Artificial Intelligence Approaches to Determine Graphite Nodularity in Ductile Iron

The complex metallurgical interrelationships in the production of ductile cast iron can lead to enormous dierences in graphite formation and local microstructure by small variations during production. Articial intelligence algorithms were used to describe graphite formation, which is inuenced by a variety of metallurgical parameters. Moreover, complex physical relationships in the formation of graphite morphology are also controlled by boundary conditions of processing, the eect of which can hardly be assessed in everyday foundry operations. The inuence of relevant input parameters can be predetermined using articial intelligence based on conditions and patterns that occur simultaneously. By predicting the local graphite formation, measures to stabilise production were dened and thereby the accuracy of structure simulations improved. In course of this work, the most important dominating variables, from initial charging to nal casting, were compiled and analysed with the help of statistical regression methods to predict the nodularity of graphite spheres. We compared the accuracy of the prediction by using Linear Regression, Gaussian Process Regression, Regression Trees, Boosted Trees, Support Vector Machines, Shallow Neural Networks and Deep Neural Networks. As input parameters we used 45 characteristics of the production process consisting of the basic information including the composition of the charge, the overheating time, the type of melting vessel, the type of the inoculant, the fading, and the solidification time. Additionally, the data of several thermal analysis, oxygen activity measurements and the final chemical analysis were included.Initial programme designs using machine learning algorithms based on neural networks achieved encouraging results. To improve the degree of accuracy, this algorithm was subsequently adapted and rened for the nodularity of graphite.