Auto-encoder based hybrid machine learning model for microwave scaled GaAs pHEMT devices

In this article, a study of performing machine learning (ML) based modeling for semiconductor devices has been developed using experimental microwave data. Characterization of gallium arsenide (GaAs) pseudomorphic high electron mobility transistors (pHEMTs) with different gate widths is used as the illustrative example to demonstrate the accuracy and effectiveness of the presented modeling procedure. The tested devices are based on the multifinger layout, in which the total gate width (W) is obtained by multiplying the number of fingers (N-f) and their length (W-0). Machines are trained with scattering (S-)parameter measurements up to 65 GHz by using the extreme gradient boosting (XGBoost) algorithm with K-fold cross-validation. Then, the output of the trained machine is utilized by the parameters such as N-f and W-0 inside the Auto-encoder (AE) model. In particular, the ML model with AE has a maximum of 99.88% prediction accuracy, despite the uncertainty inherent in the microwave measurements and the unavoidable deviations from the ideal behavior of the analyzed devices.