Zoziuk M. Application of convolutional neural network for predicting the coefficient of passage of metamaterials depending on the structure and physical composition of metamaterials

Zozyuk Maksym Application of convolutional neural network for predicting the coefficient of passage of metamaterials depending on the structure and physical composition of metamaterials. – Qualifying scientific work on manuscript rights. Dissertation for obtaining the scientific degree of Doctor of Philosophy in specialty 153 – Micro- and nanosystem engineering (field of knowledge 15 – Automation and instrument engineering). - National Technical University of Ukraine, Kyiv, 2022. The work is devoted to the development of a technique for predicting the penetration coefficient of metamaterials based on topological structure and physical composition with the use of machine learning techniques, namely artificial neural networks using convolution and auxiliary operations. Scientific and applied research, highlighted in the dissertation, is focused on the practical study of artificial convolutional neural networks for performing forecasting tasks in the fields of design, prediction of metamaterials with the required properties. The choice of these machine learning techniques is due to the inefficiency of other existing methods for the tasks of predicting properties and designing metamaterials. The possibility of predicting the transmission coefficient based on information about the structure, physical composition, and measurement conditions of metamaterials is analyzed. It is shown that such a possibility exists, and the main conditions imposed on data about metamaterials, which are required for the use of convolutional neural networks in solving problems of predicting information about metamaterials, are described. An algorithm for predicting the transmission coefficient based on the structure, physical composition of metamaterials based on a convolutional neural network using experimental data of laboratory metamaterials has been developed. It was found that data saved in ".ply" format is not suitable as input data for a neural network. The Python-based Open3D software package was used to convert the original format to the ".xyzrgb" format, which is an array where each line is a vector of six numbers; the first three are coordinates, the other three are RGB numbers. The possibility of using this type of information in convolution processes and for neural network training is shown. Conditions imposed on the data are presented: data homogeneity, which means that all parameters that affect the prediction result must either be the same or be included in the data as property information; scalability, which means that the data must be reduced to the same ranges of values and scales. An algorithm for saving information about the physical composition of metamaterials has been developed. It is shown that using information about the electromagnetic properties of chemical elements, it is possible to predict the transmission coefficient of metamaterials. 3D objects of metamaterials are built in a digital environment with information about a chemical element attached to each pixel. The algorithm for saving this information in a form convenient for training a convolutional neural network is described. Types of digital formats that can be used to save the necessary information about metamaterials are presented. The format used to store information about the 3D object in this work was analyzed. Other formats for saving information about 3D structures were analyzed. It has been studied that such formats as ".fbx", ".obj", ".stl", ".3ds" are not suitable due to the following reasons: lack of effective software functionality (to convert to the format required for the neural network); lack of saving information about additional channels of information; the presence of redundant information that cannot be separated from the necessary information. A convolutional neural network architecture has been developed for predicting the transmission coefficient based on the structure and physical composition of metamaterials. Prediction results are shown for both cases from the representation of the pass coefficient - in the form of an array of points and polynomial coefficients. The prediction results are analyzed and techniques are given to improve these results and optimize the network to reduce training execution time and save resources. Data augmentation is shown to be the most effective method for improving forecasting results. Nevertheless, performance improvement methods based on architecture changes and hyperparameter changes should be continually evaluated and used whenever possible. Keywords: convolutional neural network, metamaterial design, 3D model, pass rate, dielectric materials, dielectric constant, measurement of material parameters, approximation model, experimental-analytical method, information visualization, non-linear model, intelligent computing, computer programming, dielectrics, numerical modeling, machine learning, mathematical modeling, prediction.