Study Of The Application Of Neural Networks To Optimization Design Of Defected Ground Structures

Defected ground structures (DGS) are certain structures with artificially etched units that alter current distribution on the metallic plane of microwave circuit, and the transmission lines with DGS could behave characteristic of prohibition in special frequency range. Microwave circuits like DGS are generally analyzed by finite-difference time-domain (FDTD) method because of its great precision. However, the time-consuming process is difficult to satisfy the requirement of analysis and design of DGS.Neural networks are information-processing systems simulating human brain, and they have the ability to learn the input-to-output relationship after trained with data sets. Once trained, the neural networks can make accurate response to the input within the learning range immediately, which bring the extensive application of neural networks in a great many of areas.As for the nonlinear relationship between the dimension of etched DGS unit and the corresponding transmission coefficient, the neural networks can learn and imitate it properly. So the neural network models of DGS have been developed and embedded in the procedure to realize the optimization design.The optimization design of combinatorial nonperiodic defected ground structures (CNPDGS) featuring dual-prohibition band is accomplished in this paper applying neural networks. Optimal defected dimensions that minimize the error function are obtained by simplex algorithm according to the desired parameters of transmission coefficient. In order to improve the generalization ability of neural networks in optimization design, grouped neural networks have also been applied here. The training data are divided into groups in two ways by the similarity in certain range and each sub-network receives training independently. The final output would be obtained synthesizing those groups. Great accuracy of the group-built neural networks has been validated by the comparison with both FDTD and experimentation results, and the capacity of generalization has been greatly improved. In addition, the training time required has been shortened and training data have been made the best use of. Finally, optimization design based on the group-built neural networks has been carried out and the comparisons with simulation and experimentation demonstrate the correctness as well as effectiveness of this developed approach.