Why blast furnace decarbonisation needs physics-constrained AI

Machine learning has a reputation problem in ironmaking: models that look accurate in testing and fail the moment conditions change. For decarbonisation, that weakness is decisive.

The data you need doesn't exist yet

A purely data-driven model learns the relationships present in historical operation. But decarbonising a furnace means doing new things — a different ore blend, more hydrogen-rich gas at the tuyeres, a lower fuel rate. By definition, the history doesn't contain those conditions. A model trained only on the past is being asked to extrapolate exactly where it is least reliable.

The inside of the furnace can't be measured

Even the best-instrumented furnace measures its boundaries: top gas, shell temperatures, tuyere conditions, hot metal. The phenomena that decide whether a change works — the cohesive zone, gas distribution, the thermal reserve zone — are inside, out of reach of sensors. A model that only fits measured data cannot tell engineers what is happening where it matters.

Physics fills both gaps

Mechanism models — CFD models that solve heat, mass and momentum balance across the furnace — remain valid when operation changes, because physics doesn't change. They also compute the internal states that can't be measured. Their drawback has been speed and usability.

Physics-constrained AI combines the two: a mechanism-model database built from CFD constrains the AI, so it can respond quickly to real operating data while staying consistent with physics. That is the approach behind SIC™ BF Smart Agent X.

The practical test

For a plant engineer, the test is simple: when the model recommends a change, can you see why — and does the answer still make sense under conditions the furnace has never seen? Physics-constrained AI is built to pass that test.