Wide-Range Parametric Modeling of an Inter-Digital Bandpass Filter

This paper introduces a novel decomposition technique for parametric modeling microwave components whose geometric variable values vary over a wide range. This technique includes generating parallel data, training sub-models in parallel, and modifying sub-models in parallel. A second-derivative-based algorithm is used to divide a large range of geometric variables into a series of sub-ranges, leading to the decomposition of a highly nonlinear region into numerous sub-regions, each of which is trained to obtain an artificial neural network (ANN) model with a simple structure. However, the problem of complex multidimensional discontinuities occurs in the process of connecting all the sub-models into an overall model. Therefore, a solution to this problem is proposed by introducing a sub-model modification process, so that this overall model can be used for design optimization. The decomposition technique can enhance the accuracy with short model-development time compared with the methods that train the overall wide range directly into a model. The proposed technique is demonstrated to be valid through an example of an Inter-Digital Bandpass Filter.