Baowu's Innovative Production: AI Models Unlock the "Black Box" of Blast Furnaces

In the traditional steel industry, blast furnaces are the core of production. However, this giant reactor that creates such immense value has traditionally been like a "black box": the internal physical and chemical reactions are extremely complex, and difficult to observe or measure directly.

Since 2024, China Baowu Steel Group Corp., Ltd., (commonly known as Baowu), the world's largest steel company, has partnered with Huawei to use AI models to solve the centuries-old challenge of the "black box" in blast furnaces. This breakthrough not only marks the beginning of a paradigm shift towards intelligent traditional industries. It also brings great practical benefits: a single blast furnace creates more than CNY10 million (≈USD1.4 million) in annual benefits, setting an example for the global heavy industry to switch from experience-based operations to data-driven ones.

The dilemma of the blast furnace

Baowu produces 130 million tons of steel annually. In 2024, it ranked 44th in the Fortune Global 500, with its crude steel output ranking first in the world. Within its vast production system, the blast furnace accounts for about 70% of the total production cost. Therefore, the long-term stable operation of the blast furnace directly affects the company's profitability.

However, this furnace is actually the most uncertain and challenging part of the entire production process. The main challenges include the following four points:

Unclear furnace status: The flames in the blast furnace block external detection. When the temperature inside the furnace exceeds 2000°C, the solid, liquid, and gas states are intertwined, involving more than 5000 data dimensions. Historically, the judgment and control of the blast furnace status cannot be directly monitored. Only manual experience and indirect observation can be relied on.

Lagging operation feedback: People can only analyze limited parameters, and feedback on furnace conditions after operational adjustments can take hours. By the time issues with furnace conditions arise, the quality of the molten iron has already fluctuated. This makes it hard to respond accurately and proactively adjust.

Uncontrollable chain reactions: There are hundreds of factors that affect the furnace condition and they are tightly coupled. The composition of raw materials entering the furnace, the distribution of airflow in the furnace, and slight changes in temperature can all trigger chain reactions. Even for a top factory like Shanghai Baosteel, it is still a world-class challenge to keep a 5000-cubic-meter giant blast furnace running stably for a long time. Once the furnace condition becomes abnormal, the recovery period is often measured in days, with daily losses reaching millions of CNY for each blast furnace.

Challenges in inheriting experience: Expert knowledge relies on the long-term gathering of practical experience, making it difficult to fully quantify and replicate. As the expert team gets old and technology shifts, the passing on of experience is restricted.

Breaking through the bottleneck with an AI model for blast furnaces

To tackle this big challenge, Baowu and Huawei have established a deep cooperation model of "digital-physical world integration and joint problem solving." The two parties have formed a joint team: Baowu selects application scenarios and contributes industry knowledge based on business challenges, while Huawei leverages its full-stack technical capabilities in AI, big data, cloud computing, and other fields.

The core solution is to build an AI model for the blast furnace. Simply put, this model acts as a "brain" for the blast furnace: it transforms expert knowledge into data code (digitization), converts invisible reactions inside the furnace into predictable parameters (transparency), and ultimately achieves precise prediction and automatic control.

First, the "black box" can be observed. The joint team uses Huawei's Pangu model as the foundation, which first absorbs general knowledge from thousands of industries to lay a solid foundation, and then customizes the model based on the characteristics of blast furnace ironmaking. The AI model not only learns from Baowu's vast operational data, but also draws on universal laws of physics and chemistry in other industrial fields, showcasing a comprehensive approach. Ultimately, the model achieves a prediction accuracy of 90% for core key indicators like furnace temperature, marking the first time high-precision and real-time observation of the "black box" internal state has been achieved.

Second, experience can be replicated at scale. Baowu operates dozens of blast furnaces, each with different structures and process configurations. Developing a model from scratch for each furnace would be costly and time-consuming. To address this, the joint adopted a "pre-training base + downstream task fine-tuning" approach based on Huawei Cloud Stack. That is, the joint team fine-tuned the basic model according to the specific features of different blast furnaces, greatly shortening the rollout period.

Third, the AI application value is not limited to accurate prediction. The core capability of the model is to form a closed-loop system of continuous learning and self-optimization, which it does through the data flywheel of incremental training, prediction inference, and closed-loop control. It provides operation suggestions based on the prediction result, verifies the effect in production practices, and feeds back new data to the model for retraining, so that the capability of the model is continuously enhanced. This mechanism fundamentally transforms the past reliance on manual, reactive operational modes.

From millions in benefits to an industrial blueprint

After the efforts of the joint team, the blast furnace model has been running successfully at the production base of Baosteel for more than 10 months. Calculations show that applying this model to a single blast furnace can result in annual benefits exceeding tens of millions of CNY, stemming from reduced fuel consumption, stable iron quality, and the overall effectiveness of fewer abnormal furnace conditions.

The significance of this exploration goes beyond just for blast furnaces: The two parties are planning a steel model capability map, applying AI capabilities of prediction, vision and scientific computing onto the entire manufacturing process. Examples of parts of this process are raw materials, iron smelting, steelmaking, steel rolling, and new material development. The map covers hundreds of application scenarios such as continuous casting quality root cause analysis, hot-rolled plate type prediction, and steel surface quality inspection.

This means that the Chinese steel industry is redefining the production paradigm with AI—from experience-driven to data-driven, from passive response to proactive prediction. AI foundation models are key tools for achieving breakthrough optimizations in complex, core production processes. This integration of industrial mechanisms, expert experience, and AI technologies also provides a Chinese sample for the intelligent transformation of global heavy industries.