Решение задач материаловедения методами искусственных нейронных сетей
Аннотация
Основной целью настоящей работы, является поиск и апробация математических и программно-аппаратных методов для установления закономерностей между химическими составами и свойствами выбранных металлических и не металлических материалов для совершенствования служебных характеристик конечной продукции. Для апробации этого предположения были выбраны две сложных системы, относящихся к металлам и неметаллам. В качестве металлов были выбраны никелевые сплавы, а в качестве неметаллов – бумага и картон для полиграфической промышленности. Впервые получены базы данных о химических составах сплавов на основе никеля и их жаропрочности, содержащая информацию о 310 марках, а также о минеральном составе 255 сортов бумаги и ее отражении по 36 спектральным зонам. Созданы специализированные программные продукты для тренировки ИНС, подготовки входных данных, а также получения и представления результатов. Предложен новый подход к процедуре обучения сети, являющийся оригинальным результатом численных методов, обеспечивающий СКО на уровне 15%
The main purpose of this work is to search for and test mathematical and hardware-software methods to establish patterns between the chemical compositions and properties of selected metallic and non-metallic materials in order to improve the service characteristics of the final product. To test this assumption, two complex systems related to metals and non-metals were chosen. Nickel alloys were chosen as metals, and paper and cardboard for the printing industry were chosen as non-metals. For the first time, databases were obtained on the chemical composition of nickel-based alloys and their heat resistance, containing information on 310 grades, as well as on the mineral composition of 255 grades of paper and its reflection in 36 spectral zones. Specialized software products have been created for ANN training, preparation of input data, as well as obtaining and presenting results. A new approach to the network training procedure is proposed, which is an original result of numerical methods, providing an RMS at the level of 15%
The main purpose of this work is to search for and test mathematical and hardware-software methods to establish patterns between the chemical compositions and properties of selected metallic and non-metallic materials in order to improve the service characteristics of the final product. To test this assumption, two complex systems related to metals and non-metals were chosen. Nickel alloys were chosen as metals, and paper and cardboard for the printing industry were chosen as non-metals. For the first time, databases were obtained on the chemical composition of nickel-based alloys and their heat resistance, containing information on 310 grades, as well as on the mineral composition of 255 grades of paper and its reflection in 36 spectral zones. Specialized software products have been created for ANN training, preparation of input data, as well as obtaining and presenting results. A new approach to the network training procedure is proposed, which is an original result of numerical methods, providing an RMS at the level of 15%