| Defected ground structures (DGS) are expanded from the photonics bandgap (PBG) structures. On the microwave circuit's ground metallic plane, the defected units are etched artificially in order to change the ground current distribution and influence the frequency properties of the transmission lines. The structures of DGS are generally calculated using electromagnetic field numerical value analysis methods such as finite-difference time-domain (FDTD) method. These methods can precisely analyze the frequency properties of DGS, but they are computationally expensive. Neural networks, also called artificial neural networks (ANNs), are information processing systems which can be trained to learn any arbitrary nonlinear input–output relationships from corresponding data. They have been widely used in a number of areas such as pattern recognition, control etc. An increasing number of microwave engineers and researchers have started to take serious interest in emerging neural networks owing to speediness and precision. In this paper, ANNs are applied to study of the bandgap characteristic of DGS. The trained ANNs can quickly provide precise answers to the task they have learned, and it greatly overcomes the disadvantage of FDTD in consuming computing time. It is shown that ANNs is powerful approach for the fast analysis and precise design of DGS. The ANNs models of the periodic defected ground structures (PDGS) and a novel combinatorial nonperiodic defected ground structures (CNPDGS) with the characteristic of double stop band have been developed in this paper. The structure sizes of DGS and the frequencies are defined as the input samples of ANNs, the parameters of transmission coefficient calculated by FDTD method are sampled and defined as the output samples. For the PDGS with biggish structure sizes the transmission coefficients have the tiny dither owing to abnormal disturbance and the training samples have been smoothed by the least square method and five-spot triple smoothing method. The training time was saved and the leading trend of the transmission coefficients was reflected well and truly. The ANNs models have been trained with the Bayesian Regularization algorithm in this paper. The transmission coefficient of DGS at any arbitrary sizes and the frequencies can be obtained quickly from the trained ANNs models. The filters with CNPDGS are designed and manufactured in Hebei Semiconductor Research Institute at last, and the measured results are agreement with the results of the final ANNs models. It is shown that ANNs are very precise and effective. The ANNs program is developed using Matlab language, which can greatly simplify the complicated computation in analysis and design, and therefore the program is simple and fast. |
