Predictive Modeling and Inverse FEM Simulation for Enhancing Steelmaking Operations

This thesis implements state-of-the-art methods and introduces novel approaches in data processing, machine learning, and Finite Element Method (FEM)-based simulations to enhance the operation of a specific blast furnace in Indiana, USA. We direct our efforts into incorporating myriads of multi-sourced, incongruent, and semi structured data to build predictive tools that offer insights into furnace conditions and subsequent decision-making process. Two primary models are developed—one currently in beta testing and another transitioning from alpha to beta phase. The first model focuses on predicting hot metal silicon content, a key indicator of furnace performance. By employing a generalized data processing scheme and a robust XGBoost-based modeling pipeline, the model achieves 91% accuracy over 200 real test data predictions using only the previous operating conditions. In turn, the model can provide real-time guidance for operational adjustments that reduce risk and improve performance. The second model provides an estimate of hearth erosion, a critical concern due to the severe thermal and mechanical stresses in that zone. The efficient and minimal FEM-based model uses conductivity driven formulations, vectorized SIMD and parallelization operations and residual gradients for optimization to rapidly estimate erosion profiles without mesh reconstruction or proprietary CFD tools. It achieves 5–25°C temperature accuracy with convergence in 3-5 minutes on a 6 core CPU. Validated with industrial data, it supports long-term erosion tracking and safety planning despite steady-state assumptions. Together, these contributions offer practical, high-impact tools for improving furnace safety, reducing downtime, extending campaign life, maintaining productivity, reducing associated costs and facilitating better and on-time decision making.