The problem of prediction of the transmission coefficient using neural networks with a limited quantity of data

• Date: 2025
• Main Author: Krysenko, Pavlo
• Format: Article
• Language: Ukrainian
• Published: Kyiv National University of Construction and Architecture 2025
• Subjects: deep learning; low pass filter; convolutional neural networks; metamaterials
• Online Access: https://es-journal.in.ua/article/view/328625
• Journal Title: Environmental safety and natural resources

The article discusses available approaches to predicting the transmission coefficient of metamaterials. In the paper was proposed different approaches that create the possibility of using data from various open sources, as well as the possibility of encoding complete structural information about the composition of metamaterials. A neural network with two inputs was designed, which is based on a three-dimensional convolution operation. Using these approaches, the training of an artificial neural network was carried out, and the results of transmission coefficient prediction were presented. The nature of metamaterial use can be determined by the predicted coefficient, but the resulting root mean square error still does not allow using such a neural network as a substitute for existing approaches. The paper presents an analysis of the obtained results, in which possible approaches to solving the problem of the amount of data are proposed, as well as solving the problem of different intervals of electromagnetic radiation in the dataset using the architecture of a three-dimensional transformer.