A coupled data-driven kinetic model for phosphorus removal in BOF converter steelmaking

To address the challenge of in-process accurate prediction of phosphorus content in molten steel during top-bottom combined-blowing converter smelting, this article proposes a data-driven multi-region coupled kinetic model for phosphorus removal. First, based on actual industrial production data, three machine learning models, namely backpropagation neural network, random forest and extreme gradient boosting, were constructed to predict the terminal steel phosphorus content (temperature–sampling–oxygen phosphorus content [TSOP]). A weighted ensemble strategy was then employed to generate high-precision TSOP predictions. The predicted TSOP is fed as input to a kinetic model, which incorporates the multi-region mass transfer mechanisms of the converter's emulsion zone and slag–steel interface zone, leading to the establishment of a modified cross-region diffusion-aware phosphorus removal kinetic model. Using an industrial data-driven parameter inversion technique, systematic optimisation and calibration of key mass transfer and diffusion coefficients are conducted to effectively improve prediction accuracy, especially within the narrow, high-precision error bands critical for industrial quality control. Results show that during the temperature–sampling–carbon phase, the hit rate for phosphorus content with an absolute error of ±0.025% reaches 89.94%; during the temperature–sampling–oxygen phase, this rate increases to 96.18% for an absolute error of ±0.01%. This model synergistically integrates the advantages of data-driven and mechanism-based modelling, thereby enabling accurate characterisation of the dynamic evolution during converter phosphorus removal.