Abstract
There have been many attempts in the past to reduce the variety of steels produced without compromising the range of properties available. The objective in doing so is to reduce costs. These attempts have largely focused on experiments in which thermomechanical processing parameters are varied in order to obtain a range of properties without changing the chemical composition.
We have approached this problem differently by examining a large database on hot-rolled steels, all with a final microstructure which is a mixture of ferrite and pearlite. Alloys like these form the bulk of all the steel production in the world. The database was used to create a neural network model in a Bayesian framework with inputs consisting of the chemical composition (C, Mn, Si, P, S, Cr, Ni, Mo, Ti, Nb, V, Al, N, B, Cu), the finish deformation temperature (T F ), and the coiling temperature (T C ); the output was either the ultimate tensile strength or tensile elongation to failure. The neural network model was then combined with a genetic algorithm to seek domains of input parameters which correspond to different target properties. The results revealed ambiguity in the inputs with respect to the required property, and the trends observed within these sets of inputs were found to be metallurgically meaningful.
ACKNOWLEDGMENTS
The authors are grateful to Professor Hae-Geon Lee for the provision of laboratory facilities at the Graduate Institute of Ferrous Technology, POSTECH, Korea, and to Doctors Jae Kon Lee, Young Roc Im, and Jung Hyeung Lee of POSCO for their help with the neural network analysis.