This article is aimed to study theoretical knowledges related to deep steel desulfurization and their practical application in specific production conditions. The main goal of this study was to verify the efficiency of added CaSi to processed heats in two batches, i.e., first for deep desulfurization at the beginning of the secondary metallurgy and then in a second smaller batch at the end of that process for final modification. At the beginning of the secondary metallurgy process, it was crucial to prepare a suitable chemical composition of ladle slag with advantageous properties for good assimilation of sulfur-based inclusions from desulfurization reactions. Samples were taken of final basic oxygen furnace slag, ladle slag before the deep desulfurization process at the beginning of secondary metallurgy and ladle slag at the end of processing. Detailed cleanliness analyses of steel samples taken from mold during the casting were performed using the Automated Steel Cleanliness Analyses Tool. For better understanding and interpretation of the obtained cleanliness results and their relationship, the following defined parameters were used: ratio of “Ca/Al” oxide inclusions, percentage of “liquid” inclusions and area of CaS inclusions.

Nitrogen control in basic oxygen furnace (BOF) steelmaking is critical, as dissolved nitrogen concentrations exceeding 30–40 ppm detrimentally affect the mechanical properties and formability of low-carbon steel products; however, no prior study has applied a physics-informed machine learning model to nitrogen prediction at this process stage. A NitroPINN model was developed incorporating a multiplicative prediction structure that embeds Sievert’s law equilibrium, Wagner interaction coefficients, and Byrne–Belton surface blockage theory directly into the model. The model was trained and evaluated on 66 matched industrial heats from a top-blown 170-ton BOF converter, characterized by 16 physics-informed features, and benchmarked against ridge regression and a pure multilayer perceptron (MLP) under five-fold cross-validation. The NitroPINN achieved the lowest mean absolute error (MAE = 5.60 ppm) and mean absolute percentage error (MAPE = 27.2%) among the three models, whilst the learned equilibrium attainment factor η averaged 0.456 ± 0.028, consistent with sub-equilibrium nitrogen conditions imposed by intense CO flushing during oxygen blowing. All three models exhibited comparable overall accuracy, confirming that dataset size constitutes the principal performance bottleneck. The primary advantage of the NitroPINN lies in its physical interpretability, constraining predictions to metallurgically plausible ranges and providing a transparent decomposition into thermodynamic and kinetic contributions.