Machine Learning Approach for Predicting the Effect of Statistical Variability in Si Junctionless Nanowire Transistors

Abstract

This letter investigates the possibility to replace numerical TCAD device simulations with a multi-layer neural network (NN). We explore if it is possible to train the NN with the required accuracy in order to predict device characteristics of thousands of transistors without executing TCAD simulations. Here, in order to answer this question, we present a hierarchical multi-scale simulation study of a silicon junctionless nanowire field-effect transistor (JL-NWT) with a gate length of 150 nm and diameter of an Si channel of 8 nm. All device simulations are based on the drift-diffusion (DD) formalism with activated density gradient (DG) quantum corrections. For the purpose of this letter, we perform statistical numerical experiments of a set of 1380 automictically different JL-NWTs. Each device has a unique random distribution of discrete dopants (RDD) within the silicon body. From those statistical simulations, we extract important figures of merit (FoM) such as OFF-current (IOFF) and ON-current (ION), subthreshold slope (SS), and voltage threshold (VTH). Based on those statistical simulations, we train a multi-layer NN and we compare the obtained results with a general linear model (GLM). This shows the potential of using NN in the field of device modeling and simulation with a potential application to significantly reduce the computational cost.