Abstract
This paper presents a physics-guided machine learning (PGML) approach to model the I-V characteristics of GaN current aperture vertical field effect transistors (CAVET). By adopting the method of transfer learning and the shortcut structure, a physically guided neural network model is established. The shallow neural network with tanh as the basis function is combined with a hypernetwork that dynamically generates its weight parameters. The influence of transconductance is added to the loss function. This model can synchronously predict the output and transfer characteristics of the device. Under the condition of small samples, the prediction error is controlled within 5%, and the R2 value reaches above 0.99. The proposed PGML approach outperforms conventional approaches, ensuring physically meaningful and robust predictions for device optimization and circuit-level simulations.