PREDICTING HAZ HARDNESS WITH ARTIFICIAL NEURAL NETWORKS

The use of artificial neural networks (backpropagation networks) for p redicting heat-affected zone hardness, given the 800-500 degrees C coo ling time and chemical composition, is investigated in this study. The experimental training data are taken from a database assembled by Yur ioka ei al. Network predicted hardness values are compared with experi mental values from the entire Yurioka database and reasonable agreemen t is found (correlation factor = 0.98). The network results are also c ompared with values calculated from the regression relationships of Yu rioka and Suzuki based on the same database. Finally, an optimal netwo rk architecture (1 hidden layer, 4 hidden nodes and 40 training patter ns) is suggested.