Application of AI Toolset to EAF Steelmaking - Key experiences

The steel use case in s-X-AIPI focuses on optimizing scrap utilization in Electric Arc Furnace (EAF) steel production to enhance product quality, minimize surface defects in the subsequent casting stage, and reduce energy consumption.

A key challenge is managing tramp elements like copper (Cu), tin (Sn), and phosphorus (P), which are prevalent in lower-quality scrap and can negatively impact steel properties. Due to variations in scrap composition and metallic yield, conservative safety margins are typically applied in charge calculations, leading to inefficiencies in material use and energy consumption. To address this, the s-X-AIPI project integrates an AI-driven decision support system into the steel production process. By leveraging machine learning and predictive modeling, the system enhances scrap characterization and process optimization. If discrepancies between predicted and actual steel composition arise, the optimization system helps the human operator to adjust the scrap-mixing-related production parameters to maintain the quality standards. Additionally, the system generates operator alerts to enable timely corrective actions, ensuring consistent decision-making across shifts.­

By providing accurate scrap mix suggestions, this solution enables more efficient and sustainable steel production, particularly when using low-quality scrap. This aspect is expected to gain importance in the steel decarbonization context as scrap is the main raw material of the low-carbon steel production alternative.

Short Description of the Results Achieved

  • Using machine learning models, such as Random Forest Regression chemical composition, temperature, energy consumption, and oxygen levels of liquid steel at the endpoint of the EAF process, can be accurately predicted. These predictions are based on input data from scrap charge composition and EAF process parameters.

  • By setting predefined thresholds for deviations between predicted and actual liquid steel properties, the system identifies inconsistencies in scrap characteristics. If these deviations occur frequently, the system triggers an alarm, signaling potential discrepancies in scrap quality.

  • The system continuously evaluates scrap properties, comparing new characterizations with historical data. A Human-in-the-Loop (HITL) operator reviews and validates these updates based on their expertise before the system accepts any modifications.

  • Based on a validated scrap characterization, the system suggests optimized scrap mix recipes for scheduled heats. Additionally, it provides chemical composition predictions for upcoming heats, enabling furnace operators to strategically plan energy use and material additions, improving process efficiency.

Figure 1: The components of the prototype. *

Validation Metrics used to Assess the s-X-AIPI Toolset’s Performance

Mean Absolute Error (MAE) was used as a key validation metric to evaluate the performance of the AI models. MAE provides a straightforward measure of the average absolute difference between predicted and actual values, helping to assess the model’s accuracy. Additionally, MAE was continuously tracked and analyzed over time to identify trends, detect potential model drift, and ensure consistent performance. This ongoing monitoring helps refine the models, improving their predictive capabilities and reliability in real-world industrial applications.

Figure 2: The Self-X optimization scheme, including HITL and the external, supportive AI services. **

Accuracy in Predicting Scrap Property Deviations Compared to other Methods

Compared to the previously used statistical hypothesis-based approach, s-X-AIPI’s AI-driven method, utilizing Multiple Linear Regression (MLR), provided more accurate and reliable predictions of scrap properties. By analyzing historical data and identifying key correlations, the MLR model was able to derive more precise scrap coefficients, reducing uncertainty in scrap characterization. This improved the estimation of scrap charge influence on liquid steel composition, leading to more effective decision-making in scrap mix optimization. The enhanced accuracy and adaptability of our model contributed to more consistent process control and reduced deviations in the final steel quality. The combination of the AI alarms and knowledge is a direct help to monitor the quality of a heterogeneous raw material (scrap).

Figure 3: Type of alarms generated by AI method (left) and AM (right)

Operational Efficiency

The Self-X AI Data Pipeline has been successfully implemented, demonstrating its ability to orchestrate various system components and integrate external AI services to enhance decision-making. By sharing production-related metadata, the AI Pipeline enables real-time monitoring, anomaly detection, and optimization in steel production.

The Self-X capabilities, powered by AI, provide a resilient, self-adaptive system that can detect deviations, evaluate their impact, and take corrective actions. This approach enhances process stability and efficiency, particularly in managing scrap properties, energy consumption, and liquid steel composition.

Key measurable impacts include:

  • Improved Model Maintenance: The AI Pipeline continuously monitors production performance and ensures that the optimization models remain current, reducing the need for manual adjustments.

  • Enhanced Human-in-the-Loop (HITL) Decision Support: While AI-driven alarms highlight potential issues and suggest actions, the final decision remains with human experts, ensuring reliability and trust with AI.

  • Reduced Scrap-Related Uncertainties: AI-powered scrap characterization and real-time deviation detection allow for more accurate scrap mix optimization, minimizing production risks.

Figure 4: Scrap Mix optimizer dashboard

Human-AI collaboration in Steel Production

Workers responded positively to the AI-driven suggestions, particularly appreciating the scrap characterization, scrap mix optimization, and deviation detection with automated alarms. The AI system provides clear, actionable insights helping professionals make more informed decisions. The ability to characterize scrap and optimize its mix is particularly valuable, as it allows better control over steel composition and energy efficiency. The workers are impressed how the system identifies deviations in scrap properties and raises alarms before issues escalate.

Furthermore, the monitoring and automated alerts help workers react proactively. While AI assists in detecting issues, the final decision remains with the professionals, ensuring they retain control while benefiting from enhanced data-driven decision-making.

Overall, the integration of AI has streamlined their workflow, improved process stability, and increased confidence in managing production variability.

Looking Forward

Based on the validation results, there is high confidence in scaling this solution. The AI-driven system has demonstrated consistent accuracy in predicting scrap properties, optimizing the scrap mix, and detecting deviations. The validation process confirmed that the Self-X AI Data Pipeline effectively integrates production data with external AI services, ensuring reliable decision support for operators.

Additionally, the system’s modular and adaptive architecture makes it scalable to different steel production set-ups. While some refinements may be required to accommodate site-specific constraints, the overall results suggest that the solution can be expanded to enhance efficiency, reduce process variability, and improve decision-making in steel production at a larger scale.

*Figure from: P. Kannisto, Z. Kargar, G. Alvarez, B. Kleimt, and A. Arteaga, "Resilient, Adaptive Industrial Self-X AI Pipeline with External AI Services: A Case Study on Electric Steelmaking", Processes, vol. 12, no. 12, p. 2877, 2024. https://doi.org/10.3390/pr12122877 . The figure is original except the red annotations on the right side. Copyright 2024 © the authors. Licensed under CC BY 4.0; https://creativecommons.org/licenses/by/4.0/
**Figure from: P. Kannisto, Z. Kargar, G. Alvarez, B. Kleimt, and A. Arteaga, "Resilient, Adaptive Industrial Self-X AI Pipeline with External AI Services: A Case Study on Electric Steelmaking", Processes, vol. 12, no. 12, p. 2877, 2024. https://doi.org/10.3390/pr12122877 . The figure is original except the red annotations on the right side. Copyright 2024 © the authors. Licensed under CC BY 4.0; https://creativecommons.org/licenses/by/4.0/

Next
Next

Be part of the change! Your input is key to driving AI adoption in industry