Kohonen network modelling for the strength of thermomechanically processed HSLA steel

Abstract

Primarily from the point of view of improvement of yield strength due to additions of niobium, titanium and boron in HSLA steels, the experimental steels are divided into five classes. The data are then supplied for learning a Self Organising Map (Kohonen network). It is found that the network with six neurons possesses better capacity of prediction with unknown data. Another effort of clustering the steels according to its major strength contributing mechanisms is also made. But the capacity of the network to cluster unknown data is found to be rather poor and has failed to follow from the metallurgical principles. To avoid this limitation, Learning Vector Quantisation method is adopted to impart a certain amount of supervision in the learning process and it is found that the training pattern of the network attains a good convergence thereby leading to a good predictive ability.