Phenomenological Prediction of Defect Formation in Continuous Casting of Steel Utilizing Fundamental Modeling and Machine Learning

Defect formation during continuous casting results in the downgrading and rejection of cast products at steel plants. This article aims to predict transverse cracks and slag defects in continuous casting with decision trees utilizing phenomenological features extracted from simulations and real‐time process data. To facilitate accurate simulation of the continuous casting process under varying conditions, a transient heat transfer model called CastManager is coupled with InterDendritic Solidification (IDS). More specifically, the IDS model simulates the solidification and microstructure evolution of steels based on the transient temperature profiles provided by CastManager. From the caster of SSAB Raahe Works (Finland), slab‐specific temperature profiles and measured steel compositions are compiled as simulation inputs. Defect data collected from cast slabs and rolled plates formulate the labels for binary classification. Classification performance is evaluated with a nested and stratified fivefold cross‐validation scheme. For peritectic Nb–V and low‐carbon B–Ti microalloyed steels, average G‐mean scores of 0.93 and 0.94 are achieved in predicting transverse cracks. For low‐carbon Nb–B and B–Ti microalloyed steels, average G‐mean scores of 0.90 and 0.92 are achieved in predicting slag defects. The final decision trees are interpreted to assess their metallurgical admissibility and to identify measures for preventing defects.