Accurate endpoint control in basic oxygen furnace (BOF) steelmaking is essential for reducing production costs and improving steel quality. To overcome the limited mechanism support and poor transparency of purely data-driven models, this study proposes a physics-aware and interpretable framework for cumulative decarburization prediction based on real industrial data. Historical multi-heat data from the same converter were integrated, and an averaged full-spectrum cross-correlation method was used to estimate and correct the transport delay of off-gas signals, thereby constructing a heat-wise large-sample dataset. Key elemental features with clear physical significance were then extracted from high-dimensional flame spectra by incorporating their underlying radiation mechanisms. On this basis, a Stacking-based ensemble model was developed for cumulative decarburization prediction, and SHAP was introduced to interpret the model decision logic. Results show that the proposed framework outperforms conventional single models and purely data-driven dimensionality reduction methods. SHAP analysis further indicates that model decisions are mainly dominated by four core elemental spectral features, namely Fe, C, O, and Mn. Overall, the proposed method combines predictive performance, physical constraints, and interpretability, and provides a new solution for auxiliary soft sensing and decision support in BOF endpoint control.
An et al. (Sun,) studied this question.