Enhancing Radio Frequency Performance of Graphene Field-Effect Transistors through Machine-Learning-Based Physical Prediction and Optimization

Abstract

Graphene has excellent carrier mobility, making it highly suitable for high-speed electronics. However, controlling the carrier flow still poses significant challenges due to the lack of an energy band gap. To address this issue, researchers have explored introducing a transport gap using the pseudo-optic scheme of the massless Dirac Fermion, which has shown promising results in improving device characteristics. In this study, we demonstrate the use of a deep neural network to manage the pseudo-optic trajectories of carriers and discover an optimized geometry. The neural network operates similarly to a pinball game, where multiple intended reflections control the trajectories of steel balls. The flow of massless Dirac Fermions can be modified by well-designed reflection paths, and artificial intelligence can replace human intuition in finding an optimal design. We implement this approach using a finite-difference time-domain simulation for massless Dirac Fermions, forwarding the carriers' transmission for arbitrary gate shapes to the neural network for training. The trained neural network is then used to design an optimized gate shape for a targeted graphene field-effect transistor performance, specifically radio frequency characteristics. Using this approach, we achieved a maximum oscillation frequency (f(max) ) of up to 59.6 GHz for the machine-designed transistor, approximately 20 times higher than that of conventional graphene field-effect transistors with the same effective gate length.