Determination of the physical properties of heat treatable steels

T. Filetin, I. Žmak, D. Markučić, D. Novak:

Proceedings of the 8th Seminar of the International Federation for Heat Treatment and Surface Engineering , IFHTSE 2001, Dubrovnik- Cavtat, Croatia, 12 – 14 September 2001.

The following data: heat conductivity, specific heat, the coefficients of linear thermal expansion, the densities and other physical properties of steels versus temperature, are needed for calculation and simulation of heating and cooling processes of different heat treatment technologies. There is a restricted amount of data found in literature for steel grades in question – for the defined chemical composition of steel and temperature. In determining relevant data for the considered steels the statistical methods and/or methods of artificial intelligence for property prediction are applicable. An issue in this approach was the development of a method for predicting heat conductivity for the known chemical composition of steel at defined temperature. This paper presents the results of predicting heat conductivity depending on temperatures of different steels, using the regression analysis and by means of neural network. The regression equations for estimation of heat conductivities at different temperatures are defined. The model for statistical analysis is determined by genetic algorithm. Acceptable correlation between the input variables is estimated – the sum of element content of steel and temperature and the heat conductivities.
The static multi layer feedforward neural network is proposed to predict the heat conductivity of steels. To accelerate the convergence of the proposed static error-back propagation learning algorithm, the momentum method is applied. In the learning datasets,  41 different constructional, corrosion resistant and tool steels are used. The inputs for learning and testing were the chemical composition and the measured data from literature for heat conductivities at different temperatures (between 20 and 700  °C). The mean absolute value of error between the measured and the predicted data, and standard deviation for testing steel types are found to be acceptable. The results indicate the need for further testing of a wider dataset of steel groups and of other physical properties.

Download full article in PDF