The Electric Arc Furnace Needs More Than Automation: It Needs a Smarter Level 2

How hybrid first-principles + machine-learning models can turn EAF data into real-time decisions on energy, melting, slag, chemistry and endpoint control

Electric Arc Furnace steelmaking is becoming one of the central production routes in the transition toward more flexible and lower-carbon steelmaking. But the modern EAF is also becoming harder to operate consistently.

The furnace may process a changing mixture of scrap, DRI, HBI, pig iron or hot metal. Scrap density and chemistry vary. DRI metallization and carbon change. Power input is not perfectly converted into useful bath energy. Oxygen, burners and injected carbon add chemical energy while also changing slag and metal chemistry. Arc stability, foamy slag, air ingress, off-gas conditions, refractory state and operator actions all influence the heat.

That means the most important operating question is no longer simply:

“How many kWh per tonne should this heat receive?”

The more useful question is:

“What is the actual state of this heat right now—and what should we do next?”

That is the role of a plant-specific Level 2 process-control and decision-support system.

At ExtractMet Private Limited, our approach to EAF Level 2 development is built around a hybrid modelling philosophy: combine first-principles metallurgy with machine learning and plant data so that the model remains physically meaningful while learning the behaviour of the specific furnace.

Why EAF operation is difficult to optimize heat after heat

An EAF looks simple when reduced to its main inputs—metallic charge, electrical power, oxygen, carbon, fuel and fluxes—but the internal process state is only partially measured.

Operators do not continuously measure every important variable. During much of the heat, the plant may not directly know:

• how much of the metallic charge has actually melted;
• the true bath temperature before a measurement becomes available;
• the instantaneous carbon level of the bath;
• the mass and composition of the slag;
• the actual FeO level and foaming potential;
• how much electrical and chemical energy has reached the metal;
• how much energy is leaving through off-gas, water-cooled panels and other losses;
• how the changing slag–metal–gas reactions are affecting yield and endpoint;
• whether the current heat is ahead of, or behind, the expected melting and refining trajectory.

Traditional automation can execute sequences and regulate equipment extremely well. But the higher-level decision—what the process means and what action is optimal next—requires a process model.

That is why modern EAF automation is increasingly moving toward dynamic metallurgical models, real-time mass and energy balances, advanced state estimation, soft sensors and data-driven adaptation.

Level 1 runs the equipment. Level 2 should understand the heat.

A useful way to separate the roles is:

Level 1 automation controls fast equipment functions and regulatory loops: electrode regulation, power switching, valves, burners, oxygen and carbon devices, material handling, hydraulic systems, interlocks and sequence execution.

Level 2 process control sits above those functions. It uses process data, material information, laboratory results, models and optimization logic to estimate the furnace state, predict the endpoint and recommend—or eventually execute—higher-level operating decisions.

For an EAF, that may include decisions such as:

• What metallic charge mix is most economical for the target grade?
• How much energy should be planned for this heat?
• What oxygen, carbon and burner strategy is appropriate?
• Is melting progressing faster or slower than expected?
• Is the bath approaching flat-bath conditions?
• Is the slag condition consistent with the desired foaming practice?
• Should DRI feed rate, power, oxygen or carbon injection be adjusted?
• What temperature and carbon endpoint are now predicted?
• When should a measurement be taken?
• Is the heat likely to require corrective additions or extra power?
• Which operating pattern is repeatedly associated with yield, energy or quality losses?

A Level 2 system becomes valuable when it answers these questions in a way that is plant-specific, auditable and actionable.

The first layer: a static model that plans the heat

The static model works before, or at the beginning of, the heat.

Its purpose is to establish a physically consistent plan from the known starting conditions and production target.

Typical inputs can include:

Metallic charge

Scrap categories, DRI/HBI, pig iron, hot metal, return scrap and their respective chemistry, temperature, metallization, carbon and gangue.

Fluxes and additions

Lime, dolomite, carbon, ore, alloys and other additions.

Furnace information

Nominal heat size, heel practice, transformer and power limits, oxygen and burner capacities, tapping practice and operating constraints.

Production target

Steel grade, tapping mass, carbon, temperature, residual limits, slag requirements and downstream constraints.

The static calculation can then perform charge, elemental and energy balances to determine an initial operating strategy.

Representative outputs may include:

• optimized charge mix;
• expected metallic yield;
• electrical-energy requirement;
• oxygen, fuel and injected-carbon targets;
• flux requirement and predicted slag quantity;
• expected slag basicity and FeO range;
• expected melting and refining demand;
• predicted tapping mass, temperature and chemistry;
• approximate process-time or energy milestones.

The advantage of this approach is that the heat does not begin with a generic recipe. It begins with a heat-specific plan.

For plants using significant DRI or HBI, this becomes especially important because metallization, gangue, carbon, temperature and feeding practice materially change the energy and slag balance.

The second layer: a dynamic model that follows the heat in real time

A static plan is only the starting point. Real furnaces do not follow the plan exactly.

A dynamic model therefore advances through the heat in small time steps and updates the estimated furnace state whenever new energy, material, measurements or operating events occur.

At each time step, the model can reconcile:

• electrical power and accumulated electrical energy;
• oxygen flow;
• natural gas or other burner fuel;
• injected carbon;
• DRI or continuous metallic feed;
• flux and alloy additions;
• slag-metal reactions;
• decarburization and oxidation;
• combustion and post-combustion assumptions;
• sensible and reaction heat;
• melting of the solid charge;
• off-gas generation;
• thermal losses;
• temperature or chemistry measurements when available.

The model then updates the estimated state of the furnace.

A useful dynamic EAF state may include:

Metal phase

Liquid steel mass, remaining solid metallics, melt fraction, bath temperature, carbon and selected alloying or tramp elements.

Slag phase

Slag mass, temperature, basicity, FeO/MnO and other major oxides, plus indicators related to slag foaming.

Gas phase

CO, CO₂, H₂, H₂O, N₂ and hydrocarbons where relevant, together with gas volume, chemical energy and post-combustion behaviour.

Energy state

Electrical energy, chemical energy, sensible heat, reaction heat, melting demand and estimated losses.

The result is not simply another dashboard. It is a continuously updated virtual representation of the heat.

ExtractMet’s SmartMelt dynamic steelmaking platform is designed around this time-step philosophy: material, elemental and heat balances are updated as operating actions occur, creating a foundation that can be adapted for engineering studies, training, what-if analysis and plant-specific Level 2 or digital-twin development.

Why first-principles modelling alone is not enough

A first-principles model gives the Level 2 system its metallurgical backbone.

It can enforce:

• overall mass conservation;
• Fe, C, Si, Mn, P, S and other elemental balances;
• oxygen balance;
• slag formation;
• reaction stoichiometry;
• heat of reactions;
• sensible heat;
• melting and heating requirements;
• gas generation;
• physically meaningful limits.

This matters because an unconstrained data model can produce a numerically accurate fit while violating metallurgy outside its training range.

But a purely fundamental model also has limitations.

Some parameters are difficult to know precisely in a production furnace:

• effective arc-to-bath energy efficiency;
• changing thermal losses;
• reaction kinetics;
• slag foaming behaviour;
• carbon injection efficiency;
• oxygen utilization;
• post-combustion efficiency;
• scrap geometry and melting behaviour;
• air ingress;
• refractory and panel heat losses;
• sensor bias;
• unrecorded operator actions.

These uncertainties are exactly where plant data becomes valuable.

Why machine learning alone is not enough

Machine-learning models can identify complex correlations between historical inputs and outputs. They are useful for soft sensing, anomaly detection, endpoint prediction, residual correction and pattern recognition.

However, a black-box model trained only on historical data can struggle when:

• raw materials change;
• a new grade or operating window appears;
• the furnace undergoes a major maintenance change;
• the production route changes from scrap-rich to DRI-rich;
• sensors drift;
• operating practice moves outside the historical envelope.

For process control, the strongest architecture is often not physics versus AI.

It is physics plus AI.

The hybrid approach: let physics define the process, and let data teach the model the plant

In a hybrid Level 2 system, the first-principles model provides the conserved balances and process structure. Machine learning is then used selectively where it adds measurable value.

Examples include:

Residual correction

If the physics model systematically over- or under-predicts temperature, energy, carbon or endpoint under identifiable operating conditions, an ML model can learn the residual.

Adaptive efficiency factors

Electrical efficiency, oxygen utilization, carbon recovery or thermal-loss parameters can be adjusted from validated plant history.

Soft sensors

Signals from electrical, mechanical, gas, acoustic, vibration or off-gas systems can be combined to infer difficult-to-measure states such as melting progress or endpoint risk.

Heat classification

Historical heats can be grouped into operating regimes so the model selects more appropriate parameters or strategies.

Anomaly detection

The system can identify heats that are behaving differently from normal patterns and flag the operator before the deviation becomes expensive.

Continuous model improvement

Validated production results can be used to update model parameters in a controlled way rather than relying on permanent manual tuning.

The key principle is that machine learning should strengthen the metallurgical model—not bypass it.

What should the operator actually see?

A successful Level 2 system should reduce complexity for the operator.

The screen does not need to expose hundreds of equations. It should convert the model into clear operating guidance.

Depending on the plant, the interface may show:

• melt progress;
• predicted bath temperature;
• predicted carbon;
• slag status;
• net energy delivered to the charge;
• cumulative electrical and chemical energy;
• DRI feed performance;
• current and predicted endpoint;
• deviation from the planned heat trajectory;
• recommended power, oxygen, carbon or feed-rate adjustment;
• estimated time to next event or tapping;
• alarm or confidence status;
• comparison with similar successful heats.

For engineers and management, a second layer can provide deeper analytics: energy intensity, yield, electrode consumption, refractory indicators, charge-cost performance, tap-to-tap consistency, model error and heat-to-heat benchmarking.

Where the business value comes from

The purpose of Level 2 is not to add another software system. It is to improve decisions that already cost the plant money every heat.

A plant-specific system can be developed around targets such as:

• lower electrical-energy consumption;
• better coordination of electrical and chemical energy;
• more stable power-on and tap-to-tap time;
• improved metallic yield;
• reduced oxidation losses;
• optimized DRI feed rate;
• more consistent foamy slag practice;
• lower carbon, oxygen, flux or alloy overuse;
• reduced electrode and refractory stress;
• tighter endpoint temperature and chemistry;
• fewer corrective actions;
• improved operator-to-operator consistency;
• better visibility of emissions-related operating indicators.

The size of the benefit depends on the plant baseline, raw materials, operating constraints, instrumentation, data quality and implementation scope. A credible project should therefore begin with measured baseline KPIs and validate improvements against production data.

A practical deployment path: from historical data to real-time Level 2

The safest and most effective route is usually incremental.

1. Define the decision problem

Start with a small number of high-value questions.

For example:

• Can endpoint temperature be predicted reliably before the first measurement?
• Can DRI feed be dynamically coordinated with net available energy?
• Can power, oxygen and carbon profiles be optimized heat by heat?
• Can slag and melting progress be estimated better?
• Can the model identify why high-energy or low-yield heats occur?

2. Audit and map the data

Typical sources include PLC/DCS tags, Level 1 events, power and electrode data, oxygen/carbon/fuel flows, weigh systems, scrap and DRI records, laboratory chemistry, temperature measurements, off-gas data, delays and production results.

The data must be time-aligned and validated before model training.

3. Build the static first-principles model

Develop the heat-start mass, elemental and heat balance.

This becomes the benchmark for charge planning and the initial state for the dynamic model.

4. Build the dynamic calculation engine

Advance the heat through time and update the metal, slag, gas and energy inventories as events occur.

5. Calibrate against historical heats

Estimate uncertain parameters and test the model against operating campaigns representing the real furnace range.

6. Add machine-learning modules only where justified

Use ML for residuals, soft sensing, adaptation, prediction or classification after the physical core is functioning.

7. Run in shadow mode

Allow the system to calculate in real time without influencing operations. Compare predictions against measurements and outcomes.

8. Move to operator advisory

Once validated, provide recommendations and confidence indicators to operators.

9. Integrate with the automation hierarchy

Where the plant, cybersecurity framework and operating philosophy permit, selected Level 2 outputs can be integrated with Level 1 or supervisory systems.

This staged approach builds trust because every step can be tested before the next level of automation is introduced.

Not every EAF needs the same Level 2

One of the biggest mistakes in industrial digitalization is assuming that the same model can simply be copied from one furnace to another.

An EAF using 100% scrap is not the same modelling problem as a furnace using 70% DRI.

A furnace with continuous DRI feeding behaves differently from a bucket-charged furnace.

A stainless or alloy-steel EAF has different priorities from a carbon-steel mini-mill.

Electrical network constraints, transformer size, oxygen hardware, burner system, carbon injection, slag practice, tap weight, heel practice, raw-material quality and downstream requirements all matter.

Therefore the Level 2 must be configurable around the actual steel shop.

ExtractMet’s digital-twin portfolio reflects this philosophy: build process models around the decisions, data and constraints of the specific operation, with deployment options ranging from offline engineering and training tools to operator advisory and Level 2 integration.

What ExtractMet can develop for an EAF steel shop

A customized EAF program can include some or all of the following modules:

Static heat model

Charge optimization, mass and elemental balances, energy plan, slag plan, oxygen/carbon/flux targets and endpoint forecast.

Dynamic process model

Time-step evolution of melting, bath, slag, gas and energy state.

Hybrid ML layer

Residual prediction, model adaptation, soft sensors, heat classification and anomaly detection.

Charge-mix optimizer

Cost, chemistry, residuals, yield, energy and available inventory constraints.

Dynamic DRI/HBI feed optimizer

Feed-rate guidance based on melting capacity, energy state, slag condition and endpoint target.

Endpoint predictor

Temperature, carbon, steel mass, slag condition and tapping readiness.

Operator HMI

Heat trajectory, predicted state, recommended action, confidence and alarms.

Heat replay and troubleshooting

Reconstruct historical heats and identify root causes of high energy, long process time, low yield or endpoint deviations.

Model-management layer

Parameter versioning, validation status, model error tracking and controlled recalibration.

Digital-twin / scenario module

What-if studies for new charge mixes, DRI quality, oxygen practice, transformer operation, burner strategy, green-power scenarios and decarbonization studies.

You can explore ExtractMet’s SmartMelt platform, digital-twin portfolio, model catalogue and online model demo hub to see the modelling philosophy across iron and steel processes.

The opportunity: convert furnace data into metallurgical intelligence

Many EAF shops already generate large amounts of data.

But data alone does not optimize a furnace.

The value comes when plant data is connected to:

metallurgical understanding → real-time state estimation → prediction → decision → measured improvement.

That is the gap a well-designed Level 2 system can close.

The objective is not to replace experienced operators. It is to give them a continuously updated, heat-specific calculation layer that can see patterns across thousands of signals and historical heats while still respecting the physical laws of steelmaking.

For plants pursuing higher DRI use, more variable scrap, tighter energy targets, greater automation, green-steel production or operator-independent consistency, this becomes increasingly important.

Interested in developing a plant-specific EAF Level 2 system?

ExtractMet can work with steel plants, technology providers and engineering organizations to develop and validate static, dynamic and hybrid AI/ML-assisted EAF process-control models tailored to the furnace, charge mix, data environment and operating objectives.

Explore ExtractMet: https://www.extractmet.com/?utm_source=medium&utm_medium=article&utm_campaign=eaf_level2

Explore SmartMelt: https://www.extractmet.com/smartmelt.html?utm_source=medium&utm_medium=article&utm_campaign=eaf_level2

Digital-twin portfolio: https://www.extractmet.com/digital-twins.html?utm_source=medium&utm_medium=article&utm_campaign=eaf_level2

Model Demo Hub: https://www.extractmet.com/Models_SteelPlant.html?utm_source=medium&utm_medium=article&utm_campaign=eaf_level2

Project enquiries: info@extractmet.com

Suggested Medium tags

Electric Arc Furnace · Steelmaking · Process Control · Digital Twin · Artificial Intelligence

Suggested SEO description

*How hybrid first-principles and machine-learning models can create a plant-specific Level 2 process-control system for Electric Arc Furnace steelmaking—covering charge planning, dynamic state prediction, energy, slag, chemistry, DRI feeding and endpoint control.*

Suggested social-post teaser

Your EAF already generates data. The bigger opportunity is to turn that data into a live metallurgical state estimate.

This article explains how static models, dynamic mass-and-energy balances and machine learning can work together in a plant-specific Level 2 system for EAF steelmaking.

Read more and explore ExtractMet’s steelmaking modelling platforms: https://www.extractmet.com/?utm_source=medium&utm_medium=article&utm_campaign=eaf_level2