⚛️ quantum physics

Predictive control of blast furnace temperature in steelmaking with hybrid depth-infused quantum neural networks

This paper proposes a hybrid depth-infused quantum neural network approach that integrates quantum-enhanced feature exploration with classical regression to significantly improve blast furnace temperature prediction accuracy by over 25% and stabilize temperature control within a ±7.6-degree range, thereby optimizing steel production efficiency.

Original authors: Nayoung Lee, Minsoo Shin, Asel Sagingalieva, Arsenii Senokosov, Matvei Anoshin, Ayush Joshi Tripathi, Karan Pinto, Alexey Melnikov

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nayoung Lee, Minsoo Shin, Asel Sagingalieva, Arsenii Senokosov, Matvei Anoshin, Ayush Joshi Tripathi, Karan Pinto, Alexey Melnikov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a Blast Furnace as a giant, roaring, 24-hour-a-day kitchen oven that never sleeps. Its job is to melt iron ore into liquid iron, which is then turned into steel. This is the backbone of the global economy, but it's also a chaotic environment.

The problem? You can't see inside the oven. It's too hot, too dark, and too dangerous for humans or cameras. The only way to know what's happening inside is to look at the "exhaust" (the gas coming out) and the "soup" (the molten iron) when it finally pours out.

The biggest challenge for the chefs (the steelmakers) is keeping the temperature just right.

Too cold: The iron doesn't melt, and the oven gets clogged (like a blocked drain).
Too hot: The equipment melts, or the iron solidifies in the wrong places, causing dangerous explosions.

Currently, the chefs have to guess. They inject coal (like adding more fuel to a fire) to adjust the heat, but there's a 3-hour delay between adding the coal and seeing the temperature change. It's like trying to steer a massive ship by turning the wheel, but the ship doesn't respond until 3 hours later. Because of this delay, the temperature swings wildly, sometimes by 50 degrees above or below the target.

The Solution: A "Quantum Crystal Ball"

The authors of this paper built a new kind of "crystal ball" to predict the future temperature of the furnace so they can adjust the coal before the problem happens.

Here is how they did it, using a mix of old-school math and futuristic "Quantum" technology:

1. The Detective Work (Data Cleaning)

First, they looked at data from 580 different sensors (like thermometers, pressure gauges, and gas analyzers) from a real steel plant. It was a messy pile of information.

The Analogy: Imagine trying to find a needle in a haystack, but the haystack is made of 580 different types of hay.
The Fix: They used a smart filter (Gradient Boosting) to find the top 27 most important clues. They realized that things like the temperature of the furnace walls and the chemical makeup of the slag (waste rock) were actually better predictors than they thought.

2. The Brain: Hybrid Quantum Neural Networks

They built an AI brain to learn from these clues. But instead of just a normal computer brain, they gave it a Quantum upgrade.

The Classical Part (LSTM): Think of this as a very good student who is great at remembering a story. It looks at the last 24 hours of data (the "story") and tries to guess what happens next.
The Quantum Part (QDI Layer): This is the magic ingredient. Imagine the classical student is trying to solve a puzzle in a small room. The Quantum layer is like giving that student a hall of mirrors. It allows the AI to look at the data from many different angles simultaneously (using a concept called superposition).
The Result: The "Hybrid" brain can see complex patterns that the normal brain misses. It's like the difference between a human trying to count stars with a flashlight vs. a satellite seeing the whole galaxy at once.

3. The Pilot (Optimization)

Once the AI can predict the temperature accurately, it needs to decide what to do.

The Goal: Keep the temperature in a tiny, safe zone (between 1500°C and 1510°C).
The Action: The AI simulates thousands of scenarios in seconds: "If I add 5 tons of coal now, what happens in 3 hours? If I add 10 tons?"
The Outcome: It finds the perfect amount of coal to inject to keep the temperature steady.

The Results: From Chaos to Calm

The results were dramatic:

Prediction Accuracy: The new Quantum-Hybrid AI was 25% more accurate at predicting the temperature than the old computer models.
Temperature Stability: This is the big win. Before, the temperature swung wildly by ±50 degrees. After using this new system, the temperature stayed steady within ±7.6 degrees.

Why does this matter?

Safety: No more dangerous "hang-ups" or explosions caused by temperature spikes.
Money: Because the temperature is so stable, the steelmakers don't need to keep the furnace "extra hot" just to be safe. They can lower the target temperature, saving massive amounts of coal (which costs a lot of money).
Quality: The steel produced is more consistent and higher quality.

The Bottom Line

This paper proves that Quantum Computing isn't just a sci-fi dream for the future. Even with today's early, imperfect quantum computers (simulated on regular computers for this study), mixing them with standard AI creates a super-tool that can solve real-world, messy industrial problems.

It's like giving a human chef a pair of X-ray glasses and a time machine, allowing them to see inside the oven and adjust the fire before the food even starts to burn.

1. Problem Statement

The steelmaking industry relies heavily on the blast furnace (BF) process, where maintaining a precise and stable internal temperature is critical for efficiency, safety, and product quality. However, controlling BF temperature is notoriously difficult due to:

Non-linearity and Complexity: The internal chemical and physical processes are highly non-linear and dynamic.
Time Delays: There is a significant lag (2–3 hours) between control actions (specifically Pulverized Coal Injection (PCI)) and their effect on molten iron temperature.
Inaccessibility: The extreme environment prevents direct internal measurement; operators rely on external sensors and delayed outputs.
Instability: Current control strategies often result in large temperature variances (historically ±50°C from the target), leading to risks such as furnace blockages ("hang-ups"), solidification ("back lamination"), and excessive fuel consumption.

The core challenge is to predict future temperature fluctuations accurately and optimize PCI injection rates in real-time to maintain the temperature within a narrow target window (e.g., 1500°C–1510°C).

2. Methodology

The authors propose a Hybrid Quantum Machine Learning (HQML) framework that integrates classical deep learning with quantum computing principles to solve the prediction and optimization tasks.

A. Data Preprocessing and Feature Engineering

Dataset: One year of operational data from POSCO's Gwangyang Steel Works (2023), recorded every minute from ~580 sensors.
Cleaning: Applied outlier correction using an adjusted Inter Quartile Range (IQR) of 5 IQR (instead of the standard 1.5) to accommodate the specific anomaly characteristics of blast furnace data.
Feature Selection: Used Gradient Boosting (GB) models to determine feature importance for both temperature and PCI rate prediction. The top 25 features were selected (combining the top 19 for temperature and top 6 for PCI), including variables like tuyere velocity, gas utilization, wall temperatures, and slag composition.
Discretization: Time-series data was converted into fixed windows (24 time steps) using moving averages to reduce noise and align with the optimization horizon.

B. Hybrid Quantum Neural Network (HQNN) Architecture

The core innovation is a Hybrid Quantum Neural Network designed for temperature forecasting:

Classical Backbone: A Long Short-Term Memory (LSTM) network processes the temporal sequences of the 24 input time steps (27 features per step).
Quantum Depth-Infused (QDI) Layer: A variational quantum circuit is inserted between two fully connected layers.
- Mechanism: It encodes classical feature vectors into a high-dimensional Hilbert space using parameterized rotation gates ( $R_x, R_y$ ) and entanglement gates (CNOT).
- Configuration: The model utilizes 6 qubits with a circuit depth of 4.
- Purpose: To exploit quantum feature space exploration to capture complex, non-linear correlations that classical layers might miss, acting as a regularizer to prevent overfitting.
Output: The model predicts future temperature vectors (5 time steps ahead).

C. Optimization Framework

A two-stage model approach is used for control:

$M_{all}$ (Multi-feature Predictor): A model (similar to the HQNN but purely classical in the current implementation due to resource constraints) predicts the evolution of all 27 sensor features, including temperature, based on current states.
$M_{optim}$ (Optimization Model): A linear model that searches for the optimal sequence of PCI values.
- Process: The system simulates candidate PCI policies using $M_{all}$ , calculates the resulting temperature deviation from the target (1510°C), and uses gradient descent to update $M_{optim}$ to minimize this deviation while keeping PCI within physical bounds [0, 1].

3. Key Contributions

Novel Hybrid Architecture: Introduction of a Depth-Infused Quantum Neural Network specifically tailored for industrial time-series forecasting, demonstrating that quantum layers can enhance classical LSTMs even on simulated hardware.
End-to-End Control Pipeline: A complete framework linking high-accuracy prediction to actionable control policies (PCI optimization), moving beyond mere forecasting to active process stabilization.
Industrial Validation: Successful application of Quantum Machine Learning (QML) on real-world, proprietary industrial data from a major steel producer (POSCO), bridging the gap between theoretical QML and practical manufacturing.
Feature Discovery: Identification of previously underutilized high-value features (e.g., specific slag components like $TiO_2$ , $MnO$, $S$ ) through gradient boosting analysis.

4. Results

The study reports significant improvements in both prediction accuracy and process stability:

Prediction Accuracy:
- The Hybrid Quantum LSTM outperformed the purely classical LSTM.
- For a 1-hour ahead prediction, the Hybrid model achieved an RMSE of 4.46, compared to 5.95 for the classical model.
- This represents a ~25% reduction in prediction error.
- The hybrid model also achieved better generalization with fewer trainable parameters, suggesting the quantum layer acts as an effective regularizer.
Temperature Control (Stability):
- The optimized control policy reduced temperature variance from the historical ±50°C to ±7.6°C around the target range.
- This level of stability allows operators to lower the target temperature setpoint (reducing fuel usage) without risking furnace instability.

5. Significance and Implications

Economic Impact: Reducing temperature variance allows for lower fuel consumption. Given the injection of ~110 tons/hr of coal per furnace, even marginal efficiency gains translate to massive cost savings (estimated at 30–60 thousand KRW per ton of coal).
Safety and Reliability: Precise temperature control mitigates the risk of catastrophic events like "hang-ups" and "back lamination," enhancing worker safety and equipment lifespan.
Feasibility of QML in Industry: The paper demonstrates that NISQ-era (Noisy Intermediate-Scale Quantum) technologies, even when simulated on classical hardware, can provide tangible performance boosts in complex industrial settings. It establishes a roadmap for deploying true quantum hardware as it matures.
Scalability: The framework is generalizable to other industrial control problems involving non-linear dynamics and time delays.

Conclusion

This work validates the potential of Hybrid Quantum Machine Learning as a superior tool for predictive control in heavy industry. By combining the temporal processing power of LSTMs with the feature-mapping capabilities of variational quantum circuits, the authors achieved a breakthrough in blast furnace temperature stability, offering a practical pathway for the adoption of quantum-enhanced AI in real-world manufacturing.