Polymer quality on the extrusion line with AI — an out-of-spec MFI at the end of the roll is a lost roll.

The problem

The defect is born in minutes — the lab finds out hours later.

On a polymer extrusion line, the change that ruins the roll is not abrupt — it is a slow drift over minutes that the operator does not see because they are looking at another screen, and that the lab does not catch because it samples every hour or at the end of the roll. Three patterns repeat on the plant floor:

1. A change in virgin raw material + rework — the blend of virgin resin and recompound from the line's own scrap slightly changes the viscosity entering the screw. Die pressure creeps up, the product's MFI drifts away. The operator notices once it already shows in the part.
2. Campaign startup transient — switching from HDPE to PP, or moving up a grade within the family, the first meters are scrap "by the book." But how much scrap depends on how well the startup setpoints were guessed — and that depends on which veteran happened to be on shift the last time this grade was run.
3. Geometry/surface defect — gel, streak, bubble, out of dimension — which the operator spots when walking past the line, but which has been passing unseen for minutes by then.

And above all, the data lives in islands: the DCS knows melt pressure and temperature, the LIMS knows the MFI measured in the lab, the MES knows the recipe and the grade, the bypass rheometer (if installed) lives in its own software, and the operator's eye catches the visible defect but nothing gets written down. The knowledge of how to avoid the problem lives in the shift veteran's head — and leaves with them the day they retire.

How it fits the IRIS system

iLEAN does not touch the DCS recipe — it closes the quality loop before the roll is lost.

The problem is not a missing sensor: it is information living in islands and a quality loop that closes hours too late, in the lab. iLEAN acts as the putty that seals the gaps between DCS, LIMS, rheometer, MES and the operator's eye, and puts the veteran's intuition back on the shift's screen.

Edge watches the extruded product in line. Vision identifies the defect. Brain estimates MFI continuously. The agent proposes an adjustment. The operator signs — the DCS executes it.

The iLEAN pieces applied to polymer extrusion quality:

• Edge + Vision — a CNN-equipped terminal with a camera positioned over the extruded product (film, pipe, profile, filament, pellet) at the die exit or the calibrator. It detects surface defects (gel, streak, bubble, contamination, out of dimension) in milliseconds and triggers an actuator (traffic light, diversion to trimming, mark on the reel) before the defect spreads. It works without a network: if the plant loses WiFi, Edge keeps detecting and marking, because on a continuous extrusion line what is critical cannot depend on WiFi.
• Connect + Brain — Connect captures the process variables from the DCS (die pressure, melt temperature, screw torque, throughput), the MFI analysis from the LIMS when it arrives, the bypass rheometer if installed, and the recipe and active grade from the MES. Brain builds a continuous MFI soft sensor trained on your history, and a golden batch per grade. When the estimated MFI drifts from center, the system warns minutes before the lab would see it.
• Agent — cross-references the vision detection, the estimated MFI, the die pressure and the active recipe. When it concludes the drift is real, it proposes the concrete recipe adjustment to the DCS operator: "lower the screw by 2 rpm and raise the barrel center by 3 °C, based on 12 similar campaigns of this grade with this raw material." The operator applies it from their console — the SIS is not touched, the DCS recipe is executed by the operator, the agent proposes with the justification alongside.

See the full IRIS architecture →

Before and after

Hourly lab + operator's eye vs. in-line quality loop

Aspect	Hourly LIMS + attentive operator	With iLEAN Edge + Vision + Brain
Out-of-spec MFI detection	At the end of the roll, in the lab	Minutes after the drift begins, in line
Visual defect detection	The operator's eye, intermittent	Edge CNN over the product, continuous and marked
Campaign startup	The shift veteran's setpoints, variable scrap	Golden-batch setpoints from the last good campaign
The veteran's knowledge	Leaves with them the day they retire	Captured as a permanent capability of the plant
Operator reaction	Personal playbook + intuition	On-screen adjustment proposal with justification
Touches the SIS or the DCS recipe	n/a	Never — the operator validates and executes at their console

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with 2-5 multi-grade extrusion lines (film/pipe/profile/compounding), existing DCS, LIMS with MFI history, possibly a bypass rheometer.
• Edge + Vision + Brain pilot on the Pareto line: the one moving the largest volume or the most sensitive grade. First value expected within a few weeks.
• Reject reduction ≥ 30% in the pilot campaign, plus a shorter campaign startup transient (the transition roll that is typically scrap).
• Indicative payback between 4 and 9 months, dominated by the cost of the rolls/loads lost today to out-of-spec MFI plus the savings from optimized startup.

And the production manager's reasonable doubt

“What if the agent proposes an adjustment and makes things worse?” — hallucination is a problem of free generation, not of anchored tasks. When the AI limits itself to finding the most similar previous campaign and proposing its adjustment, the best models brought error below 1.5% ^[1]. And the proposal never enters the DCS on its own: it appears at the operator's console with the justification (which grade, which raw material, which historical pressure, which adjustment worked), and the operator decides and executes. The three-ring architecture guarantees that the SIS and the DCS recipe stay in authorized hands only — the agent contributes intuition, not decisions over the control loop.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.