Calcium-Treated Steel Cleanliness Prediction Using High-Dimensional Steelmaking Process Data

Control of calcium treatment in steel is challenging due to the reactivity of Ca and difficulty of measuring total oxygen of steel in-process to make actionable decisions. In this work, a method combining statistics and process engineering are developed using partial least squares regression (PLS) to predict non-metallic inclusion content (oxides and CaS) and composition at the end of ladle treatment and in the tundish using extensive process data and SEM/EDS-based non-metallic inclusion analysis. Total oxygen at the end of the ladle treatment can be predicted to an accuracy of 7 ppm, and the Mg/(Mg+Al) ratio in inclusions to an accuracy of 3at% providing enough data to recommend Ca addition based on a thermodynamic calculation for the Ca liquid window. Alternatively, the model can predict total inclusion fraction to 20 ppm accuracy, and accurately predict average CaS content and Ca/Al ratio of inclusions in the tundish. Model interpretability is hindered by high dimensionality and multicollinearity of the data. Non-metallic inclusion compositions correspond to the expected composition at the onset of CaS formation based on steel composition.

Automated summary

A method combining statistics and process engineering using partial least squares regression (PLS) is developed to predict non-metallic inclusion content and composition in steel. The model can accurately predict total oxygen content and Mg/(Mg+Al) ratio, providing enough data to recommend Ca addition based on thermodynamic calculation. It can also predict total inclusion fraction and accurately predict average CaS content and Ca/Al ratio in the tundish. However, model interpretability is hindered by high dimensionality and multicollinearity of the data. Non-metallic inclusion compositions correspond to the expected composition at the onset of CaS formation based on steel composition.