Recipe Copilot

Every heat is different. Why should every recipe be static?

Gradetarget band shifts
Scraplot mix and recovery
Chemistrysample to sample
Temperaturebath trajectory
Inventorywhat is actually on hand
Pricesalloy cost today

The recipe should too. Recipe Copilot evaluates each heat, recommends the best recipe action, and learns from every outcome — a measurable, auditable part of how the melt shop runs.

Heat 4712 · 42CrMo4 · EAF-2Sample 2 received 14:06
ElementNowTargetAction
Mn0.620.70–0.80+38 kg FeMn
Cr1.081.00–1.20Hold
C0.440.38–0.45Hold
Mo0.170.15–0.30+6 kg FeMo
Why: FeMn lot M-118 recovery has run 2% high this week; recommended mass is trimmed to avoid overshooting the Mn band. Confidence 0.93 based on 214 similar heats.
ApproveAdjustOverride with reason

Try it — these are live.

Approving releases the recommendation to the charge floor and books it against the heat.
Approve recommendation
Predicted Mn after addition: 0.700%
0.55%
Within target band 0.70–0.80

Illustrative response model — a real pilot calibrates it on your recovery history.

Log override
Overrides are never blocked — they are recorded, and they train the model.

Why the spreadsheet stops short.

Excel, fixed formulas and small rule-based tools are useful for standard calculations. But they depend on operator judgement, manual entry and historical assumptions that do not adapt when scrap mix, chemistry, temperature, furnace behaviour, inventory or material cost change.

The plant pays for that gap through over-alloying, chemistry corrections, delayed heat closure and inconsistent decision quality across shifts. The issue is not only cost. It is predictability, repeatability and governance of every heat decision.

Recipe treated as a static calculation

The formula does not know what the furnace is doing right now, or what the lab said ten minutes ago.

Limited link to quality, inventory and ERP

Lot recovery, price and availability live elsewhere. The recommendation is made without them.

No learning loop from actual heat outcomes

What happened after the addition never flows back into the next decision.

Weak traceability when decisions go wrong

Who changed what, why, and against which data — hard to reconstruct after the fact.

Mn trajectory · Heat 4712 · target 0.70–0.80Live
Target band 0.70–0.80%
Current Mn 0.62%Recommendation recalculates on every sample
20–40%of decision time is reducible when live data is pre-assembled for the operator
95%+of recommendations traceable to inputs, approvals and outcomes

Understand. Recommend. Learn.

Recipe Copilot runs the same loop on every heat. The operator stays in charge at every step.

1

Understand the live heat

Reads live heat context before recommending anything.

  • Grade and target chemistry
  • Charge mix and raw materials
  • Temperature and furnace status
  • Historical performance on similar heats
2

Recommend the best action

Suggests the action, shows its confidence, and explains why.

  • Additive quality and quantity
  • Timing of the addition
  • Recipe adjustment against the target band
  • Confidence score and reasoning
3

Learn from every heat

Every outcome becomes training data for the next one.

  • Actual additions and lab results
  • Delays and corrections
  • Operator approvals and overrides
  • Reason codes on every deviation
AI recommends.The operator decides.The system learns.

Business value you can measure per heat.

Every recommendation is compared against the baseline it replaced, so value is reported from production records — not estimated at the end of the quarter.

Higher yield

More saleable steel from the same charge. Every 1% of yield is output you already paid to melt.

Lower alloy cost

Optimised additions against real lot recovery, price and availability — instead of over-alloying for safety.

Better quality

Fewer off-spec heats, fewer corrections, fewer downgrades and rejections.

Faster decisions

Real-time guidance with the data already assembled. Heats close sooner, shifts decide alike.

Governed by design.

Operators accept a recommendation when they understand it, and leadership trusts a system when it can be audited. Recipe Copilot is built for both.

ApprovalsReason codesAudit trailModel confidenceFallback rulesChange control
ERP / SAPOrders, grade plan, cost centres, material master.
MES / SchedulingHeat plan, route, sequence, delay codes.
LIMS / LabChemistry, sample history, release status.
SCADA / PLCTemperature, power, oxygen, runtime signals.
InventoryAdditive availability, lot, price, supplier.
QMS / TraceabilityDeviations, approvals, audit trail.

Keep your know-how. Connect it to the live plant.

The pilot does not replace what works. It takes your existing recipe logic, connects it to live plant systems, and converts recipe management into a controlled AI layer that improves every heat — explainable to operators, auditable for leadership.

2 weeks

Data & recipe audit

Grade list, input quality, formula logic, gaps.

2–3 weeks

Model & rules setup

Chemistry prediction, constraints, fallback logic.

2 weeks

Operator workflow

Recommendation screen, approvals, reason codes.

4–6 weeks

Live pilot

Selected grades or route, measured against baseline.

2 weeks

Scale blueprint

ERP/MES rollout plan, SOPs, ROI sign-off.

Questions we get asked

Recipe optimisation, answered.

What plant leadership and metallurgists ask about AI-based recipe and charge optimisation.

What is Recipe Copilot?

Recipe Copilot is AI decision-support for charge mix and alloy addition decisions in steel and alloy melting. For each heat it reads grade and target chemistry, charge mix and raw material lots, bath temperature and furnace status, and the performance of similar past heats, then recommends which additive, how much and when — with a confidence score and its reasoning. The operator approves, adjusts or overrides, and the outcome trains the model.

How does AI charge optimisation reduce ferroalloy cost?

Over-alloying happens because operators add a safety margin against uncertainty in lot recovery, scrap chemistry and timing. Recipe Copilot narrows that uncertainty using measured recovery history for the specific additive lot, live chemistry from the lab and the current bath state, so the recommended mass sits closer to the middle of the target band. Audited plants commonly carry 0.5–1% of material cost in avoidable alloy additions.

Can it work with our existing recipe formulas?

Yes. Existing formulas and standard recipes are retained as the fallback layer and as the starting point for the model. The pilot connects that logic to live ERP, MES, LIMS, SCADA and inventory data instead of replacing it.

What chemistry elements and grades does it handle?

Recipe Copilot works on the elements that carry a specified band for the grade being made — typically C, Mn, Cr, Mo, Si, Ni, V and residuals such as Cu and P. It is used across engineering, alloy, bearing, tool, stainless and free-cutting grades. Grade lists and target bands are configured from your own specifications during the audit phase.

How are recommendations audited?

Every recommendation records its inputs, model version, confidence, the number of similar heats it drew on, the operator's decision, any reason code for an override, and the measured outcome once the next sample posts. That record is available per heat for internal audit and customer quality claims.

See Recipe Copilot on your grades.

Bring one grade list and a month of heat records. We will show where the recipe decision is leaking value and what the copilot would have recommended.