Information field prediction of hydrogen-based shaft furnace based on numerical simulation and machine learning

This study develops a computational fluid dynamic (CFD) and machine learning (ML) coupling approach to predict information fields within hydrogen-based shaft furnaces, aiming to provide accurate description and rapid prediction of intra-furnace phenomena. This approach employs CFD simulations of key operating parameters to generate a database, which is subsequently utilized to train a genetic algorithm-optimized random forest model (GA-RF) for effective prediction of in-furnace carbon content and CO volume fraction. A back propagation neural network model with dropout (BP-Dropout) is implemented to achieve accurate predictions of gas/solid-phase temperatures, metallization degree, and H 2 volume fraction within the furnace. The results demonstrate that the CFD-ML coupling approach not only achieves high predictive accuracy, with all R2 values exceeding 0.975, but also significantly reduces computational time. Specifically, whereas the conventional CFD simulation requires one hour, the proposed CFD-ML method completes the prediction in only one minute. This highlights its potential for real-time application in intelligent metallurgy and provides a powerful tool for optimizing hydrogen-based shaft furnace ironmaking processes. • First developed a CFD-ML approach for accurate and rapid prediction of hydrogen-based shaft furnace 2D information fields. • Replacement of CFD calculations by CFD-ML coupling for substantial time scale reduction. • Adopts Genetic Algorithm to optimize hyperparameters of RF model and enhance prediction performance. • Adopts Dropout to optimize BP neural network model prediction performance.