EVRAZ deploys a neural network to optimize steel heating and cut gas use in its rolling mill

EVRAZ has described how, in the hot rolling mill at NTMK, it replaced manual furnace tuning with a recommendation system built on a mathematical model and a neural network. The goal was not automation for the sake of a report, but cutting excess gas use when heating billets, without hurting rolled-product quality.

Why the gas was being wasted

In hot rolling, billets pass through reheating furnaces before they reach the mill, and on paper this stage looks routine. In practice, every furnace behaves differently: the design, the condition of the burners, lining wear and even the billet's path to the first stand all vary. The result is also affected by steel grade, cross-section, temperature before charging, scheduled roll changes and unplanned stoppages. For a shop with a wide product range, this means a single mode almost never suits every consecutive batch.

Previously, operators set temperature, heating time and gas flow by hand, relying on the instructions, their experience and the current state of the equipment. Formally the rules were the same for everyone, but in a real shift they had to be adapted constantly to the situation. As a result, different crews showed different fuel consumption, and when switching between product mixes, overconsumption became almost the norm. EVRAZ describes the problem directly as a gap between the document and the real shop floor.

«Follow the technological instruction».

What the model is made of

The team quickly realized that a pure ML model would not work here. Billet temperature is measured only at furnace entry and exit, and there is almost no data on how the metal heats up during the cycle. So the solution is built on a physical foundation: a numerical heating model based on the heat conduction law, which accounts for heat from the burners, convection, radiative and contact heat exchange, and the properties of the metal at different temperatures. The model also accounts separately for scale formation, which effectively creates a heat-insulating layer on the billet surface.

To make the system work in a real shop, the model had to be fitted to specific furnaces and actual production modes. During tuning, details surfaced that are rarely visible in abstract diagrams: air infiltration from the shop, the effect of billet spacing, downtimes known in advance during roll changes, and even difficulties interpreting pyrometer data after the roughing stand. So the system accounts not only for theory but also for a set of factors that typically get lost in standard instructions:

• differences in furnace geometry and burner placement
• air infiltration from the shop into the furnace atmosphere
• billet spacing and the gaps between billets
• downtimes known in advance during mill roll changes
• the特点 of dozens of steel grades and different cross-section shapes

The results of the numerical simulation were then packed into a neural network. It was trained on calculated data from tens of thousands of heating scenarios, extracted from a historical archive spanning several years. This hybrid approach gave the system two things at once: physical plausibility and speed sufficient for soft real time. The result is a digital twin of the heating process that does not merely forecast temperature but tells operators how to set the furnace zones and burners for a specific batch of billets.

Validation in production

Before deployment, the model was checked against two main criteria. First, the calculated billet temperature was compared with pyrometer data right at the furnace discharge, not after the roughing stand, where the picture is already distorted by subsequent cooling. Second, the team used archived tests of instrumented billets with sensors inside the metal. Even though the furnace has since undergone a major overhaul, the data helped confirm that the model correctly reflects the real heating profile.

EVRAZ also separately assessed how accurately the system calculates the required energy consumption over time. It compared actual and model values using the coefficient of determination R². The company reports a value of 0.75, which was enough to show production and management that the model is adequate.

After the technical work, the team built a working interface: operators see a zone-by-zone diagram of the furnace and recommendations for burner settings. The pilot in one shop was judged a success. The exact savings figures are not disclosed, but the system is already being prepared for rollout to other furnaces at the plant.

What this means

This case shows well where industrial AI is actually heading. The most useful effect here came not from a general-purpose chatbot, but from a combination of a digital twin, historical data and hints that operators can understand. If such systems start to be replicated, metallurgy could save energy not through rigid restrictions but through tuning the process more precisely for each specific batch.