AI digital twin for a glass fiber furnace
ML engineer
A virtual sensor that keeps reading underglass temperature when the physical thermocouple fails, at 4.37°F RMSE in the middle of melt.
4.37°FRMSE, middle of melt at 100 TPD
The problem
Molten glass sits at roughly 2,150°F, and its temperature governs fibre diameter, breakage rate, energy cost and quality. The thermocouples measuring it are submerged in that glass: they drift, degrade and fail, which leaves operators blind on the most important variable in the process.
Separately, when the plant changes production rate the furnace takes hours to re-stabilise, so operators react after quality has already drifted rather than before.
The data
Live process historian, twelve months of continuous operation — two-hour averages plus a twenty-minute high-resolution set covering transitions.
44 raw process tags expanded to 110 engineered features: lags 1–12, rolling means and standard deviations, first and second deltas, the oxygen-to-gas stoichiometric ratio, crown thermal gradient and underglass spread.
Signals include four crown thermocouples, three underglass thermocouples as the target, gas and oxygen flow, eight bushing well temperatures, cullet percentage, glass level, furnace pressure and pull rate.
Substantial work went into cleaning: missing channels, drifting sensors, unit inconsistencies between pulls, and physically impossible readings during faults.
Approach
Three connected pieces. A virtual sensor — six gradient-boosted models predicting underglass temperature at the doghouse, the middle of melt and the throat, using only crown thermocouples, gas flow, oxygen flow and cullet percentage. A forecaster producing rolling multi-step temperature trajectories. And a prescriber, an optimiser working backward from a target temperature to a recommended gas and oxygen ramp.
Predict the change, not the value. Early models predicting absolute temperature scored well by learning the mean rather than the dynamics. Reframing the target as ΔT was the single decision that made the system work.
Six models rather than three, because production rate splits the problem into two different physical systems. At higher rate the crown runs hotter while the glass runs cooler — more throughput means less time in the hot zone — and crown and glass temperature are nearly uncorrelated within a single production regime.
Results
RMSE by model, location and production rate:
| Model | Location | Rate | RMSE |
|---|---|---|---|
| UG T04 | Middle of melt (primary) | 100 TPD | 4.37°F |
| UG T06 | Throat | 100 TPD | 3.76°F |
| UG T02 | Doghouse | 100 TPD | 5.26°F |
| UG T04 | Middle of melt | 120 TPD | 5.62°F |
| UG T06 | Throat | 120 TPD | 6.40°F |
| UG T02 | Doghouse | 120 TPD | 7.06°F |
Temperature travels as a wave down the furnace, from doghouse to middle of melt to throat. The strongest predictors were not absolute temperatures but rates of change at neighbouring thermocouples — the model learned to watch the wave coming.
Limitations and what I'd do differently
Physics constrains pre-positioning to roughly 25–30°F of achievable shift in an eight-hour window at safe ramp rates. The rest of the gap closes only when pull rate itself changes, which is a production decision rather than a control one.
The models are fitted per production regime. A rate the plant has not run is outside their range, and nothing in the current setup warns when that happens.
RMSE is reported against held-out operation from the same twelve months. It is not a test against a different campaign, a rebuilt furnace or a different cullet mix, and it should not be read as one.
Artifacts
This work is owned by NOIC. Plant location, flow rates, crown temperatures and production tonnage are operating parameters for a live commercial furnace and are proprietary — they are not published here. Every finding above holds without them.