Simulation of iron oxide reduction using machine learning an‎d response surface methodology

10/28/2025 12:00:00 AM

The direct reduction (DR) process plays a pivotal an‎d strategic role in the production of sponge iron for electric arc furnaces an‎d the steel industry. Consequently, its development an‎d optimization have always been of special significance. In recent years, AI has emerged as a powerful tool, enabling engineers to model an‎d predict the behavior of complex systems in this field. Various machine learning (ML) algorithms have been introduced to date, capable of identifying behavioral patterns among system data an‎d supporting process development. Iron ore reduction is a type of non-catalytic gas-solid reaction that takes place industrially inside shaft furnaces. To investigate an‎d analyze the reduction behavior an‎d reaction mechanisms on a laboratory scale, pellets are tested in a TGA, an‎d the extent of the reaction is determined based on weight changes. In this thesis, two distinct databases from two separate studies were utilized. In the first study, a new dataset was employed for ML to examine the effect of the initial weight of iron ore pellets. Two models, MLP an‎d RBF, were developed; comparison of performance metrics identified the RBF model as optimal, with a test accuracy of R² = 0.9923. Additionally, RSM revealed that in this study, the metallization degree was linearly related to initial pellet weight (R² = 0.9657), an‎d with an increase in initial pellet weight, the reaction rate decreased. In the second study, a more comprehensive database—previously investigated—was used, for which an optimized MLP model with test accuracy of R² = 0.9828 was developed. This study eva‎luated the performance of advanced ML models, including CatBoost, XGBoost, an‎d LightGBM, for predicting the metallization degree of iron ore during DR. The XGBoost model demonstrated the highest accuracy among the models (R² = 0.9905 on test data). Interpretability analysis using SHAP values indicated that reaction time an‎d temperature are the most influential variables for predicting the metallization degree; increased temperature an‎d reaction time lead to a significant improvement in metallization, while pellet size plays a minimal role. Additionally, a variety of three-dimensional plots were generated to assess the influence of different parameters. The results showed that temperature has a highly non-linear impact on metallization degree, an‎d as temperature increases, the reduction rate accelerates. Conversely, increased hydrogen an‎d decreased carbon monoxide concentrations in the reducing gas mixture also enhance the iron ore reduction rate. Keywords: Direct reduction, iron ore, iron oxide, machine learning, simulation

احيا مستقيم , شبكه عصبي

direct reduction , neural network

Shayan Goudarzi

Norollah Kasiri