Prediction of Characteristics Using a Convolutional Neural Network Based on Experimental Data on the Structure and Composition of Metamaterials

This work proposes an algorithm for properties predicting metamaterials depending on their structure, physical properties of the components of metamaterials, and their characteristics. In this context, the term ”properties” means the result of interacting with the irradiation of a material with electromagnetic exposure of a certain frequency or spectral composition to determine the transmittance/reflection coefficients of the metamaterial. The model is based on the construction of metamaterial in form of a 3D object, the presentation of physical properties in the form of additional components in the object’s vectors, the presentation of experimental data in the form of polynomial coefficients, or the points on the chart of dependencies. Despite the small amount of data, a sufficiently small error rate was obtained for both cases, and the prediction results of experimental data are presented. The amount of experimental data can be increased by supplementary parameters which characterize the conditions under which the experimental data were obtained - polarization, angle of incidence, the intensity of irradiation, etc. The main issues may arise during the preparation of data for neural network learning due to difficulties in converting 3D formats into the required array of data and taking into account all the circumstances, dielectric and magnetic permeabilities, and specific conductivity.