ProductionNewVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

Machine learning recommends the minimum alloy addition for each heat at a steel mill (Gerdau, Fero Labs)

Added 2026-10-19 · Last updated 2026-10-19

Role

Production (process optimization: recipe and additive amounts per batch) · Typical users: process engineer, metallurgist, quality engineer, furnace and ladle operator, continuous improvement lead

The problem

In batch processes such as steelmaking, every batch starts from different raw materials, so plants set conservative recipe limits to make sure every batch meets spec. Quality stays high, but the plant overspends on additives. At Gerdau's Petersburg mill (about 200 products and 25 grades, over 4,000 heats a year, hundreds of process readings), scrap quality, alloy recovery, slag and rolling mill temperatures all vary, and the alloy addition range for each grade was set at a conservative level. The capability analyses used to tune those ranges were done only once or twice a year, on a small subset of the data, and assume linear relationships and normally distributed data.

How AI helps

A white-box machine learning model, trained on melt shop, rolling mill and test lab data (over 20,000 heats covering the mill's whole grade portfolio), predicts the final yield and tensile strength while the steel is still in the ladle. After each ladle sample, it recommends the lowest-cost alloy addition that still meets the specification with a set confidence, within about 30 seconds of the spectrometer reading. The ladle operator sees the recommendation on a screen in the pulpit and makes the addition. The quality team owns the model constraints and the safety buffers added to the spec limits. The model shows confidence bands and rules itself out on new, rare or hard-to-make grades, where operators keep the standard grade limits.

Step by step

Pick one high-volume product family where an additive, ingredient or energy input is expensive and the spec is measured by a lab test; record the baseline cost per unit and the failure rate.
Build one table per batch: raw material analysis, additions, key process readings and the final lab results, with at least a year of history.
Train and back-test a model offline: does it predict the final lab result well enough, and where are its confidence bands too wide (rare grades, noisy sensors)?
Have the quality team set the constraints and a safety buffer on the spec limits, and the cases where the model must not be used.
Run in advisory mode: the operator sees the recommended amount on a screen where the decision is made and decides; log every recommendation, the amount actually added and the final result.
Track cost per unit, quality variation and failures against the baseline; relax the buffers grade by grade only after hundreds of batches, and retrain the model regularly (about monthly at Gerdau).

Tools

Fero Labs (no-code, white-box machine learning for process data): used at Gerdau
Other tools to evaluate: process analytics and optimization software that reads your historian and lab data (for example Seeq or Braincube), or your MES or historian vendor's analytics module
For a first offline test: any LLM with a data-analysis or code tool (ChatGPT, Claude, Copilot, Gemini) on an export of past batches, with the prompt below

Result

Gerdau Petersburg mill (Virginia, USA), with Fero Labs, live since 2020: thousands of heats cast with ML-optimized chemistry; alloy cost down $3 per ton on average while meeting quality specs; alloy additions down about 8%; more than 500,000 pounds of manganese, carbon, niobium and vanadium conserved; 15% less quality variation (AISTech 2023 paper co-written by Gerdau and Fero Labs; Fero Labs, March 2023). Moving the operator screen to a central monitor in the ladle furnace pulpit led to a 2.5-fold increase in tonnage cast with the recommendations and a threefold improvement in profits within one month. The mill kept deliberate buffers on the spec limits, which in some cases cut the per-ton savings potential by up to 55%; the paper puts the true potential alloy reduction at twice the 8% achieved. Keeping the model up to date takes less than 10 minutes a week, with retraining about once a month. Gerdau has expanded Fero to five other plants in North America (Fero Labs case study). The figures are published by the vendor and the mill.

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

The quality team, not the model, owns the spec limits and the safety buffers; recommendations stay advisory and the operator decides on every batch.
Models are weak on new or rarely made grades and on noisy sensor data: use the confidence bands, and fall back to standard limits where the model rules itself out.
Savings depend on the process upstream leaving room to add less: at Gerdau, the furnace operators had to change their initial additions so the ladle operators could follow the lower recommendations.
The figures are reported by the vendor and the mill, for one steel mill; your process, data and margins will differ.
Do not upload confidential recipes or process data to a public chatbot; use an approved enterprise tool.

Try it yourself

You are a process engineer analysing past production batches to see whether we overdose an expensive additive.

I will paste or attach a table with one row per batch: [batch ID, product or grade, raw material analysis, amount of the additive added, key process readings, final lab result(s)], plus the specification limits per grade [grade, lower limit, upper limit].

Do this, using code on the data:
1. Check the data: count batches per grade, missing values and obvious outliers; list them, do not fix them silently.
2. Per grade, show the distribution of the final lab result against the spec limits and how far above the lower limit most batches finish.
3. Fit a simple, explainable model of the final result from the additive amount and the other inputs; report its error on held-out batches and which inputs matter most.
4. List the grades where the data suggests we could add less while staying inside spec with a margin, and the grades where the data is too thin or too noisy to say.
5. Write the questions and the trial plan I should take to the quality team before any change.

Do not recommend changing any recipe or setpoint directly: this is an offline analysis for discussion.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 8 Production use cases →