For most of the twentieth century, quality in float glass was a matter of experienced operators looking carefully at a ribbon moving fast. That is changing. The industry now has the instrumentation to see what the eye misses, and the compute to act on it.
The first deployments were unglamorous. A camera on the tin bath, a model that predicts deviation before the operator can see it. The returns were not in defect counts so much as in stability: a line that used to drift with the shift now holds its gauge through the night.
Seeing the ribbon
Modern inspection is no longer a single camera. You will find spectral imaging for bubble and inclusion detection, laser triangulation for thickness, and full-field interferometry for roll-out and distortion. Each of these is a useful sensor. The value comes from combining them, and from acting on the combination within a control loop rather than after a shift report.
Machine learning earns its place in three specific jobs. First, classification: separating a real defect from a lens artefact, a scratch from a reflection, at line speed. Second, prediction: forecasting when a furnace condition will push the ribbon out of specification. Third, process discovery: mining years of furnace data for relationships that were never written down, then testing them physically.
Classification is solved. Prediction is useful. Discovery is where the real gains are, because that is the only one of the three that produces something nobody had already thought of.
The control loop is the hard part
Most glass AI programmes do not fail on modelling. They fail on integration. A prediction that arrives forty seconds late, or that cannot be expressed in terms the operator can act on, is a dashboard rather than a control system.
The plants making real progress have a simple rule: every model must have a named human decision attached to it. If nobody can say what the operator does differently when the model is right, the model is not finished.
What it looks like in practice
Across the plants we have visited this year, three patterns repeat. Continuous ribbon thickness correction driven by a closed loop. Bubble detection feeding a furnace-side gas composition adjustment minutes rather than hours later. And a digital twin of the melting end used to rehearse a setpoint change before it is committed to a tank that costs more than the building around it.
None of this is autonomous glassmaking, and the people running it are clear about that. But a line operated with better information is a line that makes more of what it is supposed to make, from the same energy, in the same furnace.
What to watch next
Two things. The first is the falling cost of edge inference, which puts inspection on the line rather than in the control room. The second is the arrival of standardised process data — if glassmakers can share shape without sharing process, the learning curve stops being site-specific and starts being industry-wide.
