Beyond Endpoint Guesswork: Hybrid Level 2 Intelligence for BOF / LD Steelmaking

How static heat planning, dynamic blow control, first-principles metallurgy and machine learning can work together to improve endpoint consistency, reduce corrective actions and create a practical digital-twin foundation for oxygen steelmaking

By ExtractMet Private Limited
Metallurgical engineering • Process intelligence • Advanced process control • Digital twins

Explore ExtractMet’s process-control and digital-twin capabilities

Figure 1 — A hybrid BOF Level 2 system combines metallurgical models, live plant data and data-driven intelligence to turn process measurements into actionable control guidance.

A Basic Oxygen Furnace can transform hot metal into steel in minutes. Yet the last part of the blow can still become a high-stakes exercise in uncertainty.

The operator may know the hot-metal analysis, scrap weight, oxygen flow, lance profile, flux additions and laboratory targets. The automation system may collect thousands of tags. The historian may preserve years of heats. But the variables that matter most at a given instant — the true bath carbon, bath temperature, reaction progress, slag FeO, remaining oxygen requirement and likelihood of hitting the endpoint — are not always measured continuously.

That gap between what the plant measures and what the process is actually doing is where a Level 2 process-control system creates value.

For BOF/LD steelmaking, a serious Level 2 solution should do more than display trends. It should calculate the metallurgical state of the heat, update that state as new information arrives, estimate what is not directly measured, predict the likely endpoint, and recommend an operating action within plant constraints.

At ExtractMet, we believe the most practical route is a hybrid architecture: first-principles metallurgy provides the auditable backbone, while machine learning learns plant-specific deviations, nonlinearities and operating patterns that are difficult to represent perfectly from equations alone.

See ExtractMet’s BOF and steelmaking digital-twin portfolio

Figure 2 — Process modelling becomes commercially useful when it is connected to plant data, operating logic and decisions rather than remaining an isolated simulation.

The BOF control problem is not a shortage of data

Modern oxygen steelmaking already produces substantial information: hot-metal weight and chemistry, scrap and coolant additions, flux analyses, oxygen and bottom-gas flows, lance position, vessel events, off-gas measurements where installed, sublance or immersion-probe measurements where available, tapping results, laboratory chemistry and production history.

The difficulty is that these signals arrive at different frequencies, with different delays and different levels of uncertainty. Some are measurements. Some are laboratory results. Some are operator-entered values. Some are inferred from equipment states. And some of the most important process variables are hidden.

This means the real question is not:

“How much data do we have?”

It is:

“How do we convert the data we have into a continuously updated metallurgical state — and then into the next best decision?”

That is the role of Level 2.

What Level 2 should do in a BOF shop

At the simplest level, Level 1 automation executes equipment sequencing, interlocks and regulatory control. Level 2 sits above that layer and provides process intelligence: metallurgical calculations, setpoint generation, state estimation, endpoint prediction, optimization, operator guidance, heat tracking and integration with production and quality data.

The exact architecture depends on the plant, but a practical BOF Level 2 system may connect:

• Level 1 / PLC / DCS process tags and equipment states;
• hot-metal, scrap, flux, coolant and alloy data;
• laboratory and LIMS results;
• production recipes, steel grades and target constraints;
• historian and heat-event data;
• static and dynamic metallurgical models;
• data reconciliation and soft sensors;
• machine-learning correction models;
• operator advisory screens and heat-replay tools; and
• approved setpoint interfaces where closed-loop operation is justified.

Figure 3 — A Level 2 layer should calculate, reconcile, predict and recommend. Plant deployment remains subject to site-specific validation, cybersecurity review and approved automation interfaces.

Explore ExtractMet’s model catalogue

Static control: make the best heat plan before the blow begins

The first job of a Level 2 system is to calculate a credible starting strategy.

A static BOF model uses the information known before or at the beginning of the heat to estimate the material and thermal requirements needed to reach the target. Depending on plant practice, the calculation can include:

• hot-metal weight, temperature and C–Si–Mn–P–S chemistry;
• scrap quantity and scrap categories;
• ore, DRI, coolant or other solid additions;
• lime, dolomite and other flux analyses;
• carry-over slag assumptions;
• vessel thermal state and heat-loss parameters;
• target steel weight, carbon and temperature;
• target slag basicity and phosphorus-removal requirements; and
• plant-specific constraints on oxygen, flux, coolant and blowing practice.

The model can then calculate or recommend quantities such as:

• theoretical and calibrated oxygen demand;
• expected scrap-melting and cooling requirement;
• lime and dolomite additions;
• expected slag quantity and basicity;
• predicted endpoint carbon and temperature;
• estimated metallic yield and iron oxidation;
• expected gas generation; and
• the initial lance, stirring and addition strategy to be passed to the operating practice.

The strength of a first-principles static model is transparency. Every tonne of iron, every kilogram of carbon, every mole of oxygen and every major heat term must go somewhere. Mass, elemental and energy conservation create an auditable backbone that plant metallurgists can inspect and challenge.

But no plant behaves exactly like a textbook balance. Scrap geometry changes. Flux reactivity changes. Hot-metal sampling has uncertainty. Vessel age changes heat loss. Slag carry-over varies. Reaction distribution is not perfectly known. That is why the static model should be calibrated against plant history rather than treated as an untouchable theoretical calculation.

Dynamic control: keep recalculating while the heat is evolving

The static model gives the best plan available at the start. The dynamic model answers a different question:

Given what has actually happened during this heat, where is the process now?

A dynamic BOF model advances the heat through time or event steps. As oxygen is blown and materials are added, it updates the estimated state of the metal, slag and gas phases.

Depending on available instrumentation, the dynamic update may use:

• cumulative and instantaneous oxygen flow;
• lance height and blowing pattern;
• bottom-gas stirring flow;
• timed flux, ore, coolant or other additions;
• off-gas CO, CO₂, O₂ and flow information;
• vessel pressure and gas-system data;
• sublance measurements;
• immersion temperature or sample results;
• acoustic, vibration or image-derived indicators; and
• operator or equipment events.

The dynamic calculation can continuously revise estimates of bath carbon, bath temperature, slag mass, slag FeO, reaction progress, gas evolution and remaining oxygen demand.

Figure 4 — Static control establishes the heat plan; dynamic control repeatedly re-estimates the process state; post-heat comparison provides the calibration and learning loop.

This is an important distinction. A dynamic model should not merely recalculate the original static prediction with a new timestamp. It should reconcile the plan with what the heat actually did.

For example, if a heat shows a different decarburization trajectory from the expected one, the system can revise the predicted endpoint. If a sublance result arrives, the model can re-anchor its state estimate. If off-gas information indicates a change in oxidation behaviour, the dynamic model can incorporate that information. If sensors are unavailable, the system can still advance through a physics-based state estimate but should report lower confidence rather than pretending that uncertainty does not exist.

Why first-principles alone is not enough — and pure AI is not enough either

There are two tempting extremes in industrial process modelling.

The first is to build an extremely detailed first-principles model and assume that every plant deviation can be captured through equations. The second is to feed historical data into a powerful machine-learning algorithm and assume that the model will learn everything automatically.

Both approaches have limitations.

A purely first-principles model can struggle with plant-specific unknowns: effective reaction distributions, uncertain heat losses, variable scrap melting, flux dissolution, slag carry-over, sensor bias and operational behaviour. Increasing equation complexity does not always increase plant accuracy.

A purely data-driven model has the opposite weakness. It may fit historical heats very well but can become unreliable when raw materials, steel grades, vessel condition, operating practices or sensor behaviour move outside the training envelope. A black-box prediction may also be difficult for operating teams to interpret when it disagrees with metallurgical expectations.

The hybrid approach is designed to combine the advantages of both.

Figure 5 — Physics provides conservation laws and metallurgical structure. Machine learning provides plant-specific correction, adaptation and soft sensing. The combined state estimate supports decisions.

In a hybrid BOF model, first principles can provide

• mass and elemental balances for Fe, C, Si, Mn, P, S and selected species;
• heat balance and sensible-heat calculations;
• oxygen allocation to oxidation reactions;
• decarburization and gas-generation logic;
• slag formation and FeO evolution;
• basicity and flux requirements;
• scrap-melting and coolant effects;
• slag–metal–gas reaction constraints; and
• physically consistent state variables through the heat.

Machine learning can then add value in targeted roles

• Residual correction: learn the systematic difference between model prediction and plant outcome.
• Adaptive parameter estimation: tune uncertain model coefficients by campaign, grade, vessel age or operating regime.
• Soft sensing: estimate endpoint carbon, temperature, FeO or other hidden states from available signals.
• Anomaly detection: identify heats, sensors or events that differ materially from the normal operating envelope.
• Risk classification: estimate the probability of reblow, endpoint miss, slopping or another defined operational event.
• Confidence estimation: indicate when the model is operating inside or outside the domain represented by the training data.

This is more useful than attaching “AI” to a dashboard. The machine-learning layer should have a defined metallurgical role and should be evaluated against plant KPIs.

The best hybrid design uses ML to correct physics — not to hide it

One powerful architecture is residual learning.

Suppose the first-principles model predicts endpoint temperature of (T_{FP}). Historical data shows that under certain combinations of scrap category, hot-metal silicon, vessel campaign and blowing practice, the model has a systematic bias. A machine-learning model can learn the residual:

Residual = Actual endpoint − First-principles prediction

The final hybrid estimate becomes:

Hybrid prediction = First-principles prediction + ML residual correction

The same concept can be used for carbon, oxygen demand, slag FeO or selected calibration parameters.

This architecture has several practical advantages. The ML model does not need to learn the entire BOF from scratch. The physics model handles conservation and broad metallurgical behaviour; the ML layer concentrates on the plant-specific error structure. When operating conditions move, engineers can inspect whether the issue originates in plant data, physics assumptions or the learned residual.

Other hybrid architectures are also possible: ML-estimated parameters inside the physics model, physics-informed constraints in the loss function, ensemble models, Bayesian updating, Kalman-style state estimation, or separate expert models combined through a decision layer.

The right architecture should be selected from the plant problem — not from the popularity of a particular algorithm.

What should the dynamic model actually track?

A commercially useful BOF dynamic model should be organized around process states that operators and metallurgists care about.

Metal bath

• liquid steel mass;
• remaining unmelted solids where relevant;
• carbon and selected elemental composition;
• bath temperature;
• iron oxidation and metallic yield; and
• endpoint trajectory.

Slag

• slag mass;
• CaO, SiO₂, FeO, MgO, MnO and other relevant components;
• basicity;
• phosphorus-removal potential where modeled;
• flux dissolution state; and
• slopping / foaming-related indicators where suitable data exist.

Gas phase

• CO and CO₂ generation;
• oxygen utilization;
• post-combustion assumptions or estimates;
• off-gas flow and chemistry where measured; and
• gas-based indicators of reaction progress.

Energy

• heat released by oxidation;
• sensible heat of hot metal and additions;
• scrap and coolant melting demand;
• heat carried by slag, steel and gas;
• estimated vessel/process heat loss; and
• predicted endpoint temperature.

A useful model should also preserve heat history. Operators should be able to replay a heat and ask: Why did the model recommend this oxygen amount? Why did the temperature prediction move? Which input caused the residual? When did the process depart from the expected trajectory?

That traceability is essential for adoption.

An illustrative heat: how the hybrid system thinks

Consider a heat in which the hot-metal silicon is higher than the recent average and the scrap mix contains a larger fraction of heavier pieces.

A conventional recipe may treat this as a small variation. A hybrid Level 2 system can reason through the consequences in stages.

Before the blow, the static model sees the additional oxidation heat associated with the changed hot-metal chemistry and recalculates oxygen, flux and thermal balance. It also considers the expected effect on slag generation and the cooling/melting requirement.

During the blow, the dynamic model compares actual oxygen delivery and additions against the plan. If live process signals indicate that the decarburization or thermal trajectory differs from expectation, the state estimate moves accordingly.

The machine-learning layer recognizes whether similar combinations of chemistry, scrap category, vessel condition and operating practice historically produced a systematic model residual.

Near the endpoint, the system does not simply repeat the original target oxygen volume. It calculates the remaining requirement from the updated state and provides a prediction with a confidence level.

After tapping, the measured endpoint is compared with the prediction. The residual is logged, the heat is classified, and controlled model-adaptation logic can update the calibration dataset.

That closed learning loop is what transforms a one-time mathematical model into a plant process-intelligence system.

The real KPI is not model accuracy — it is shop-floor performance

A Level 2 project can generate impressive machine-learning statistics and still create little industrial value. The success criteria should therefore be agreed before implementation and linked to the steel shop.

Typical BOF KPIs can include:

• endpoint carbon hit rate;
• endpoint temperature hit rate;
• double-hit rate for carbon and temperature;
• reblow or corrective-action frequency;
• oxygen overblow / underblow variability;
• heat and tap-to-tap time variability;
• lime and flux consumption;
• slag FeO and basicity consistency;
• metallic yield and iron loss;
• slopping frequency or severity;
• operator intervention rate;
• repeatability across shifts and crews; and
• model availability, confidence and data-quality performance.

Figure 6 — A Level 2 project should be judged by operational KPIs, not by “AI accuracy” in isolation.

The business case then becomes measurable. Better endpoint control can reduce unnecessary corrections. More stable slag control can support yield and refining consistency. Better heat prediction can reduce decision variability. Better heat replay can shorten troubleshooting time. The specific financial value must be calculated from the plant baseline rather than assumed in advance.

Current industry direction: Level 2 is becoming more integrated, not less important

Recent BOF automation projects continue to combine process models, Level 2 optimization and increasingly rich sensor information. Commercial systems now integrate metallurgical models with sublance and/or off-gas data, generate oxygen and material setpoints, and connect multiple meltshop units within unified digital architectures.

In parallel, peer-reviewed research continues to advance both first-principles dynamic modelling and data-driven endpoint prediction. The direction is clear: the opportunity is not “physics versus AI.” It is physics plus plant data plus adaptive prediction plus disciplined deployment.

That is exactly the design space in which a hybrid Level 2 system can be most effective.

Where ExtractMet fits

ExtractMet Private Limited works on metallurgical process optimization, mathematical modelling, plant-data analysis, advanced process control and digital twins for iron, steel and other high-temperature processes.

For BOF/LD steelmaking, a plant-specific development program can be structured around modules such as:

• static mass, elemental and heat-balance model;
• charge / oxygen / flux calculation engine;
• dynamic time-resolved BOF model;
• endpoint carbon and temperature prediction;
• slag mass, FeO and basicity estimation;
• off-gas reconciliation where suitable instrumentation exists;
• machine-learning residual correction and soft sensors;
• heat-to-heat adaptive calibration;
• heat replay and root-cause analysis;
• operator advisory dashboard;
• model validation and KPI reporting; and
• interface definition for Level 1, historian, LIMS or MES integration.

ExtractMet’s broader digital-twin approach begins with mass, elemental and heat balances, thermodynamics, kinetics and process logic, then calibrates those models using plant measurements and augments them — where useful — with data reconciliation, statistical methods, AI/ML and optimization.

Explore ExtractMet’s digital-twin architecture and BOF model

Figure 7 — BOF process intelligence should connect the oxygen blow, metal bath, slag and gas system to endpoint decisions.

A practical deployment path

A plant does not need to jump directly from spreadsheets to closed-loop automation. A staged route is usually easier to validate and easier for operating teams to trust.

1. Data audit and problem definition

Choose one measurable decision problem. Map all relevant tags, laboratory data, material analyses, timing, missing values and data-quality limitations. Define the KPI and the current baseline.

2. Build and validate the static model

Develop the mass, elemental, heat and reaction calculations. Calibrate uncertain parameters against representative historical heats. Confirm unit consistency and conservation.

3. Build dynamic heat replay

Reconstruct the heat event by event. Track oxygen, additions, reaction progress, bath, slag and gas states. Compare calculated trajectories with available measurements.

4. Add ML only where it creates incremental value

Train residual models, soft sensors, parameter estimators or risk classifiers using time-aware validation. Test on unseen heats and operating regimes. Define confidence and out-of-domain logic.

5. Deploy as an advisory Level 2 application

Provide operators and process engineers with state estimates, endpoint predictions, recommended actions, heat replay and explanation. Measure real operating performance before increasing automation authority.

6. Integrate and sustain

Connect approved interfaces to Level 1, historian, LIMS or MES. Establish model version control, retraining rules, data-quality monitoring, cybersecurity review and ownership by the plant team.

Figure 8 — A focused pilot can evolve from offline engineering model to advisory Level 2 and, where justified, deeper plant integration.

Open ExtractMet’s SmartMelt dynamic steelmaking platform

Why a focused pilot is often the best first step

The most effective starting point is not “digitalize the BOF.” That is too broad.

A better question is:

Which repeated BOF decision is currently expensive, variable and measurable?

Examples could include endpoint carbon and temperature, reblow reduction, flux optimization, slag FeO consistency, oxygen prediction, slopping risk, heat-time variability or adaptation to changing hot-metal / scrap conditions.

Once one use case is validated, the calculation framework can be extended to adjacent decisions. The static model becomes a foundation for the dynamic model. The dynamic model becomes a foundation for soft sensing. The state estimator becomes a foundation for optimization. The validated calculation engine becomes a credible basis for a larger Level 2 or digital-twin roadmap.

Figure 9 — The objective is not to remove metallurgical expertise from the control room. It is to give operators and process engineers a better, more consistent estimate of what the furnace is doing and what action is justified next.

The opportunity: make every heat more explainable, predictable and controllable

The BOF will always be a fast, nonlinear, high-temperature metallurgical reactor. Raw materials will vary. Sensors will have limitations. Steel grades will change. Operators will encounter heats that do not follow the average pattern.

That is precisely why Level 2 process control should not depend on a single model philosophy.

First-principles models provide the metallurgical structure.
Plant data provides reality.
Machine learning provides adaptation.
Dynamic state estimation provides context.
The operator and plant control system provide the decision environment.

When these are designed together, a BOF Level 2 system can become more than an endpoint calculator. It can become a living process-intelligence layer for oxygen steelmaking.

If your BOF shop is working on any of the following, we would be interested in discussing the plant problem:

• static or dynamic BOF endpoint models;
• modernization or replacement of an existing Level 2 calculation layer;
• first-principles + machine-learning hybrid modelling;
• carbon and temperature soft sensors;
• oxygen / flux / coolant optimization;
• slag and FeO control;
• off-gas based dynamic control;
• heat replay, troubleshooting and operator advisory tools;
• Level 2 / digital-twin pilot development; or
• plant-specific steelmaking modelling and process optimization.

Bring us one measurable BOF problem. We can help structure the data, build the metallurgical calculation backbone, add machine learning where it earns its place, validate the model on plant heats and define a practical path toward Level 2 deployment.

Discuss a BOF / LD Level 2 project with ExtractMet

Explore the Online Steel-Plant Model Demo Hub

Technical context and further reading

The following references are useful background for readers interested in the technical basis of static, dynamic and data-driven BOF control. They are cited as industry/research context and not as performance claims for ExtractMet.

1. Dynamic Modeling and Simulation of Basic Oxygen Furnace (BOF) Operation, Processes (2020): https://doi.org/10.3390/pr8040483
2. End-Point Static Control of BOF Steelmaking Based on Wavelet Transform Weighted Twin Support Vector Regression (2019): https://doi.org/10.1155/2019/7408725
3. Whale Optimization End-point Control Model for 260 tons BOF Steelmaking, ISIJ International (2022): https://doi.org/10.2355/isijinternational.ISIJINT-2021-517
4. End-Point Dynamic Control Model for 260 Tons BOF Steelmaking, steel research international (2023): https://doi.org/10.1002/srin.202200872
5. BOF endpoint forecasting via informer architecture for multivariate time series data, Metallurgical Research & Technology (2026): https://doi.org/10.1051/metal/2025119
6. Recent industrial example of integrated BOF Level 2 optimization and dynamic endpoint control: https://www.primetals.com/en/news/advanced-systems-implemented-at-rizhao-steel-enable-full-plant-automation/
7. Recent India BOF project including Level 1 and Level 2 automation: https://www.sms-group.com/en-cn/press-and-media/press-releases/press-release-detail/jsw-steel-dolvi-works-selects-sms-group-for-major-expansion-of-its-steelmaking-facility