Data-driven modeling of mechanical properties of steel sheets obtained from an industrial production route

The steel industry faces significant challenges due to the necessary transition towards greener manufacturing processes. One approach to facilitate this transition is to increase the recycling rate by using more scrap metal instead of pure iron ore. However, this introduces various elements that affect the microstructure and, consequently, the mechanical properties of the steel. \\ Adjusting process parameters offers a way to counterbalance these impacts. Unfortunately, the optimal adjustments are often not well understood, and a comprehensive physical model of the entire production process is not available or feasible due to the complexity of the process, which involves numerous influencing factors and complex interactions. In response to these challenges, this thesis implements a data-driven approach. Various machine learning (ML) methods are used to create models that predict mechanical properties based on process parameters and chemical composition. This includes classic black-box models that learn from statistical data gathered throughout the entire process, from the production of liquid steel through all necessary production steps to produce high-quality, cold-rolled, batch-annealed low-carbon steel coils. Additionally, grey-box models are applied, where physical knowledge is combined with ML to potentially improve performance, trustworthiness, and understanding of the resulting models. A critical aspect of this work is thoroughly evaluating the strengths and limitations of different model types when applied to the available data and features, particularly in terms of prediction quality and interpretability. While the used grey-box models provide mathematical equations for estimating the target value, facilitating straightforward interpretation. Also classic black-box models can achieve similar interpretability using state-of-the-art feature importance analysis techniques that offer local explainability of any model.