Multi-task deep learning for simultaneous prediction of steel purity and carbon capture rate using membrane separation technology in integrated steelmaking processes

Abstract

Steel production significantly contributes to global CO₂ emissions, demanding simultaneous optimization of product quality and environmental performance. Current prediction models address steel purity and carbon capture separately, missing opportunities for integrated process optimization. This work presents the first comparison of single-task learning (STL) versus multi-task learning (MTL) for simultaneous prediction of iron purity classification and carbon capture rates in membrane-integrated steelmaking processes. Deep neural networks (DNNs) were trained on 1,473 validated simulation data points with 30 input features covering raw materials, operating conditions, and membrane specifications. The MTL architecture employed shared hidden layers with task-specific output branches, utilizing ReLU activation functions and Adam optimization. STL achieved 97.62% accuracy with perfect recall for iron purity classification, while MTL demonstrated superior carbon capture prediction (R² = 0.9948 vs 0.9902), representing 30% improvement through shared process learning. Feature importance analysis revealed air flow rate as the dominant factor for iron purity, while membrane feed pressure controlled carbon capture performance. Results demonstrate strategic model selection for steel optimization: STL for critical quality control requiring zero false negatives; MTL for integrated processes leveraging parameter interactions. This framework enables simultaneous steel quality and environmental enhancement, advancing sustainable steelmaking and multi-objective optimization in process industries.