Optimization of Radio-frequency Properties of Graphene FET enabled by Deep Learning

Abstract

The possibility of a novel design of the gate to increase radio-frequency (RF) properties of graphene filed-effect transistors (GFETs) that use graphene pseudo-optics by using a deep learning (DL) was explored. The transmission probabilities for arbitrary gate shapes were calculated by using the finite-difference-time-domain method for massless Dirac fermions and used to train a deep learning. The trained DL predicts the trajectories of massless Dirac fermions according to graphene pseudo-optics. Furthermore, the trained DL can design an optimized gate shape for a targeted graphene pseudo-optic response that increases RF properties of GFETs. The DL-designed GFET has the cutoff frequency f_T=46 GHz and the maximum oscillation frequency f_max=49 GHz, and the conventional GFET has f_T=33.7 GHz and f_max=7.8 GHz with gate length L_g=300 nm. The results showed that the trained DL can increase RF properties of GFETs and solve the existing problem that GFETs have much lower f_max than f_T.