A Guangxi steel company's AI transformation case study

Ask AI · How does the Liugang AI large model optimize energy consumption and costs in steelmaking?

In 2025, a historic event occurred in China’s steel industry—something that hadn’t happened in over twenty years—the amount of steel used in manufacturing exceeded that used in construction for the first time.

The proportion of steel used in construction dropped from 58% in 2020 to 49%, while manufacturing’s share rose to 51%.

Over the past two decades, building skyscrapers, roads, and bridges has been China’s biggest steel buyer. Now, automotive, shipbuilding, new energy, and aerospace are taking over that position.

This isn’t just a change in buyers; the types of steel used in construction—rebar, with few varieties and simple specifications—are driven by volume and price; high-tech manufacturing steel, with many varieties, strict performance standards, and short delivery cycles, emphasizes precision and management capabilities.

The competitive logic of the entire steel industry is changing, and at the same time, the market size is shrinking.

According to data from the National Bureau of Statistics, in 2025, the national crude steel output reached 961 million tons, a year-on-year decrease of 4.4%; apparent consumption of crude steel was 829 million tons, down 7.1% year-on-year.

Simultaneous declines in production and consumption are uncommon in China’s steel industry over the past twenty years.

Li Yiren, Vice President of the China Iron and Steel Association, believes that the industry has entered a new stage of “reduction adjustment, stock optimization, and quality improvement and upgrading.”

Demand is changing, total volume is decreasing, and the steel industry no longer relies on building blast furnaces for growth. Steel companies need to squeeze efficiency from every existing process.

How to improve the pass rate of each furnace? How to reduce waste in each scheduling? Can transportation be more energy-efficient?

In the past, these questions were answered through master craftsmen’s experience and lean management. Now, there is a new tool.

On March 31, in Nanning, the “Guangxi Traditional Manufacturing Artificial Intelligence Innovation Application and Liugang Large Model Release Conference” was held. Guangxi Liuzhou Steel Group Co., Ltd. (hereinafter “Liugang Group”) jointly released Guangxi’s first steel industry large model—the “Xuan Tie” Steel Large Model—with Huawei and China Mobile Guangxi, and simultaneously established the “Guangxi Liugang Artificial Intelligence Research and Innovation Center.”

Liugang Group is a steel enterprise with annual revenue exceeding 27k yuan, a top 500 Chinese enterprise, Vice President unit of the China Iron and Steel Industry Association, and has been rated A+ in competitiveness among Chinese steel companies for many years.

This veteran steel company employs 27k people, with over 250k sets of equipment operating daily, producing more than 1,000 types of steel products.

According to data disclosed at the release, the “Xuan Tie” Steel Large Model now covers multiple core links from ironmaking to rolling. Backed by the “Xuan Tie” model’s empowerment, just in the full-process intelligent steelmaking, Liugang’s crude steel production cost can be reduced by nearly 100 million yuan annually.

When output no longer grows, cost reduction becomes growth.

The next story is about how this Guangxi steel company uses AI to find space for quality improvement and efficiency in each process.

After 961 million tons

In 2025, the nationwide crude steel output of 961 million tons marks the second consecutive year of decline.

But reaching a production peak is only superficial; deeper changes are happening downstream. The steel used in real estate continues to shrink, while manufacturing fields like automotive, shipbuilding, and new energy demand higher varieties and performance of steel.

Meanwhile, steel industry carbon emissions account for over 15% of the country’s total, and the pressure to reduce emissions is forcing changes in production methods.

According to the China Iron and Steel Industry Association, over 95% of steel companies have incorporated digital transformation into their overall development strategy.

This high proportion is partly because steel production processes are inherently complex—many steps, numerous parameters, sensitive to precision—leaving less room for manual management optimization.

From iron ore to finished steel, it’s a long process.

Iron ore is first smelted into pig iron in a blast furnace, then transferred to a converter to be refined into steel. The steel undergoes refining to adjust composition and temperature, then continuous casting into billets, and finally rolling into various specifications of steel plates, coils, or profiles.

This process involves dozens of steps, each with many parameters needing real-time control—temperature differences of a few degrees, alloy additions of a few grams—affecting steel quality.

Saving a little energy here, increasing yield there—multiplied by millions of tons of annual production—adds up to significant savings.

This makes steel production naturally suitable for AI intervention, given the many steps, parameters, large data volume, and sensitivity to accuracy.

Liugang Group was founded in 1958 as a small steel plant with only a few hundred employees.

Over more than sixty years, it has grown into one of Guangxi’s largest manufacturing enterprises, ranking among the top 50 global steel companies.

Liugang’s products have supplied major projects like the Hong Kong-Zhuhai-Macau Bridge, China’s “Sky Eye,” and the Pinglu Canal, and also supply steel for BYD, BAIC, Midea, and other companies.

Liang Bin, Secretary of the Party Committee and Chairman of Liugang Group, mentioned at the release that the “14th Five-Year Plan” was the most challenging period in China’s steel industry development history. During this low point, Liugang completed product restructuring and equipment upgrades, achieving profitability by 2025.

However, despite profitability in the industry downturn, the space for traditional management methods to improve efficiency is limited.

Shen Min, Vice General Manager of Liugang Group, introduced at the release that over 90% of Liugang’s production lines have now been automated; traditional automation has been well established.

In recent years, Liugang has also implemented some AI applications, but they are fragmented—each workshop deploying its own model, with no data sharing between workshops, and algorithms managed separately.

For example, the steelmaking workshop has its own model, and the rolling workshop has another, but there’s no coordination between them, making production scheduling and quality control disconnected.

This is the starting point for Liugang’s joint development of the “Xuan Tie” Steel Large Model with Huawei—to integrate scattered AI capabilities into a unified platform.

The “Xuan Tie” Steel Large Model is based on Huawei’s Pangu large model.

First, it is trained with steel industry data to learn the basic laws of steel production; then, it is fine-tuned with Liugang’s years of accumulated production data to adapt to Liugang’s specific conditions; finally, lightweight sub-models are generated for each specific process scenario to execute.

In simple terms, one “brain” manages overall coordination, while a group of “small brains” handle individual processes.

This system covers six links: ironmaking, steelmaking, rolling, logistics, environmental protection, and safety, with over 20 application scenarios planned—20 are already mature and clear, and N are new scenarios expanding based on actual production.

In terms of computing power, the entire system uses a fully domestically developed stack, including Huawei’s Kunpeng processors and Ascend AI chips.

Among the first batch of intelligent applications released with the “Xuan Tie” model, two key applications—LF refining optimization and intelligent slab grouping—have been deployed and put into use at the Fangchenggang base.

How does AI reach the workshop?

Liugang Fangchenggang base is about 200 km from Nanning, adjacent to Beibu Gulf.

This is Liugang’s coastal production base, with a U-shaped layout: raw materials arrive at one end, go through ironmaking, steelmaking, and rolling in sequence, and finished products exit at the other end, with no backtracking, minimizing transportation distance.

Multiple AI applications based on the “Xuan Tie” model are deployed across various workshops in this base.

Following the steel production process, the first step is ironwater transportation. After the blast furnace produces pig iron, it must be quickly sent to the converter for steelmaking.

Fangchenggang has 10 locomotives and 51 ironwater tankers, each transporting about 3,600 tons of molten iron daily. The molten iron temperature exceeds 1,400°C, and the journey from blast furnace to converter covers about 2 km, passing through three railway crossings, taking roughly 28 minutes.

Previously, this transportation segment was one of the most labor-dependent parts of the base.

Liao Liuqiang, director of the iron transport workshop at Fangchenggang port, explained that there were four main issues: over 70 workers on shifts including drivers, shunters, and dispatchers; communication relied on walkie-talkies, often with delays; open-top tankers caused continuous temperature drop—by the time it reached the converter, temperature was insufficient, requiring extra heating and increasing costs; hook-up and parking operations were manual, with high safety risks near 1,400°C molten iron.

Starting in 2020, Liugang partnered with Huawei and China Mobile Guangxi to begin intelligent upgrades of this transportation line.

Moke Yi, the train driver team leader, experienced the entire transformation.

He explained that the main steps were threefold:

First, locate the ironwater tanks by installing positioning devices, so dispatchers can see real-time locations on a screen, eliminating the need for back-and-forth radio communication.

Second, maintain water temperature by adding automatic insulation covers, reducing temperature loss by 35°C.

Third, achieve driverless operation: all 10 locomotives are equipped with autonomous driving systems, and with an intelligent dispatch system that updates transportation plans every minute, the locomotives automatically handle hooking, transporting, and unhooking.

Moke Yi said that in the past, drivers worked 24-hour shifts; now, the locomotives run themselves, and drivers have shifted to monitoring in the control room.

Liao Liuqiang added that after the intelligent transportation system went live, the turnover rate of water tanks increased from 3.5 to 4.5 tanks per day—about a 30% improvement. The temperature drop was reduced by 35°C, meaning more scrap steel could be added per ton of steelmaking, directly lowering raw material costs.

The original team of over 70 people now only needs about 10 in the control room for monitoring and maintenance. The entire system generates about 35 million yuan in annual benefits.

After the molten iron reaches the converter, the next step is steelmaking.

The core of converter steelmaking is blowing oxygen into the molten iron to remove excess carbon and impurities.

This process requires real-time judgment of furnace temperature and carbon content to decide when to add materials, when to stop blowing, and when to tap the steel.

Traditionally, operators “watch the fire”: standing in front of the furnace, wearing sunglasses, staring at the flames.

White, glaring flames indicate high temperature and nearly complete carbon burn; red, dull flames suggest insufficient temperature; long, forked sparks indicate high carbon content; short, round, or disappearing sparks mean carbon removal is nearly complete.

An experienced furnace operator can judge within seconds by observing the flames.

But this skill heavily depends on personal experience, making it hard to standardize and pass down, and influenced by physical condition—fatigue on night shifts reduces accuracy.

Lu Gang’s intelligent manufacturing expert Luo Chunhong explained that now, Liugang’s AI fire-watching system uses high-definition cameras installed at the furnace mouth to capture real-time images of flames, analyzing color, shape, brightness with image recognition algorithms, operating 24/7 without fatigue.

Luo also noted a change: in the past, dispatchers had to manually collect data from sintering, blast furnace, and rolling workshops; now, at the Fangchenggang intelligent manufacturing center, a large display shows the entire base’s production status.

After the steel is tapped from the converter, it goes into the refining furnace (LF furnace) for final temperature and composition adjustments.

Refining is complex due to many variables: Liugang has hundreds of steel grades, each with different composition requirements and process paths.

Previously, refining relied mainly on manual judgment—testing steel composition, then deciding how much alloy to add and how long to blow argon, based on experience. Results came after the fact, often as verification rather than control.

Liugang and Huawei jointly developed an LF refining optimization scheme that models the fundamental physical and chemical laws of steel, using AI prediction models to forecast temperature and composition changes based on real-time data, and solvers to determine optimal operation parameters.

This integration shifts from “post-verification” to “process control.”

The scheme has shortened refining time by 2 minutes and reduced alloy addition costs by 2 yuan per ton.

Overall, Liugang has deployed 33 AI models covering converter blowing, LF refining, ladle argon blowing, and other processes.

These models have increased converter productivity by 8.5%.

Once the steel is refined, it is cast into billets and enters the rolling process.

The billets are rolled into various specifications of steel plates or coils. In brief, rolling involves two steps: hot rolling at over 1,000°C to roughly shape and size the steel; then cold rolling at room temperature for finer thickness and surface quality.

The working environments are entirely different. Visiting Fangchenggang’s cold rolling mill, the first impression is cleanliness: the floor is coated with green epoxy, and rows of silver coils are stored in the yard, with blue projection markings indicating “Unmanned Crane Area.” Looking up, a yellow overhead crane moves along tracks, automatically gripping a coil and slowly transporting it to the designated spot—all without operators.

Cold rolling expert Su Hui explained that the plant has a “cloud vision” heavy-duty gantry robot jointly built by Liugang, Huawei, and China Mobile Guangxi.

Its job is to lift rolling mills from storage to production lines. Previously, this required three workers on-site, working in high-temperature, dusty environments.

In 2020, Liugang first upgraded this with 5G remote control.

But remote control still required human supervision, so the team continued to innovate, equipping the robot with stereo vision and AI algorithms.

Su Hui said that during development, the team and Huawei iterated over 10k lines of code, processed over 1 million images, ultimately increasing the grab success rate from 85% to 99.5%, with accuracy within 15 millimeters. After the upgrade, personnel costs for this job dropped by 83%.

At the entrance of the galvanizing line in the cold rolling plant, there’s also a dual-line gantry intelligent bundling robot.

Before steel coils are put online, they need to be unbundled. Previously, this was manual work.

Now, this robot is equipped with millimeter-level laser positioning, with positioning errors less than 0.5 mm, and a success rate of 99.5%. It can automatically unbundle 60 coils per shift, covering four production lines.

Liugang’s cold rolling plant produces thin sheets and galvanized sheets. The Guangxi Fangchenggang base also has another key production line—the 3,800 mm wide heavy plate line, which started operation in November 2024, covering shipbuilding, marine engineering, wind power, and other fields.

A characteristic of heavy plate orders, called “Chinese medicine orders” within the industry, is that they are diverse, with complex specifications and small batch sizes. Each order has different requirements for length, width, and thickness.

Chen Peizhi, head of the model technology for the heavy plate line at the main hot rolling plant, explained that previously, engineers had to manually match customer contracts with billets in the system, figuring out which orders could share a billet and how to arrange cuts to minimize waste—processing thousands of contracts could take hours and still struggle to find the global optimal solution.

Now, Liugang has developed an intelligent slab grouping and billet assembly system based on Huawei’s Tianchou solver.

The system converts human experience into calculable rules and models, considering delivery deadlines, production line load, energy consumption, and other goals, completing the grouping plan for over 1,000 contracts within minutes.

Chen said that over 90% of contracts can now be automatically matched, with yield and utilization rates improved by more than 1%.

A 1% increase may seem small, but for a line producing over 250k tons annually, that 1% saves more than 20,000 tons of steel each year.

Next three years

Li Bian announced at the March 31 release that the company plans to invest over 3 billion yuan in the next three years to advance the “Digital and Intelligent Liugang” project.

For a steel company, 3 billion yuan is a significant investment, reflecting management’s clear expectations of AI-driven cost reduction returns.

Where will this 3 billion be spent?

According to Li Bian, part will go toward infrastructure for computing power and large model platforms; part toward developing and deploying specific scenarios; and part toward personnel.

Liugang has launched a “Ten Thousand AI Employees” plan, which has two layers: first, enabling real employees to develop AI skills. So far, Liugang employees have independently built 2,082 AI assistants.

For example, an employee at the Fangchenggang sales center, with no prior IT development background, used large model capabilities combined with the company’s sales knowledge base to create an intelligent quotation assistant, solving the low efficiency and error-prone manual quoting process.

The second layer involves deploying intelligent agents across production stages, allowing AI to participate as “digital employees” in scheduling, dispatching, quality inspection, and more.

Liugang’s goal is to achieve over 80% full-process automation within three years, build more than 10 industrial intelligent agents at the production line level, develop over 30 high-quality industrial data sets, and implement 20 more benchmark AI scenarios by 2027.

Liugang is not the only steel enterprise doing this.

Li Yiren mentioned at the release that just a week prior, the China Iron and Steel Industry Association held a closed-door seminar on “AI + Steel Industry” development in Nanjing, with leading companies like Baowu, Ansteel, Shougang, Hebei Steel, and Nanjing Steel participating.

Currently, the association is organizing Liugang, Baowu, Ansteel, Nanjing Steel, and others to compile the “AI + Steel Industry Implementation Guide” at the request of the Ministry of Industry and Information Technology.

Huawei is also increasing its investment in the steel industry.

Huawei’s Steel, Nonferrous, and Military Industry Group was established in 2025, and has already cooperated with over 300 companies worldwide, including more than 120 steel and nonferrous metal enterprises, developing over 100 AI scenarios covering the entire smelting process.

Huawei Vice President Jiang Wangcheng introduced other industry deployments: Baowu relies on the Pangu large model for blast furnace temperature prediction, reducing costs by 5-10 yuan per ton of pig iron; Conch Cement uses predictive models for overall energy consumption optimization, decreasing standard coal consumption by 1%, and reducing annual carbon emissions by over 4,500 tons per line.

Huawei’s Steel and Nonferrous Military Industry CEO Shi Miao stated that, amid rapid AI technological evolution, the “Xuan Tie” Steel Large Model architecture and industry partner development system provide certainty to address uncertainties in technological progress.

For Liugang, the efficiency gains from AI also have an extra significance: the company is accelerating its outward expansion.

Guangxi is China’s only province directly connected to ASEAN countries via land and sea. Relying on the coastal advantage of the Fangchenggang base, Liugang’s exports have grown rapidly in recent years, expanding from traditional bars to heavy plates, ship plates, automotive steel, and other varieties, with markets covering ASEAN, the Middle East, and South America.

The 3,800 mm wide heavy plate line, scheduled to start in November 2024, fills a gap in Guangxi’s heavy plate production. Its products have obtained major global classification society certifications and EU CE certification.

Liang Lei, Deputy Secretary-General of the Guangxi Zhuang Autonomous Region Government, stated at the release that Guangxi is promoting a development path of “R&D in Beijing, Shanghai, Guangzhou + integration in Guangxi + ASEAN application,” with cutting-edge technology developed in major cities, integrated and implemented locally in Guangxi, and then exported to ASEAN countries.

Currently, over 100 leading AI companies are based in Nanning, with the core industrial output value of AI in the industrial field surpassing 89 billion yuan.

At the “Xuan Tie” Steel Large Model release event, Liugang and Huawei, along with China Mobile Guangxi, signed a deepening cooperation agreement.

According to the plan, by 2027, the application scenarios of the “Xuan Tie” Steel Large Model will cover the main aspects of Liugang’s production and operation.

In the future, Liugang aims to, together with Huawei and other partners, leverage AI capabilities like the “Xuan Tie” model to seize opportunities and gradually develop into a nationally leading, internationally competitive smart factory with global influence.

(Author: Feng Kehan)