Elucidating Graphite Formation Mechanisms of Ductile Cast Iron by Means of Artificial Intelligence

Ductile cast iron (GJS), as a traditional structural material, is still widely used today. The complex metallurgical interrelationships in ductile cast iron production can lead to enormous differences in the formation of graphite and the local microstructure by small variations during production. Elucidating the mechanism of graphite formation is important for controlling graphite morphology and improving ductile cast iron performance. However, complex physical relationships in the formation of graphite morphology are controlled by adjusting the processing boundary conditions, of which the effect can hardly be assessed in real daily foundry operations. In this thesis, well-designed experiments on laboratory and industrial scale were therefore performed to elucidate the mechanism of graphite formation, although the graphite formation is influenced by a variety of metallurgical parameters. More than 600 met allographic samples were analyzed from test specimens cast under precisely controlled manufacturing conditions. The samples were produced under laboratory conditions at the Austrian Foundry Research Institute and under industrial boundary conditions in three foundries. Furthermore, artificial intelligence algorithms were used to elucidate graphite formation based on the results obtained from well-designed experiments. The influence of relevant input parameters can be predetermined using artificial intelligence based on conditions and patterns that occur simultaneously. By predicting the local formation of graphite, measures to stabilize the production were defined and, thereby, the prediction accuracy of the graphite structure was improved. Throughout this work, the most important dominant variables, from initial charging to final casting, were compiled and analyzed with the help of statistical regression and classification methods to predict nodularity, sphere density, and amount of ferrite, respectively. Different machine learning algorithms were tested to predict nodularity. They were compared according to the root mean square error (RMSE) and the factor of determination (R2). Linear Regression (LR), Gaussian Process Regression (GPR), Regression Trees (RT), Boosted Trees (BT), Support Vector Machines (SVM), Shallow Neural Networks (SNN) and Deep Neural Networks (DNN) were applied to the prepared dataset to predict nodularity. The highest accuracy could be achieved by DNN. Therefore, the final models to predict nodularity, sphere density, and the amount of ferrite were based on this supervised modeling technique. To predict the amount of ferrite, an additional classification model was used before the regression model. This is necessary because the model is only valid for ferritic-perlitic GJS grades. If the model assumes that the sample does not consist of a purely ferritic-perlitic phase, the regression model is not used to determine the ferrite content. Combining the classification model with the regression model achieved a higher accuracy. The innovation of this paper is that it describes regression and classification methods to predict the microstructure of GJS using 45 input parameters including parameters of the chemistry, the metallurgy, the cooling rate, and time-based parameters. For new castings, without information on the local cooling rate, preceding thermal modeling with commercial simulation software used in the foundry industry can be used instead of temperature measurement in the castings. According to the trained model, for a particular sample, graphite nodularity is strongly dependent on solidification time, overheating time over 1500°C, fading, type and amount of inoculant, as well as on the Mg and Ce content.