Neural network (NN) modeling for semiconductor devices has been proposed and under research for a few years, yet most proposed models use a fully-connected neural network without a good physical explanation. Simulated data rather than experimental ones are often used as a resource. We present here a physical-information based neural network using experimental data [1] to model Carbon Nanotube Field Effect Transistor behavior with varying channel length (Lch), contact length (Lc), drain source voltage (Vds) and gate-source voltage (Vgs). We’ve also created a data cleaning method for experimental data with large hysteresis. The model was able to fit training data well and also provides good estimation to test conditions.

We build up our NN model using a combination of two parts: one part using Vgs and Lch to model charge carrier density, and another using all parameters to model the velocity of charge carriers. Since current is based on both charge carrier density and saturation velocity, we use the the result of these two models to predict current. We model charge carrier density using log(Ids) as input data with respect to Lch and Lc. For the charge carrrier velocity model, we created a layered structure to group electric field distribution with relevant inputs of Vds, Vgs and Lch into each layer, then add the effect of Lc as another input that controls charge carrier collection. We seperate these two processes in the model because charge carrier density is exponentially affected by Vgs, while saturation velocity is much less varying. Using only Ids tends to an inaccurate model. After training, this two-part model is able to provide good fit of Ids and log(Ids) at the same time. We have also tested the model on un-seen conditions and the model provides reasonable estimation.