Neural network model for ballistic carbon nanotube transistors

Abstract

In this paper we present a Neural Network (NN) model for the ballistic carbon nanotube transistors. In comparison with the state of the art theoretical reference CNT model implemented in FETToy, our proposed model is a SPICE-compatible model and has a faster speed while maintaining the accuracy within less than 2% in terms of RMS error. The results show that, NN model has smaller RMS errors in calculated current under various conditions such as the oxide thickness, the nanotube diameter, gate-source voltage, the oxide permittivity and the source Fermi level, than the existing analytical models published by others.

A mathematical space mapping model for ballistic carbon nanotube field-effect transistors

“…[9] to solve this problem. In this model, three neural networks have been applied for predicting the parameters V A , k, and V TH of a well-known ''level 1'' MOSFET model in SPICE such that its output would be equal to the one of FETToy.…”

“…The complete model is constructed by inserting these parameters into the previous model, which is [9], Fig. 2 are a comparison between the proposed and FETToy models made for various CNTFETs with different values of carbon nanotube diameters.…”

“…The CNT diameters were selected in the range of 0.5 nm B d B 1.5 nm, a range for practically used CNTs. The values of the parameters used in this comparison are the same as those considered in [9], i.e., T = 300 K, C Tot = 172 aF-lm -1, a G = 0.93, a D = 0.035, e r (-SiO 2 ) = 3.9, t ox = 1.5 nm, and E F = -0.32 eV. The figure was obtained under constant V GS = 0.5 V. As seen from the figure, the dependency of I DS to CNT diameter can be appropriately modeled by the proposed model.…”

“…In a number of modified models, analytical mathematical model [1,8], models based on neural networks (NNs) [10] or models based on neural space mapping (NSM) [9,12] have been proposed for the direct calculation of output current. In [11], a circuit model was presented that could implement the FETToy model in the device simulators if an analytical solution is available for its charge integral equations.…”

“…The model presented in Ref. [9] is one of the models used this concept. In this model, a well-known ''level 1'' MOSFET model in SPICE [13] is modified by three neural networks such that its output would be equal to the one from FETToy.…”