A Neural Network Approach to the Prediction of Submerged Arc Weld Metal Chemistry

A neural network technique has been employed to predict submerged arc weld metal chemistry, using a database for which a previous linear regression model had been developed. Thus, a comparison may be made of neural network and regression approaches. Weld metal chemistry is a complex function of interactions between the welding electrode, baseplate and shielding medium. It is demonstrated that simplifying assumptions made in the regression analysis, i.e. that the final weld metal composition is largely dependent on the plate and wire chemistries, are justified when the dataset is restricted to similar weld process variables. However, for a wider range of flux chemistries, for example, the neural network recognises the more complex interrelationships within the data. Similarly, complex models are generated by a neural network to predict properties such as weld metal toughness for which a simple relationship between input variables can not be derived.