A Novel Transfer Learning Approach for Efficient RF Device Behavior Model Extraction

In this work, a novel approach that combines artificial neural networks (ANNs) and polynomial modeling to efficiently model the behavior of nonlinear RF devices is introduced. Utilizing the strengths of ANNs in knowledge retention and transfer, the proposed approach effectively incorporates new operational states such as different dc biases, frequencies, dies, and devices into existing models. This adaptive capability, in addition to the dynamic sampling techniques that only capture crucial dynamic characteristics across various operational states, significantly enhances the behavioral model's extrapolation ability across various operational states and decreases the amount of measurement data needed. This approach is verified with a 0.25- mu m real gallium nitride (GaN) high electron mobility transistor (HEMT) device measured at different dies, sizes, dc biases, and frequencies. The results show that by utilizing only 30% of the conventional measurement data, the established polynomial equation-based model achieved a normalized mean squared error (NMSE) below - 30 dB. The proposed model technique also has the advantage of easily being implemented in computer-aided design (CAD) software and allows for the fast design of RF circuits in industry applications while ensuring high accuracy and adaptability.