AI for Steel & Iron Making Process Control

Blast Furnace Optimization

The blast furnace is arguably the most complex reactor in industrial metallurgy — a countercurrent heat exchanger, chemical reactor, and moving bed packed column, all operating simultaneously across a 30m height with residence times of 6-8 hours. Classical control relies on operator experience, rule-of-thumb burden calculations, and delayed feedback (hot metal composition is known only at tapping, 4-6 hours after the corresponding charge was loaded).

AI changes the feedback loop from reactive to predictive.

Open data/blast-furnace-data.csv in the code panel. Each row represents a 15-minute operational snapshot: burden composition (Fe content, basicity, coke rate), blast parameters (volume, temperature, moisture, oxygen enrichment), gas analysis (top gas CO, CO2, H2), thermal indicators (ETD — Estimated Thermal Datum), and hot metal composition at the next tap (Si, S, temperature).

Silicon Prediction

Hot metal silicon is the primary indicator of furnace thermal state. High Si (> 0.8%) means the furnace is running hot — wasting coke and producing off-spec iron. Low Si (< 0.3%) risks chilling, skull formation, and irregular descent. Operators target 0.4-0.6% Si for BOF feed.

The challenge: Si at tapping reflects conditions 4-6 hours earlier. By the time you see high Si, the root cause is already 4 charges behind you.

A predictive model trained on 2+ years of operational data forecasts Si at next tap from current blast parameters and gas analysis:

Features (real-time): blast_volume, blast_temp, O2_enrichment,
  moisture_injection, top_gas_CO, top_gas_CO2, top_gas_H2,
  stock_line_probe_temps (8 points), ETD
Features (lagged): burden_Fe_content[-4h], coke_rate[-4h],
  basicity[-4h], Si[-1tap], Si[-2tap]
Target: Si_next_tap
Model: LSTM (captures temporal dependencies in sequential charges)

At Tata Steel Jamshedpur's H Blast Furnace (India's largest at 4,060 m³), an LSTM-based silicon predictor achieved MAE of 0.06% Si — sufficient to give operators a 2-hour advance warning of thermal excursions. This translated to a 2.5 kg/THM reduction in coke rate over 6 months.

Coke Rate Reduction

Coke is the single largest cost in ironmaking — typically 40-45% of hot metal cost. Every kg/THM (per tonne of hot metal) reduction in coke rate saves approximately ₹3-4/THM, or ₹30-40 crore annually for a large furnace producing 10,000 THM/day.

AI-assisted coke rate optimization works by:

Burden Distribution

At JSW Dolvi's blast furnaces, a digital twin of the burden profile uses DEM (Discrete Element Method) simulation calibrated with stock-line probe data to predict radial ore/coke distribution after each charge. The model recommends charging matrix adjustments to maintain the target gas flow pattern — central flow for stability, wall flow for hot metal temperature control.

BOF Endpoint Prediction

Open data/bof-heat-data.json — each record represents a BOF heat with charge mix (hot metal weight, scrap weight, flux additions), blowing profile (oxygen flow rate, lance height over time), sublance measurements, and endpoint chemistry/temperature.

Carbon and Temperature at Turndown

The BOF converts hot metal to steel in 15-20 minutes. The critical endpoint: carbon < 0.04% (for most grades) and temperature within ±10°C of target. Miss the carbon and you need reblow (delays, refractory wear). Miss the temperature and you have casting problems downstream.

The AI dynamic model uses real-time off-gas analysis (CO, CO2 flow rates from the gas cleaning plant) to track decarburization progress continuously. The cumulative carbon removal is:

C_removed(t) = ∫₀ᵗ [Q_CO(τ) × (12/28) + Q_CO2(τ) × (12/44)] dτ / W_steel

Combined with the initial carbon (from hot metal composition) and scrap carbon, the model tracks bath carbon in real-time. An LSTM trained on 10,000+ heats learns the decarburization kinetics — the shift from slag-metal reaction control to mass transfer control at ~0.3% C — and predicts endpoint carbon 3-5 minutes before turndown.

At SAIL Bhilai's SMS-3 (3 × 150T BOFs), AI-assisted endpoint control reduced reblow frequency from 8% to 2.5% and hit-rate (carbon + temperature both on target) improved from 78% to 91%.

Continuous Casting: Breakout Prediction and Defect Minimization

Open data/continuous-casting-data.csv — each row captures casting parameters: speed, superheat, mould level, mould thermocouple readings (36+ points), SEN (Submerged Entry Nozzle) position, water flow rates, and downstream quality (slab surface defects, internal cracks from macro-etch).

Breakout Prediction

A breakout — liquid steel breaking through the solidified shell in the mould — is the most catastrophic event in continuous casting. It destroys equipment, halts production for 12-24 hours, and costs ₹5-10 crore per event. Breakouts are preceded by characteristic thermocouple signatures: a "sticker" breakout shows sequential temperature drops as the sticking point moves down the mould.

Traditional breakout detection systems (BDS) use rule-based logic on thermocouple patterns — they catch 70-80% of stickers but generate 5-10 false alarms per genuine event. False alarms trigger speed reductions that reduce caster throughput.

An AI-based BDS trained on historical breakout events and normal casting data:

Input: mould_thermocouple_matrix (36 TCs × 60 seconds sliding window)
       casting_speed, superheat, steel_grade, mould_oscillation_params
Target: breakout_probability (0 to 1)
Model: 1D-CNN on thermocouple time series + gradient boosted meta-features
Threshold: alarm at P > 0.85

At RINL Vizag's caster, the AI-based BDS caught 95% of breakout precursors with < 1 false alarm per month — a dramatic improvement over the rule-based system's 5-10 false alarms per month. Over two years, zero breakouts occurred.

Superheat Control

Superheat (temperature above liquidus) at the mould directly affects solidification structure. Too high (> 30°C) causes coarse columnar grains and centreline segregation. Too low (< 10°C) risks nozzle clogging from frozen steel.

AI predicts superheat at the mould from ladle temperature, tundish geometry, casting speed, and steel grade. The model accounts for heat losses in the ladle-tundish-mould system and recommends ladle furnace aim temperature to hit the target 15-25°C superheat at the mould.

Surface and Internal Defect Prediction

Surface cracks (longitudinal, transverse, star cracks) and internal defects (centreline segregation, off-corner cracks) depend on the interplay of superheat, casting speed, mould taper, cooling rates, and steel composition — particularly carbon content in the peritectic range (0.09-0.17% C) where δ→γ transformation causes volume contraction.

A predictive model flags high-risk parameter combinations before they produce defective slabs:

High-risk indicators:
  - Peritectic carbon (0.09-0.17%) + superheat > 25°C → longitudinal cracks
  - Casting speed change > 0.2 m/min per minute → shell thinning
  - Mould level fluctuation > ±5mm → slag entrapment
  - Asymmetric thermocouple pattern → SEN misalignment

Digital Twins for Process Metallurgy

A digital twin is not a dashboard — it is a physics-informed model that runs in parallel with the real process, continuously calibrated by sensor data, and capable of "what-if" simulations.

For Indian integrated steel plants, the most valuable digital twins are:

The common platform challenge in India: sensor reliability. Blast furnace thermocouples fail, caster mould level sensors drift, BOF off-gas analysers need frequent calibration. Any digital twin deployment must include sensor health monitoring — the AI model that detects when its own inputs are unreliable.

Key Takeaways