Machine learning approach for predicting tramp elements in the basic oxygen furnace based on the compiled steel scrap...

Area of Science:

• Metallurgy and Materials Science
• Data Science and Machine Learning

Background:

• Steel production relies on pig iron and scrap, but scrap composition variability complicates predicting final chemical analysis.
• Current scrap mix compilation often depends on experience and trials, leading to potential inefficiencies.

Purpose of the Study:

• To develop a machine learning model for predicting chemical element content in steel converters.
• To enable optimized scrap mix utilization and maintain melt quality despite variable scrap availability.

Main Methods:

• Utilized a dataset of approximately 115,000 heats for model development.
• Applied the XGBoost machine learning algorithm to predict tramp element concentrations (Cu, Cr, Mo, P, Ni, Sn, S).
• Implemented an online model with a synchronous interface for real-time simulation and optimization.

Main Results:

• Successfully predicted the content of tramp elements (copper, chromium, molybdenum, phosphorus, nickel, tin, and sulfur) at the end of the basic oxygen furnace process.
• Demonstrated that predictions can be made using routinely collected, pre-existing data without additional sensors.
• The online model facilitates the simulation of new input material combinations to maintain melt quality.

Conclusions:

• Machine learning, specifically XGBoost, offers a viable solution for predicting chemical composition in steel converters.
• Accurate scrap management and high-quality upstream data are crucial for effective model performance.
• The developed model supports adaptive scrap selection to ensure consistent steel quality.