An Enhanced Automated Neuro-Model Generation Algorithm for Parametric Modeling of Microwave Components

This paper presents an enhanced automated model generation (AMG) algorithm using neural networks (NNs). AMG performs automated sampling of training/testing data and automated NN structure adaptation to obtain a compact NN model of user-required accuracy with suitable amount of data. The proposed enhanced AMG performs data sampling in a stage-wise manner. In each stage, instead of sampling along all dimensions of the input space, the proposed AMG distinguishes the input dimension which influences the nonlinearity of the model behavior most in current stage, then generates additional training samples along this input dimension. Compared to existing AMG, the proposed algorithm can reduce the amount of data needed for NN modeling, especially in the case of multidimensional parametric modeling of microwave devices. Examples including automated modeling of MOSFET and bandpass filter are presented to demonstrate the validity of the proposed AMG.