ANN model for compact modeling of Thin-TFET

(Depreciated, see Pi-NN)

1. Model Overview

Feed forward artificial neuron networks (ANN) are used to model emerging electronic devices. This ANN modeling paradigm uses a special designed loss function to achieve much better accuracy in the sub-threshold region than standard quadratic loss function. Besides this, these ANN models are easy to use and guarantee to be infinitely differentiable everywhere.

We have trained two different types of model: one using logistic sigmoid functions as activation functions, and the other using hyperbolic tangent functions as activation functions. Brief descriptions are followed:

a) Model with logistic sigmoid function (sig model)

This model has only one hidden layer, which has 5 to 7 neurons in different versions. The first and third quadrants are modeled independently, then add them together.

• Advantage: excellent accuracy in the deep sub-threshold region (i.e. Vtg < 50 mV)
• Disadvantage: nonlinear turn-on around Vds = 0
• When to use: when deep sub-threshold region is important. Since sig model is light-weighted compared to tan model, if you can live with the nonlinearity around Vds = 0, sig model is your friend.

The sig model comes with different versions for both first and third quadrants:

• For the 1st quadrant, there are one versions:
  1. sig_1q_7: which has 7 neurons in the hidden layer
• For the 3nd quadrant, there are two versions:
  1. sig_3q_5: which has 5 neurons in the hidden layer
  2. sig_3q_7: which has 7 neurons in the hidden layer

Recommend to use sig_1q_7 and sig_3q_5.

b) Model with hyperbolic tangent functions (tan model)

This model uses hyperbolic tangent functions. Unlike sig model, tan model use a single ANN to model both first and third quadrants.

• Advantage: compared to sig model, this model has no unintentional nonlinear turn-on around Vds = 0
• Disadvantage: current densities oscillate around 0 in the deep sub-threshold (Vgs < 50 mV)
• When to use: when the linearity of Id-Vds around Vds = 0 is important

The tan model comes with three different versions:

1. tan_2h: which has 2 hidden layers, each has 9 neurons
2. tan_3h: which has 3 hidden layers, each has 9 neurons
3. tan_4h: which has 4 hidden layers, each has 9 neurons

Recommend to start with tan_2h, the simplest version yet almost as accurate as its bigger brothers (i.e. tan_3h and tan_4h)

2. Use the model (in Python)

The interfaces to use this model are in neuralFET.py. Here is a short tutorial of how to use both sig model and tan model. The main goal we would like to achieve is: for given Vgs and Vds, get current density Id.

a) Use sig model

First, read the model file, and create a function accordingly. For example, if we would like to use sig_1q_7 for the first quadrant, and sig_3q_5 for the thrid quadrant, then:

sig_ann = make_sig('sig_1q_7/', 'sig_3q_5/', 2)	

The last "2" indicate there is n-1 hidden layer(s). Now you can get current density from Vgs and Vds by:

id = sig_ann(Vgs, Vds) 
# Vgs and Vds have unit of V, id has unit of A/m

b) Use tan model

Just like sig model, if we would like to use tan_2h version:

tan_ann = make_tan('tan_2h/', 3)	

Remember, when you switch to tan_3h or tan_4h, you need to change the second argument to 4 or 5 respectively. To get current density, simply use:

id = tan_ann(Vgs, Vds)
# Vgs and Vds have unit of V, id has unit of A/m 	

---

There is a plot function can be very handy to plot out the whole model at once, the signature of this function is in ffnn.py:

data = plot(current, x_list, c_list, TYPE = "t", LOG = False)

Current is the model function, such as sig_ann or tan_ann; For family plot, x_list is a list of Vds and c_list is a list of Vgs; For transfer plot, x_list is a list of Vgs and c_list is a list of Vds; TYPE = "t" for transfer plot, TYPE = "f" for family plot; LOG = False means linear plot, LOG = True means semi-log plot. plot function returns all the plotted data points in a 2D list.

To save this 2D list into csv file, there is also a saveData function to help do so. This function is in ffnn.py.

Enjoy using this model!