FORECASTING PROPERTIES OF CARBON MATERIALS WITH THE USE OF NEURAL NETWORKS

Abstract

Nanoporous carbon materials (NCM) due to the large specific surface area and unique physical and chemical properties are widely used in various branches of science and production as catalytic, electrochemical and sorption materials.
Taking into account the widespread application and despite the deep and long history of their study carbon materials are very interesting research object, because of the quite large variety of raw materials, activation methods and chemical treatment of the surface. That's why defining of optimal parameters and conditions of technological processes of chemical activation and temperature treatment of plant raw materials for porous carbon materials obtaining with predetermined parameters of the porous structure is an urgent scientific and practical task. Optimization of the methods for nanoporous carbon materials obtaining requires a large number of experiments that require significant material and time costs. Therefore, the development of a mathematical model for the dependence of the characteristic of NCM porous structure on the technological conditions of obtaining is an urgent task.
The difficulty of physical processes in the investigated material causes significant difficulties in implementing a mathematical model with the help of mathematical dependencies in the form of a system of equations. Using the neural networks gives the possibility to avoid this problem.
The neural network structure is taken and it is considered that after training it will play the role of a mathematical model of a physical object which then can be used for prediction. The created neural network plays the role of an approximator, which makes it possible to determine the properties of the physical system at arbitrary values.
The ability of neural networks to approximate multidimensional functions is based on Kolmogorov's theory.
According to Kolmogorov's theorem the only one hidden with the number of neurons 2N+1, were N dimension of function is sufficient for approximation of an arbitrary function.
Numerical experiments have shown that a multilayered neural network can be used to predict the physical properties of nanoporous carbon materials.

Keywords: nanoporous materials, multilayered neural networks, prediction of material properties