Artificial intelligence has entered glass manufacturing, but not yet as a fully integrated raw-materials-to-QC intelligence layer. What exists today is better understood as a set of useful, partly connected capabilities: supervisory control, furnace vision, predictive furnace-quality modeling, inspection analytics, digital twins, and maintenance analytics.
Some of these tools already have a credible business case. Others are promising but still difficult to justify unless the plant has a costly, recurring problem: excessive energy use, unstable quality, chronic defects, poor pack-to-melt, high false rejection, slow job changes, or dependence on a small number of expert operators.
From a practitioner’s perspective, the question is not whether AI can be applied to glassmaking. It can. The better question is where it pays.
Where the ROI Is Clearest
The strongest commercial case still belongs to advanced supervisory control rather than fashionable AI.
Systems such as Glass Service’s ESIII are mature examples of model-based supervisory control. Glass Service describes ESIII as supervisory-level software for furnace and forehearth control, with applications including melting, boosting, batch feeding, energy optimization, and process stabilization. Public Glass Service material reports more than 400 installations worldwide, while a 2025 presentation cites energy savings in the range of 2–4%. Those figures should be treated as vendor-reported performance claims, not universal guarantees. Actual savings will depend on furnace condition, instrumentation, operating discipline, baseline control quality, energy cost, and product mix.
Even with that qualification, the economic logic is sound. Energy is measurable. Pull is measurable. temperature stability, job-change recovery, and operator intervention are measurable. If advanced control reduces specific energy, stabilizes the process, or reduces the cost of operating near quality limits, the value can be recurring and significant.
This is a practical warning. Many glass plants should not begin their “AI journey” with AI. If the plant does not yet have reliable instrumentation, stable combustion, useful historian data, consistent inspection feedback, and well-tuned supervisory control, the best return may come from fixing those foundations first.
AI should not be used to decorate weak process control.
Furnace Vision: Turning Observation into Measurement
Furnace vision is one of the more useful developments. Near-infrared imaging and AI-based image analysis are being applied to batch-line position, batch coverage, batch movement, batch islands, bubbling behavior, temperature mapping, fogging, and build-up detection. Glass Service has publicly described AI furnace-vision modules connected to furnace control environments, while AMETEK LAND has published work on machine learning and neural networks for batch tracking and batch-line determination from furnace imaging.
This is more than a camera upgrade. Batch blanket behavior is a furnace-state indicator. It can tell the operator something about charger balance, melting load, heat transfer, batch conversion, foam behavior, and residence conditions.
The ROI is usually indirect but real. Furnace vision pays when it changes decisions: better charger adjustment, earlier detection of batch carryover, better tuning of combustion or boosting, faster diagnosis of instability, and more objective furnace reviews. It is much weaker when the plant merely archives images without converting them into operating information.
Predictive Quality: The Right Direction, Still Emerging
Predictive furnace-quality systems are among the most important developments because they address the central fact of glassmaking: quality is delayed.
A defect seen now may have originated from a material, thermal, hydraulic, redox, refining, or operating condition many hours or even days earlier. Predicting that delayed consequence is more valuable than simply holding a few temperatures steady.
CelSian’s Celfos is a public example of this direction. CelSian and trade press reporting describe Celfos as an AI-based predictive quality system that uses furnace data and CFD-derived insight to predict future glass quality up to 48 hours ahead, including risks associated with defects such as cords, seeds, and bubbles. This is a vendor-described capability and should be treated as an emerging application rather than a generally proven industry standard.
The potential ROI is large. Avoiding one serious defect episode, customer complaint, pull reduction, or extended reject period can justify considerable analytical effort. But predictive quality has to meet a high credibility standard. It must show what changed, when it changed, what defect family is at risk, what historical cases resemble the present trajectory, and what corrective actions are plausible.
A prediction without an actionable explanation is just another alarm.
Inspection AI: Close to the Money
AI-assisted inspection has a clearer short-loop value proposition. Systems that improve defect classification, reduce false rejection, identify section or mould drift, and communicate actionable information to hot-end personnel can directly improve pack-to-melt.
IRIS Inspection Machines, for example, describes iBot as an AI-powered assistant connected to its inspection machines, intended to analyze production data, support defect classification, provide real-time alerts, and help operators respond to process drift. As with all vendor claims, actual value depends on installation quality, defect mix, operator response, false-reject baseline, and integration with plant routines.
This area is close to the money. False rejects are costly because the plant has already paid to batch, melt, condition, form, anneal, inspect, and handle glass that is not sold. Better inspection analytics can reduce waste, shorten response time, and improve defect discipline.
But inspection AI has limits. It can often say that a defect family is increasing, where it appears, and which machine, section, mould, or timing pattern is involved. It usually cannot, by itself, prove that the cause was a sand PSD shift, soda ash density change, batch moisture excursion, charger imbalance, furnace convection change, or forehearth lag many hours upstream.
Inspection AI is useful. It is not, by itself, a full root-cause system.
Digital Twins: Powerful, but Not Automatically Profitable
Digital twins and CFD-linked operating models are valuable, especially for furnace studies, commissioning, pull changes, energy analysis, troubleshooting, and what-if simulations. They can provide physical structure that pure machine learning lacks.
AGC, for example, announced in 2023 that it had developed CADTANK Online Computation and Optimization Assistant, or COCOA, as a digital-twin technology for the glass melting process. AGC described the system as integrating an online simulator with a digital prototyping tool, with operational verification scheduled at AGC float furnaces. The stated purpose was to provide better insight into furnace state and allow more rapid study of operating conditions.
This is promising, but a digital twin does not automatically produce ROI. It must be maintained, synchronized, validated, and used in operating decisions. A digital twin that supports furnace studies and process decisions can be valuable. A digital twin that becomes a specialist engineering tool used only occasionally is harder to justify.
The best future systems will likely be hybrid: physics-based models to describe what must be true, plant data to describe what is actually happening, and AI to learn patterns that are not captured cleanly by first principles.
The Missing Bridge: Raw Materials to QC
The least developed area remains the most important one: connecting raw materials to final quality.
Modern batch plants can dose accurately. Historians can collect thousands of tags. Inspection systems can classify defects. Furnaces can be modeled. Cameras can track batch behavior. Yet the industry still lacks a widely available, standard commercial layer that can coherently say:
A raw-material or batch-plant change altered the material entering the furnace; that changed melting, refining, convection, residence history, or conditioning; and that later appeared as a measurable defect population at QC.
That is the prize.
To do this, AI needs more than data. It needs plant structure. It needs material genealogy, batch-lot tracking, silo state, charger behavior, furnace residence logic, pull history, cullet changes, forehearth lag, forming context, annealing history, and defect taxonomy.
Without that structure, AI becomes an expensive correlation engine. With it, AI can become a delayed-causality engine.
The Sensible First Step: Offline AI, Online Advisory
For most glass plants, the near-term opportunity is not autonomous AI control. It is offline AI trained on historical plant behavior, then deployed online as an advisory layer.
This system would not directly move the furnace. It would monitor whether the plant is still inside the operating envelope historically associated with good glass. It would ask whether raw materials, batch-plant behavior, furnace response, forehearth control, forming behavior, and QC trends remain coherent with the desired quality outcome.
This is not a replacement for MPC. It sits above MPC. MPC asks how to move manipulated variables to hold targets. The AI advisory layer asks whether the current targets, assumptions, and process state still make sense.
The Practitioner’s ROI Hierarchy
The highest-confidence ROI is in supervisory control, energy optimization, boosting optimization, forehearth stability, inspection improvement, and false-reject reduction — all measurable.
The next tier is furnace vision and AI-assisted process observability. These pay when they improve response time and reduce misdiagnosis.
The emerging tier is predictive quality and digital twins. These may be highly valuable, but their ROI depends on accuracy, explanation, and the plant’s ability to act.
The strategic but least mature tier is raw-materials-to-QC intelligence. This may ultimately be the largest opportunity, but it requires the hardest foundation: time-aligned material genealogy and defect-causality ontology.
Conclusion
AI in glassmaking is real, but its value is uneven. The clearest returns are found where the application is close to measurable loss: energy, yield, rejects, downtime, job-change recovery, and inspection performance.
The larger opportunity is not another dashboard. It is an intelligence layer that understands how material becomes glass over time, how process history becomes quality, and how early deviations become later defects.
The machine learning is not the hardest part. The hard part is giving the machine a faithful representation of the plant.
That makes the difference between AI as analytics theatre and AI as a practical tool for glassmaking.