Data-driven optimisation of rebar manufacturing using deep neural networks: Manganese reduction and mechanical property control

This study presents the development of deep neural network (DNN) models using a large-scale industrial dataset containing 39,389 data points to optimise rebar manufacturing processes. The primary objective is to predict the mechanical properties of hot-rolled steel bars based on rolling process parameters and liquid steel chemical composition. The predicted properties include yield strength (YS), ultimate tensile strength (UTS), UTS/YS ratio, percentage elongation (A) and elongation under maximum load (Agt). By accurately forecasting these properties, the models determine the minimum required manganese (Mn) content and reduce unnecessary Mn additions. An integrated dual-DNN framework was developed, featuring two models, each featuring two hidden layers with 100 nodes. The first model utilises 16 input parameters to optimise manganese consumption, while the second model evaluates 27 inputs, incorporating both chemical and production variables to calculate quenching water pressures and predict final properties. This 27-parameter model achieved high predictive accuracy with R ² values of 0.9789 for YS and 0.9701 for UTS. The developed models were implemented in a commercial-scale steel plant. The implementation resulted in a 12% reduction in manganese consumption and significantly reduced process variability. This optimisation improves cost efficiency and resource utilisation while preventing approximately 1020 tons CO ₂ -eq/year, based on the GWP20 climate-change impact value reported for conventional ferromanganese production, in a 1 million ton/year steel plant. Therefore, the proposed approach contributes to sustainable and green steelmaking practices.