Application of Machine Learning Algorithms for Prediction of Iron Content in Iron Ore Flotation Tailings

Iron ore concentration by flotation involves multiple interdependent variables, making monitoring and control key to ensuring operational efficiency and product quality. However, many measurements still depend on laboratory analysis, resulting in high response time and delay in operational adjustments. In this work, three algorithms based on decision trees, XGBoost, LightGBM and Random Forest, and two based on neural networks, TabNet and MLP, were compared to predict the iron content in flotation tailings. TabNet had the lowest median MAE, with 1.82, and a mean of 1.85, indicating better performance in absolute error. However, the LightGBM obtained a very close result, with a median of 1.86 and a mean of 1.88, in addition to less dispersion of errors. XGBoost and Random Forest performed mid-range, while MLP underperformed. Thus, LightGBM stood out for the best balance between predictive performance and stability.