PROCESS OPTIMIZATION / 2026
Data-Driven Process Optimization in Steel Manufacturing
Applying time-series analysis and reinforcement learning to reduce energy consumption in electric arc furnace operations.
ABSTRACT
Data-Driven Process Optimization in Steel Manufacturing
Reinforcement learning offers a framework for autonomous process optimization, but industrial deployment faces safety and sample efficiency constraints. We present a constrained RL approach for electric arc furnace optimization using a high-fidelity simulator for pre-training and safe policy transfer to production. Results show 12% energy reduction with zero safety violations during learning.
ARTICLE
MAIN CONTENT
Electric arc furnaces (EAFs) consume approximately 400 kWh per tonne of steel. Power profile optimization — the trajectory of electrical power input over the heat — represents the largest lever for energy reduction. However, the optimization space is high-dimensional (power, oxygen, carbon injection, scrap charging timing) and constrained by safety limits (roof temperature, electrode stability, refractory wear). We frame EAF optimization as a constrained Markov Decision Process. The state space includes real-time sensor data (electrical, thermal, gas, mechanical). The action space comprises power setpoints, oxygen lance flow, carbon injection rate, and charging decisions. Constraints enforce thermal limits, electrical stability, and product quality specifications.
FIGURE
Constrained RL Training Pipeline
Simulator pre-training with safety constraints, followed by conservative policy transfer.
CONTINUED
DISCUSSION
Training proceeds in three phases: (1) physics-informed simulator pre-training with domain randomization, (2) conservative policy optimization with constraint satisfaction guarantees, (3) production deployment with human-in-the-loop oversight and gradual autonomy increase. The simulator integrates a reduced-order thermal model, electrical circuit model, and scrap melting kinetics — validated against 2,000+ production heats. Production results at a 120-tonne EAF: 12% specific energy reduction (48 kWh/t), 8% tap-to-tap time reduction, zero safety constraint violations during 6-month deployment. The policy generalizes across scrap grades and ambient conditions without retraining.
CONCLUSION
CONCLUSION
Constrained reinforcement learning with simulator pre-training enables safe, sample-efficient optimization of industrial processes. The approach is transferable to other high-temperature processes (cement kilns, glass furnaces) where safety constraints and sample efficiency are paramount.
RELATED RESEARCH
RELATED\nPUBLICATIONS.
Explore additional research in industrial intelligence and digital twins.
CONTACT
DISCUSS THIS\nRESEARCH.
We welcome collaboration inquiries, peer review, and industrial validation partnerships.