Efficient behavioral model extraction of nonlinear active devices using adaptive sampling with compact nonlinearity measure

Description of nonlinear active devices is very complex, and depends on many input variables. Therefore, extraction of behavioral models based on traditional Designs of Experiments, such as factorial or Latin hypercube, may be unacceptably expensive in terms of sample evaluation time. In order to limit the total number of samples required to obtain accurate behavioral models, an adaptive sampling strategy may be used. It is based on surrogate models that are extracted for each sampling iteration. As nonlinear description consists also of many output variables, a common synthetic quantity is proposed to limit the surrogate modeling cost. It is defined as a total change of all the output quantities. The approach was evaluated in measurements of a 0.15 μm pHEMT model. The modeling accuracy is improved, while significant modeling-cost reduction can be observed.