltage (I-V) characteristics of short-channel bulk MOSFETs at 4 K was examined. Combining two neural network (NN) models and the least-squares method enabled the extraction of model parameters from only 4 experimentally obtained Id-Vg characteristics of 2 cryogenic MOSFETs, contributing to reducing time and financial costs for cryogenic device modeling. The proposed protocol provided lower root-mean-squared errors than the case where the least-squares method was solely employed. The superiority of the proposed protocol became evident as the number of model parameters increased, attributed to NNs providing good initial guesses for the least-squares method. As a result, cryogenic MOSFETs’ characteristics were fit by BSIM4 with the 6.6% error. This study also emphasizes the superiority of generative NN models over regression NN models for model parameter extraction by demonstrating the parameter extraction for highly correlated model parameters relating to threshold voltage.