Decision support system for optimizing BOF tapping cut-off in low-P steel production using machine learning and constraint programming

Controlling phosphorus in ultra-low phosphorus steel often requires an early basic oxygen furnace (BOF) tapping cut-off to limit carryover slag, reducing tapping quantity (<i>TQ</i>) and increasing yield variability. This study develops a BOF tapping cut-off decision support system (BTDSS) that combines machine learning, similarity retrieval, and constraint programming to recommend a risk-controlled tapping cut-off. From 9,980 raw BOF and Ruhrstahl Heraeus (RH) heats, 6,281 heats were retained after filtering missing values, outliers, and metallurgically inconsistent data. RH end-point phosphorus (%P_RHend) was predicted from 26 pre-tapping features using XGBoost with recursive feature elimination and cross-validation, achieving an adjusted R² of 0.9774 and a 10.41% mean absolute percentage error. For a query heat, comparable heats are retrieved within each steel grade group by principal component analysis of 12 pre-tapping variables and cosine similarity to form an empirical <i>TQ</i> distribution. Gaussian and Gaussian mixture models are compared using Akaike and Bayesian information criteria. When multimodality is detected, the higher-mean mode is retained and tail trimming is applied. Trimmed quantiles determine the predicted tapping quantity (<i>TQ_pred</i>) and risk-controlled maximum tapping quantity (<i>TQ_max</i>). A phosphorus safety margin, based on the model prediction, error buffer, steel grade phosphorus limit, and similar-heat phosphorus distribution, determines the upper quantile used for <i>TQ_max</i>. Constraint programming searches the feasible region bounded by <i>TQ_pred</i> and <i>TQ_max</i> for tapping guidance. Across grade groups, <i>TQ_pred</i> was 0.34% higher than the similar-heat mean and predicted tapping yield (<i>TY</i>) reached 95.1%. The recommended <i>TQ_max</i> provided 2.2% headroom, with a maximum <i>TY</i> of 96.9%. Retrospective validation using 1,736 new heats showed that the BTDSS-derived recommendations increased predicted <i>TY</i> from 94.9% to 96.2% and reduced the yield range from 13.2% to 4.5%. In the <i>TQ_max</i> scenario, the additional recoverable molten steel was 13,107 tonne, indicating that BTDSS can provide risk-aware tapping guidance with potential plant-scale benefits.