Surrogate Model-Based Filter Optimization by a Field-Circuit Model Mapping

In this article, an artificial neural network (ANN) model development technique is described for efficient and fast high-frequency structure design and optimization. Unlike the well-documented space mapping (SM) and aggressive SM technology, we map equivalent circuit (EC) model parameters to field model geometric parameters through neural modeling. First, a complete electromagnetic (EM) structure is segmented into a series of different discontinuities. Then, the EC model corresponding to each discontinuity is derived from a set of calibrated circuit parameters. Next, couplings of different orders between the discontinuities are represented as a part of ECs. Finally, the complete EC model of a full-wave EM structure is developed. All the circuit parameters are extracted against different combinations of critical geometric parameters of the target EM structure. This dataset is used to develop the ANN model for mapping the EC model parameters to EM model geometric parameters. At this stage, the circuit model can be used for optimization purposes. The optimized circuit parameters are then mapped back to the geometric parameters in connection with the predesignated performance. In this work, a dual-band resonant-aperture (RA) rectangular waveguide filter and a third-order nonradiative dielectric (NRD) waveguide filter are shown as examples to demonstrate the proposed methodology. Both examples show a good agreement between simulation and measurement results.