AI in Steel: Why Scheduling Beats Visual Inspection as First Win

Introduction: The Heaviest Industry for AI

Steel and non-ferrous metals are the most energy-intensive, data-fragmented, and latest process industries to be truly touched by AI. Unlike automotive or semiconductors, steelmaking is constrained by a 1,600°C furnace interior that cannot be seen, dozens of tightly coupled process steps, and tens of thousands of sensors with inconsistent data formats. The core question is not whether AI can make steel, but where AI lands first, in what sequence, and what it actually saves.

Four Mountains Weighing on Steel

Four fundamental barriers explain why steel's intelligent transformation lags behind other sectors. These mountains are also the entry points AI eventually breaks through:

Energy & Carbon Mountain: The blast furnace accounts for over 60% of manufacturing cost and carbon emissions; energy fluctuations mean direct financial loss.

Data Fragmentation Mountain: Sensors and level-2 models are weak links; multi-source heterogeneous data lacks unified formats.

Logistics & Inventory Mountain: Major costs often lie outside the production line; molten iron transport and raw-material inventory have long relied on experience.

Mechanism Black-Box Mountain: Reactions inside the 1,600°C furnace are invisible; operations have long depended on veteran workers' tacit experience.

A telling comparison: the blast furnace is the core process, consuming 60%+ of cost and emissions. Shougang Qian'an used to transport molten iron from ironmaking to steelmaking in torpedo ladles taking over 40 minutes ; after deploying an automated molten-iron allocation and scheduling large model, the time dropped to under 15 minutes , saving 70 million RMB annually and cutting 40,200 tonnes of CO₂ . The first mountain AI moved was the one everyone overlooked — scheduling.

Core Concept: The Three-Layer Cake of Steel AI

Steel AI applications can be sliced by proximity to process control into three layers. Understanding this cake explains why some mills only add cameras while others dare call themselves "steel large models" or "full-domain intelligent agents."

Bottom Layer: Data, Compute & Industrial OS

This least "intelligent" but most critical layer ingests massive sensor data and connects heterogeneous systems. China Baowu launched the "2526" project in February 2025, fully deploying DeepSeek and building the domestic steel industry's largest, most advanced AI compute center. Hebei Province concentrated resources to build a "Steel Vertical Large Model Application Service Platform" connecting 31 major steel enterprises covering over 85% of provincial capacity, letting SMEs call scene-validated professional models at low cost. Without this layer, upper layers starve.

Middle Layer: Steel Large Models & Intelligent Agents

This is the true brain of "Steel + AI": it coordinates mechanism models, vision models, and knowledge graphs for cross-process global optimization. HBIS's "WiseBao" is the nation's first vertical steel large model, using a "large model + small models + agents" system to achieve a full-process "perceive–plan–execute–optimize" closed loop. Nanjing Iron & Steel (Nangang) with Huawei released the "Yuanye Steel Large Model" with a dual-driven architecture: "left-brain precise solving, right-brain probabilistic perception." China Steel Research (CISRI) with Baidu Cloud launched the first "Metallurgical Process Perception Large Model." This layer's hallmark is "global coordination" — pulling previously siloed process optimizations into a single collaborative net.

Top Layer: Scenario Execution

Blast furnace control, surface defect inspection, production scheduling, energy efficiency, equipment predictive maintenance — each completes a domain-specific "perceive–decide–execute" closed loop. Baosteel's blast furnace AI based on Huawei's Pangu model moved to No.3 blast furnace for batch application in August 2025. Handan Steel's surface defect vision system achieves 100% detection for edge cracks, longitudinal cracks, etc. Yangchun New Steel's equipment intelligent control system reaches >98% full-process data collection rate. An iron rule governs this layer: AI sits atop mechanism models, only closing loops, never making final calls — temperature prediction can be modeled, but "how much charge to add, whether to tap" remains with mechanism models and on-site engineers.

The Closed Loop: Mechanism Models Are Forever the Foundation

Sensors continuously collect furnace temperature, composition, vibration, current, etc.; models predict and infer; feedback goes to process control forming a loop; new data flows back for retraining. Baosteel calls this "incremental training – prediction inference – closed-loop control" self-optimizing system. But the foundation is always the mechanism model trained on decades of metallurgical theory and expert experience: AI can "learn while using" but cannot turn physicochemical laws into a black box. Whoever solidifies both "mechanism + data" layers runs a stable loop.

Path: Testing Steel with the "AI+Industry Five Questions"

The entire 20-episode series uses the same ruler — "AI+Industry Five Questions" : Where does data come from? Who owns decision rights? How is responsibility assigned? What are the measurement units? Where are the boundaries? Applied to steel, it instantly separates "real closed-loop cost reduction" from "showroom demos."

1. Data Source: Sensors (temperature/composition/vibration/current), process history databases, equipment ledgers, and provincial shared platforms aggregating peer data. Steel's data is scattered and messy — the biggest barrier. Hebei's public service platform connecting 31 mills covering 85% capacity solved "model selection difficulty, high usage cost, slow deployment" together.

2. Decision Rights: AI gives operation suggestions and closed-loop control, but final process parameter adjustments are confirmed by engineers in the loop. Baosteel's blast furnace AI uses "prediction + control dual-drive" with model adoption rate >90%, yet "charging and tapping" authority stays with humans and mechanism models.

3. Responsibility: Steel is continuous production, safety-sensitive; model errors mean breakouts, scrap, shutdowns. Hence AI sits "above sign-off" — only closes loops, never decides; Yongyang Special Steel codifies veteran experience into "digital assets" but anomalies still require human confirmation.

4. Measurement Units: Look at fuel consumption per tonne of steel (kg), Grade-1 molten iron rate (80%→99%), defect detection rate (98.5%/100%), scheduling time (48h→30min), annual cost reduction (millions/tens of millions), not isolated "model accuracy."

5. Boundaries: Mechanism models are the bottom line; AI does not replace physical laws. Visual inspection suits "data ready, standards unified" steps; process optimization suits "experience codifiable" steps. Forcing "end-to-end black-box steelmaking" is neither compliant nor economical.

Benchmark Players: Where They Sit and How Strong the Signal

Six standard-bearers mapped to layers with quantified outcomes:

Baowu/Baosteel Blast Furnace AI (Top Layer): Furnace temperature prediction hit rate 95%, silicon content 92%, 2-hour ahead 90%+; fuel per tonne iron -2kg; hot-rolling surface inspection 96%. Source: Baowu 2526 / Xinhua.

Nangang Yuanye Steel Large Model (Middle Layer): Coal blending 1-2 days → 1-2 minutes; coke cost per tonne -5~10 RMB, annual saving 7.5M; Grade-1 molten iron rate 80%→99%; unit output energy consumption -12%. Source: China Metallurgical News / Baidu Baike.

HBIS WiseBao + Tangsteel (Middle + Top): Order scheduling 48h→30min; inventory turnover -50%, on-time delivery 100%; Shougang Qian'an molten iron transport 40min→15min saving 70M. Source: Hebei Daily / MIIT.

Yangchun New Steel (Yangjiang, neighbor Yunfu) (Top Equipment): Spare parts procurement 88M→half; fault duration 5,900→3,800 min; fault prediction accuracy +85%; data collection rate 98%. Source: Yangjiang Gov / Provincial Science Award.

Rizhao Steel Equipment AI Inspection (Top Equipment + Knowledge): Repair plan 3h→0.5h; spare parts mismatch 30%→5%; knowledge reuse 40%→85%; new hire training 6 months→2 months. Source: China Daily.

Jingye / Yongyang Blast Furnace Diagnosis + Maintenance (Top Energy + Equipment): Coke per tonne pig iron -2.1kg, fuel -3.5kg, daily output +100 tonnes; Yongyang half-year 23 early warnings, maintenance -10~20%, annual saving 11M. Source: Hebei Industry & IT / Yongyang Special Steel.

A notable industry signal: among Hebei's 55 smelting sites, over half deeply apply AI large models; 3 steel vertical large models passed national filing; 160 professional models span the industrial chain; average production efficiency up ~20%; comprehensive energy consumption per tonne steel down to 550 kg standard coal, below national average. Steel AI has crossed from "pilot show" to "scale replication" — and the replication lever is precisely public goods like provincial shared platforms that "lay the data foundation first."

Sober View: Three Romanticized Misconceptions

Writing about steel AI easily falls into "robots replace steelworkers" or "black-box fully automatic steelmaking" narratives. Three traps must be flagged, two directly bearing on regional investment attraction.

1. AI Deployment Order Is Decided by Data Availability, Not Value Size

The first scaled deployments in steel AI are not "seen" visual inspection but "calculated" vehicle/vessel scheduling and production scheduling — because production-line data is hard to obtain and formats inconsistent, while molten iron transport, raw-material inventory, and order scheduling data already sit in ERP/MES systems. Shougang Qian'an's 70M saving came from "data already there." Conversely, visual inspection must wait for sensors, cameras, and labeling systems to catch up. This is the most counter-intuitive point and the one that must be written into investment proposals.

2. Mechanism Models Remain the Foundation; AI Does Not Make Final Calls

China Baowu expert Wang Songtao explicitly states the industry's biggest shortfall is digital infrastructure — sensors, instruments, industrial control systems, level-2 models, and other basic capabilities still need improvement. Hebei's 20% average efficiency gain relies on "mechanism model + data model" dual-drive, not discarding mechanism. Forcing "end-to-end black-box steelmaking" is non-compliant and uneconomical: a single breakout or scrap loss far exceeds manual operation cost.

3. Regional Investment Must Distinguish "What Can Be Promised" from "What Cannot"

Yunfu Metal Intelligent Manufacturing possesses ~15 million tonnes/year steel capacity and Xi River 5,000-tonne-class navigation logistics advantages — real chips for attracting metal deep processing (aluminum profiles, stainless steel kitchenware, prefabricated building components). Promisable: "energy and scheduling data ready, logistics cost quantifiably optimizable, connectable to provincial compute and shared model platforms." Not promisable: "production-line visual inspection in one step" — that requires first filling sensor and process data foundations, a multi-year investment. Three self-checks for metal processing firms landing AI:

Do energy & scheduling first (data ready), then defect tracing (needs process data), don't start with vision.

Checklist: What Steel/Metal AI Projects Must Ask First

In one sentence: Don't rush the "robots replace steelworkers" narrative; first ask if the data foundation is laid. Don't hype "end-to-end black-box steelmaking"; first clarify the boundaries of mechanism and human-in-the-loop. Three executable checks:

Check Data Availability First. Are energy, scheduling, logistics data in systems? Are production-line sensors and level-2 models complete? Without ready data, upper large models spin empty; Yangchun New Steel pushed data collection to 98% to support prediction accuracy +85%.

Write Mechanism & Human-in-the-Loop Boundaries into Process. AI advises and closes loops, but charging and tapping final judgment stays with engineers; strike "fully automatic black box" from proposals, first verify "mechanism + data" dual-drive is solid.

Investment Attraction: Look at "What Can Be Promised". Regional metal intelligent manufacturing can use "15M tonnes steel + Xi River 5,000-tonne logistics" to promise energy & scheduling optimization and provincial shared platform access; visual inspection demands must flag data foundation investment cycle, don't trade immature promises for projects.

Steel is the "heaviest" piece for AI landing: it bears hard constraints of energy, carbon, and continuous production, and most tests the patience of "data foundation first, mechanism models as bottom." Its paradox is the series' sharpest tension — the industry most emphasizing mechanism and safety is the first to let large models sit in the scheduling layer; and what determines how far it goes is not how fancy the algorithm is, but how solid the data foundation is laid. Next episode, we turn from "steel in the furnace" to "reactions in the pipes" — AI+Chemical, the next fortress of process industry.