There are considerable demands for the development of weld metals for high-strength low-alloy steels. To deal with this, a neural network was trained and tested on a set of data obtained on weld metals for steels of the type used for shipbuilding. The input variables for the network were the chemical elements and the weld cooling rate. The outputs consisted of the yield and ultimate tensile strengths, elongation and reduction of area. Many models were created using different network configurations and initial conditions. An appropriate committee of models was then assembled by testing the ability of the models to generalise on unseen data. The neural network technique used is due to MacKay, with a Bayesian framework and hence allows the estimation of error bars which warn the user when data are sparse or locally noisy. The method revealed significant trends describing the dependence of mechanical properties on weld composition and cooling rate.
Science and Technology of Welding and Joining, Vol. 6, 2001, 116-124.
Download PDF file