A Tandem Neural Network-Based Inverse Modeling Method for Fast and Accurate Estimation of Electromagnetic Model Parameters

Abstract

In this paper, a neural network method is proposed for better estimation of simulation material parameters. Such improved estimations increase the accuracy of EM simulations. Estimating simulation material parameters from an electrical response is an ill-posed problem since multiple parameter combinations can exhibit almost the same electrical response. Such problems can be more effectively addressed by specific neural network topologies, such as tandem neural networks. Hence, a solution following this approach is proposed. As an example, the substrate model parameters of the cascaded T-resonators and the Beatty resonator are estimated using their scattering parameters. The proposed approach supports a wide range of microwave structures through a structure-independent training data generation process. As an alternative to the proposed tandem neural network that combines convolutional, transposed convolutional and fully connected neural networks, a second tandem neural network employing only fully connected neural networks was also realized as a reference model and the performance of the two is compared. The performance of the two is evaluated using synthetic test data obtained from EM simulations. The mean squared error of the proposed model is approximately $0.5 \times 10^{-3}$, which is significantly lower than that of the reference model. The proposed method can be applied to measurement-based modeling, de-embedding fixture effects at high frequencies, production lot monitoring, among others.