Prédiction de la teneur finale en phosphore de l'acier dans le four à arc électrique (EAF) par l'application de réseaux de neurones artificiels

Abstract

The scrap-based electric arc furnace (EAF) is of increasing interest in steelmaking as it significantly reduces environmental emissions by recycling steel, thereby minimizing the need for raw material extraction and lowering greenhouse gas emissions associated with traditional steelmaking processes. However, it faces scientific and technical challenges, particularly in managing impurities such as phosphorus. This research work aimed to develop a robust predictive model based on machine learning or deep learning techniques to estimate the phosphorus content of steel as a function of input process parameters at the end of the EAF process. Several models were tested, and an artificial neural network (ANN) was selected for its high prediction accuracy of phosphorus content. First, data were collected on key parameters such as the chemical composition and weight of the scrap, the amount of lime added as fluxing agent, and the volume of oxygen injected. The data were provided by the industrial partner of the project FinklSteel (Sorel Forge). Once collected, the data were organized and preprocessed for use in model development. Several machine learning models were evaluated, but the artificial neural network model proved to be the most suitable for the problem studied. This model was optimized by adjusting various hyperparameters to maximize its performance. The model's performance was evaluated using several metrics, including MSE (0.00016), RMSE (0.0049), and correlation and determination coefficients R and R² (99% each). The results demonstrate that ANNs can provide highly accurate predictions, aiding in the optimization of industrial dephosphorization operations. However, one of the challenges of ANN models is their "black box" nature, making it difficult to interpret their predictions. To overcome this limitation, thermodynamic calculations were applied to better understand and interpret the physical and chemical phenomena of dephosphorization. These calculations were performed using the thermodynamic software FactSage version 8.3 and its optimized databases, allowing the relationship between input variables and dephosphorization outcomes to be more comprehensible. The comparison between the FactSage calculations and the correlations obtained in this study elucidated the net effect of each input parameter on the thermodynamics and kinetics of the dephosphorization process. This study contributes to the advancement of efficient phosphorus removal in EAF, enhancing the sustainability of steel recycling processes.

Date	17 Aug 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Elmira Moosavi (Supervisor) & Mohammad Jahazi (Co-supervisor)