The production of steel coils with scrap material using an electric furnace (EAF) results in a very low CO2 emission compared to traditional production in blast furnace followed by basic oxygen steelmaking, but introduces many foreign elements by scrap. The impact of these foreign elements on the mechanical properties, such as the plastic strain ratio (r-value), is in many cases not understood entirely and the role of nano-precipitates are not captured by the process analysis. Predicting the r-value that determines the deep drawing capability of steel coils is a prerequisite for producing high-quality flat steel by EAF route.

In this work we apply AI regression models for predicting the r-value of steel coils from chemical composition and process parameters. The data from steel production and tensile tests was provided by voestalpine Stahl GmbH and includes a full chemical analysis, as well as many parameters measured during all working steps of the process and the resulting mechanical properties. As a prerequisite for training of AI models, the data needs to be understood, analyzed, checked, and unreasonable data be removed (data cleaning). Additionally, methods for data fusion are investigated. The result is a machine-readable dataset fit for various modelling tasks. The used models include Random Forest Regression, Support Vector Regression, Artificial Neural Networks and Extreme Gradient Boost. We document the effort for hyper-parameter tuning and training for each model type and compare their prediction accuracy. Based on the insights gained, we present strength and limitations of different model types with the available data and number of features. We discuss possible improvements like introducing prior physical knowledge.