Predicting the effectiveness of meta-screen screening using a neural network approach

This paper proposes an approach to predicting the effectiveness of meta-screen screening based on an artificial neural network. The meta-screen is considered as a periodic structure of resonator copper plates on a dielectric substrate. An artificial neural network was trained for the task of predicting the effectiveness of meta-screen screening, based on computer modeling data as reference values, which demonstrated the potential applicability of the method in real conditions. Based on numerical simulation, a sample is formed that includes the values of the shielding efficiency in dB at the target frequency for various combinations of parameters. To approximate the nonlinear mapping of vector X to the value of screening efficiency, a multilayer fully connected neural network is used, trained in a supervised mode using the metric of the average absolute percentage error. The influence of hyperparameters and network structure on the convergence and accuracy of the model is analyzed. It is shown that with the optimal choice of architecture, the test prediction error does not exceed 7.01%, and the calculated values of the shielding efficiency are in good agreement with the results of electrodynamic modeling. The results obtained demonstrate the possibility of using the developed model in the tasks of accelerated parametric synthesis of meta-screens and automation of procedures for predicting screening efficiency.