NEURAL NETWORK MODELLING FOR PREDICTION OF FERRITE NUMBER IN STAINLESS STEEL WELDS

Abstract

As Neural Network (NN) has been very useful tool for recognizing a pattern in complex data, here the complex
data means the complex Non-Linear inter-relationships and inter dependence among the various materials/ process
parameters of the weld deposits concerned.
There is a significant role of the ferrite content in determining the fabrication and service performance of welded structures.
Because several properties can be predicted by estimating ferrite content. Research has also discovered that a minimum
ferrite content in steel is very necessary to improve the hot cracking resistance whereas higher ferrite leads to higher
mechanical and corrosion resistance of the ferritic-austenitic types, especially the “duplex” and “Super-duplex “types.
Though it has become the routine practice to predict the ferrite content from the schaffeler, Delong or WRC-1992 diagram.
All is designed on the basis on Ni-equivalents and Cr-equivalents .But different diagrams considers different respective
equivalent of Ni and Cr. But the limitations of using these diagrams lie in their linear or “pseudo-linear” relations and also
they don’t take in to consideration every chemical compound and the interaction between them. The relation between the
variables (dependent and independent) are more complex, i.e. Non-linear behavior. this problems can be overcome by using
a Neural Network modeling.we attempted Network architectures like Multilayer perceptron method, (MLP) ,and Radial
basis function (RBF), for building a neural networks and algorithms like Back propogation, (BP), conjugate gradient decent
(CG), quick propagation (QP),Levenberg Marquardt (LM), Delta bar delta (DD) and such many for training the NN model .
By applying Neural Network Modeling, we have trained several best optimized models for prediction of ferrite Number (
output) as a function of Chemical Composition (input) and Mechanical properties (Charpy toughness, Yield strength, %
elongation and Ultimate tensile strength) as a function of Chemical composition and ferrite number. Neural Network models
predicted the output well tuned with the experimental data and have also shown the Metallurgical trends. Successfully
trained NN model has been very useful tool for the cost reduction in the welding research and practice engineering field in
the terms of material, money and time saving aspects.