TAGRA 2025

High-quality steel production is influenced by numerous factors, including chemical composition (particularly carbon content), process parameters (heat-treatment durations, charge amounts, preheating values), scrap quality, and the control mechanisms utilized during the production cycle. The relationships among these variables and the impact of process parameters on steel quality can be identified using machine learning methods. This study presents a data-driven approach for determining steel quality classes by using three months of steel production data obtained from a manufacturing company in Türkiye. The operational dataset consists of 39 variables, including preheating values, scrap charge information, process durations, and chemical composition measurements. On the raw dataset, missing-value analysis, removal of inconsistent records, scaling of numerical variables, encoding of categorical variables, and splitting into training (80%) and test (20%) sets were performed. Ultimately, machine learning models—Support Vector Machine (SVM), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost) and Random Forest (RF)—were trained using 3,817 observations to predict steel quality classes (B420 – S220). The results indicated that the quality class was clearly separated by the carbon content in the chemical composition; when carbon analysis values were included in the models, RF, XGBoost, LightGBM, and SVM all achieved 100% accuracy and AUC = 1.00. This finding demonstrates that carbon content is the primary determinant of the quality decision. To further examine the effect of carbon, the carbon variable was removed from the dataset, and the models were retrained using 5-fold cross-validation. In this case, the accuracy of tree-based methods decreased to approximately 86%, while the SVM model achieved 65% accuracy. ROC analyses performed without cross-validation showed that AUC values ranged between 0.93 and 0.95 in the absence of the carbon variable. This indicates that process parameters contribute to quality classification to a certain extent.