Compare
Physics-based vs data-only models for a smelting furnace: why a hybrid is usually best.
A pure machine-learning model can fit the past but cannot explain why. A pure physics model explains but drifts from the furnace you actually run.
| Data-only (machine learning) | Physics-based | Hybrid: physics calibrated on data | |
|---|---|---|---|
| Built from | Historical plant data | Mass and energy balance, thermodynamics, kinetics | Physics first, corrected with plant history |
| Explains why | Rarely | Yes | Yes |
| Outside past operating range | Unreliable | Grounded in physics | Grounded in physics |
| Matches your furnace | On data it has seen | Only after tuning | Calibrated every heat |
| Data needed | A lot, and clean | Inputs and assays | Inputs, assays and heat history |
| Trusted for new recipes | Risky | Better | Best of both |
When each makes sense
- Data-only: narrow predictions inside a well-covered operating range, such as a soft sensor.
- Physics-only: design studies and first estimates before much data exists.
- Hybrid: day-to-day estimates of energy, grade and recovery, and scenarios for recipe changes that go beyond past practice.
Questions to ask of any furnace model
- Does it close the mass and energy balance?
- Does it show a range on each estimate?
- Is it re-calibrated as new heats arrive, and are versions logged?
- Can it explain which driver caused a change?
See mass and energy balance and digital twins in the supply chain.
How Furnace Twin does it
Furnace Twin is a physics model of your furnace (mass and energy balance, slag chemistry, reduction kinetics, an electrical model and furnace zones) calibrated on your heat history and re-calibrated as new heats arrive. It estimates energy, grade and recovery every heat and flags drift with its likely cause, such as basicity, reductant or electrodes. A scenario engine compares reductant, basicity, temperature and feed changes side by side so only the best goes to a short trial on the real furnace. Anyone on shift can ask questions in plain language and get answers that cite records, model runs and literature, and shift handover can be captured by voice in Hindi or English. Every model version is evaluated and logged. Delivery runs baseline, calibrate, then operate, starting with one furnace.
Questions
Is machine learning enough for a furnace model?
For narrow predictions in a familiar operating range it can be. For explaining causes and testing new recipes, a physics model calibrated on plant data is more reliable.
What is a hybrid digital twin?
A physics-based model, such as mass and energy balance and slag chemistry, corrected with plant history so its estimates match the actual furnace.
See it on your own data. Furnace Twin — Ask the furnace, see the proof. Book a 30-minute working session with an engineer.