Modeling of changes in heat resistance of nickel-based alloys using bayesian artificial neural networks

Abstract

Resource design of gas turbine engines and installations requires extensive information about the heat resistance of nickel-based superalloys, from which the most critical parts of aircraft and marine engines, pumps of gas-oil pumping stations and power plants are made. The problems are that the data on the heat resistance obtained as a result of testing each alloy under study are quite limited. In the present paper, the task of modelling changes in the heat resistance of nickel-based superalloy on the basis of available experimental data is solved. To solve the task, the most modern approach, the neural network modeling method, was applied. The input data are chemical compositions of heat-resistant nickel-based superalloys and the values of their heat resistance obtained experimentally. The output data are the calculated values of heat resistance modeled by an artificial neural network. In the course of the work, transformations of the input data were carried out to reduce the standard deviation of the modeling of the output data. The choice of the neural network configuration was made in order to achieve the highest possible accuracy. As a result, a neural network of direct error propagation was used, with 27 neurons on the input layer, 13 neurons in the hidden layer and 1 neuron in the output layer. To validate the results of the predictions, a group of alloys with the maximum number of known experimental values of heat resistance was randomly selected before the input of data into the network. After preparing the data, selecting the configuration and training the network, the chemical compositions of the selected group were loaded and their heat resistance values were calculated. Comparison of the obtained data with the experimental data showed high efficiency of the method. As a result, data on the change of heat resistance for the studied alloys were obtained and an analytical expression describing the obtained dependences was formulated. © 2020, Institute for Metals Superplasticity Problems of Russian Academy of Sciences. All rights reserved.