Hot rolled steel control with AI — the window is narrow and the defect shows up with the coil already cold.

The problem

The deviation happens on the mill in seconds — but you pay for it in the coil yard or at the customer.

The quality of a hot rolled coil is the intersection of four pieces of data that are almost never on the same screen:

1. Which steel went in — composition and heat certificate of the slab. In the melt shop system, delayed and in another building.
2. How it went through the mill — entry and finishing temperature, speed, reduction per pass, roll loads. On the pulpit panels, in front of the operator and nowhere else.
3. How it cooled — cooling on the run-out table and coiling temperature, which set the microstructure and the final flatness.
4. What the strip looks like coming out — rolled-in scale, periodic roll marks, scratches, edge laps. A glance from the operator at the moving strip and little more.

The quality manager knows this, but most defects only get confirmed once the coil has cooled (and then you have to decide whether to downgrade the whole thing, cut it or reassign it to a lesser destination) or at the customer (when the strip is pickled, formed or painted and the mark comes out). By then the next heat has already been rolled with the same parameters. And the data that would explain what happened lives on a panel nobody logged and in a handwritten shift report. The classic system works 99% of the time. That 1% is what the monthly committee pays for.

How it fits the IRIS system

iLEAN does not replace the mill's control system — it seals the cracks between heat, curve and coil.

The problem on a hot strip mill is not a lack of information: it is information living on separate islands — the panel, the shift report and the heat certificate — that never reaches, cross-referenced, the person who decides where the coil goes. iLEAN acts as the putty that fills those gaps, without asking you to change the mill control system, to open the loop, or to negotiate an integration with the line builder.

Connect reads the mill panels without touching the control system. Edge classifies the surface on the hot strip. The agent cross-references heat, curve and coil in the central memory and builds the dossier. The person decides where the coil goes.

The three iLEAN pieces applied to hot rolling:

• Edge — a computer-vision terminal (CNN) over the moving strip, with a camera and lighting adapted to the radiation of the hot strip. It classifies rolled-in scale, periodic roll marks, longitudinal scratches and edge laps, and flags the questionable stretch with its position in meters of coil. It works with no network.
• Connect — non-invasive reading of the mill panels: a camera aimed at the pulpit display or the mimic panel transcribes temperatures, speeds and loads into the central memory, without touching the PLC. And it captures everything else: heat certificate, rolling schedule, shift report, roll change, a heads-up from a customer with a tighter spec.
• Agent — cross-references heat ↔ mill curve ↔ coil in the central memory: it links the surface signature to the actual temperature and reduction window each stretch was rolled in, alerts the quality manager through whatever channel they use when the curve approaches the edge of the window, and builds the per-coil dossier automatically. The person decides and signs; no coil is downgraded on its own.

See the full IRIS architecture →

Before and after

Panels and shift reports vs. cross-referenced control with iLEAN

Aspect	Mill panels + shift report	With iLEAN Edge + Connect + Agent
Mill curve (temperature, speed, reduction)	On the pulpit panel, with no cross-referenced record	Read off the panel without touching the control, attached to each coil
Heat certificate	In the melt shop system, in another building	Cross-referenced with the curve and with the strip surface
Surface defect	A glance at the strip and confirmation once the coil is cold	Classified on the strip, with its position in meters of coil
Dimensional deviation	Discovered on cooling, already in the coil yard	Alert when the curve approaches the edge of the window
Operation without a network	n/a	Edge keeps classifying with its own light
File for the customer	Rebuilt by hand after a complaint, from shift memory	Per-coil dossier, automatic, with heat + curve + image

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Hot strip mill supplying sheet processors: automotive, white goods, tube, structural steel or slitting services.
• Edge pilot on the run-out table (surface camera with optics for hot strip + non-invasive reading of the mill panels + integration with the heat certificate). First value expected within a few weeks.
• Reduction of downgraded coils and of customer complaints against the historical baseline ≥ 25% — a conservative estimate.
• Indicative payback between 5 and 10 months, depending on the average value of a downgraded coil and the number of surface or flatness complaints over the last year.
• The hard lever is cutting out the bad stretch instead of downgrading the whole coil, and coming to the complaint with a dossier rather than with a hand-made reconstruction.

And the quality manager's reasonable doubt

“What if the AI flags a stretch as defective when it was fine?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely cross-references surface signature + mill curve + heat composition, the best models brought error below 1.5% ^[1]. And even so, the flagged stretch goes to a second review on the cold coil, not straight to a downgrade. The person decides; no coil is downgraded on its own. The three safety rings are there for exactly this.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.