Physics-Informed AI for Blast Furnace Silicon Prediction

Integrated steel mills are increasingly turning to physics-informed artificial intelligence (PIAI) to optimize blast furnace operations. This article provides a comprehensive overview of how steelmakers combine fundamental metallurgical laws with machine learning to accurately predict hot-metal silicon content. By using silicon as a primary thermal indicator, these hybrid AI models give operators early warnings of temperature drops, allowing them to take corrective action and prevent catastrophic hearth chilling.

The blast furnace is the heart of an integrated steel mill, converting iron ore, coke, and limestone into liquid iron, known as hot metal. Because the internal environment of a blast furnace reaches temperatures up to 2,000 degrees Celsius, direct continuous measurement of the internal thermal state is nearly impossible. Instead, operators rely on the silicon content of the tapped hot metal as a proxy for the furnace's thermal energy. High silicon levels indicate a hot furnace,which wastes expensive coke, while low silicon levels signal a cooling furnace.

If a furnace loses too much heat, it risks a severe operational failure known as hearth chilling. Hearth chilling occurs when the temperature drops so low that the molten iron and slag begin to solidify at the bottom of the furnace. Recovering from a chilled hearth can take weeks of downtime and cost millions of dollars in lost production and repairs. Conversely, keeping the furnace overly hot to avoid chilling consumes excessive fuel and increases carbon emissions. Predicting silicon content hours in advance is the key to maintaining this delicate thermal balance.

Historically, mills used two separate methods to predict silicon content: physics-based models and pure data-driven machine learning. Physics-based models use thermodynamic and kinetic equations to simulate the chemical reactions inside the furnace, such as the reduction of silica into silicon. While scientifically sound, these models struggle to keep up with real-time operational changes and non-linear complexities. On the other hand, pure AI models analyze historical sensor data to find patterns. However, because they do not understand the laws of physics, pure AI models can produce physically impossible predictions when encountering unusual furnace conditions.

Physics-informed artificial intelligence solves this problem by merging both approaches. PIAI embeds fundamental metallurgical principles—such as conservation of mass, heat transfer, and chemical equilibrium—directly into the architecture of neural networks.

Steelmakers implement PIAI for silicon prediction and hearth chilling prevention through a structured, real-time process:

1. Multi-Source Data Ingestion The system continuously collects data from hundreds of furnace sensors. This includes top gas analysis (carbon monoxide, carbon dioxide, and hydrogen levels), hot blast parameters (temperature, pressure, oxygen enrichment, and moisture), pulverized coal injection rates, and burden charging weights from the top of the furnace.

2. Applying Physics Constraints Before the AI makes a prediction, the data is processed through physical equations. Mass and energy balance equations calculate the theoretical heat available in the lower zone of the furnace. The model also accounts for the time delay—often several hours—between materials entering the top of the furnace and melting into the hearth.

3. Constrained Neural Network Training The AI model, often a Long Short-Term Memory (LSTM) network designed for time-series data, is trained on historical furnace performance. Crucially, the AI's "loss function"—the mathematical way the AI grades its own accuracy—is modified. The AI is penalized not only for statistical errors in its silicon predictions but also for violating thermodynamic laws. For example, if the AI predicts a rise in silicon when the energy balance equation clearly shows a net heat loss, the model corrects itself.

4. Real-Time Forecasting Once deployed, the PIAI model predicts the silicon content of the hot metal 2 to 6 hours before it is tapped from the furnace. Because the model is anchored by physics, its predictions remain reliable even when raw material quality fluctuates or when the furnace is operating outside of its normal historical parameters.

5. Preventing Hearth Chilling When the PIAI model forecasts a sharp downward trend in silicon content, it triggers an early warning for a cooling hearth. Because the prediction happens hours in advance, operators have enough lead time to intervene. The system can recommend specific adjustments, such as increasing the pulverized coal injection rate, raising the temperature of the hot blast, or reducing the moisture levels in the blast.

By bridging the gap between theoretical metallurgy and data science, physics-informed AI gives blast furnace operators unprecedented visibility into the hearth. This technology successfully prevents hearth chilling, stabilizes iron quality for downstream steelmaking, and significantly reduces fuel consumption in integrated steel mills.