Abstract
An accurate ensemble learning (EL) based hybrid empirical model for gallium nitride high electron mobility transistors (GaN HEMTs) is presented in this paper. The EL is used to construct the 10-element intrinsic parameter extraction model, linked with device S-parameters using the empirical formulation theory for multi-bias conditions. To obtain a good prediction accuracy and generalization over the complete range of multi-biasing, the EL method intends to achieve the optimized set of all hyperparameters such as method (Boosting or Bagging), minimum leaf size, number of learners, and the learning rate for the development of a model. The trained EL-based model is verified and made scalable in terms of geometry by predicting the intrinsic parameters at 50 % multi-bias testing conditions for GaN HEMTs of different sizes. The proposed approach expedites parameter extraction with pragmatic accuracy. Eventually, the predicted 10-element intrinsic parameters were used to validate the small-signal modeling accuracy on multi-biasing conditions on bias preserved as a test set. A good agreement between the measured and the simulated S-parameters over a frequency range of 1–18 GHz demonstrates the method's efficacy.