Krysenko P. I. Data enrichment for predicting the properties 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 “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv, 2024.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 “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv, 2024. The work is devoted to the development of data enrichment methods to improve the prediction of the transmission coefficient of metamaterials based on the topological structure and composition, the parameters of metamaterial surface irradiation, as well as to improve the possibilities of solving the inverse problem using artificial neural networks. The dissertation researched the scientific and applied aspects of the use of convolutional neural networks for solving problems in the field of designing and predicting the properties of metamaterials and the possibility of setting the parameters of metamaterial surface irradiation to increase the amount of available data. In modern research in the field of science and technology, deep learning, in particular, the application of convolutional neural networks neural networks, turns out to be an effective means of developing effective methods of analyzing and predicting the properties of metamaterials in the context of their design. The use of neural networks in the design of metamaterials involves their application to find materials and the topology of metamaterials according to given characteristics. This process begins with training neural networks based on a large amount of input data, which includes metamaterial parameters such as size, shape, and composition, and desired properties such as absorption, reflectance, and transmission of light. Thanks to neural networks, it is possible to automate and optimize the design process of metamaterials, to search for complex relationships between material parameters and functional response, which is difficult to achieve with traditional methods. It is also important to note the possibility of analyzing data on the composition and structure of metamaterials and using them to predict electromagnetic properties. Data on the topology and composition of these materials were used to obtain information about metamaterials. Software environments were used to develop digital code and build 3D objects of metamaterials with defined properties. The conversion of data from ".ply" format to ".xyzrgb" format was carried out using the Python-based Open3D software package. An algorithm has been developed for predicting the properties of metamaterials based on their structure and composition and the conditions of experimental studies, using a convolutional neural network, and methods of encoding additional information into an artificial neural network. A method was also developed for efficient storage of information about the composition of metamaterials, which allows predicting their electromagnetic properties. Methods of presenting the properties of metamaterials in a form convenient for convolutional neural network training have also been investigated. A comparative analysis of the effectiveness of different methods was carried out, indicating that the representation of characteristics in the form of polynomial coefficients, although faster, is not the best for predicting the characteristics of metamaterials. Formats for saving information about 3D structures have been analyzed. In this work, it was investigated that such formats as ".3ds", ".obj", ".fbx", ".stl" are not suitable due to the following disadvantages: lack of fast and simple software (for efficient conversion of metamaterial structures into format for training an artificial neural network); impossibility of storing additional information in channels about the structural composition of metamaterials; the impossibility of removing proprietary information that is coded and mixed with data necessary for training. A three-dimensional convolutional neural network learning pipeline has been developed for predicting the frequency electromagnetic characteristics of metamaterials based on the structure and composition of metamaterials and the possibility of establishing information about the conditions of experimental research. Prediction results are shown for different data samples divided by the scale of exposure frequency, taking into account information about the conditions of experimental studies. Ways to improve prediction quality were suggested. Ways of solving the inverse problem for generating the structure according to the given parameters were also proposed
