Epoxy resin control with AI — a batch out of spec is not an incident, it is a customer lost.

The problem

The recipe is always the same. The batch never is.

Anyone running a resin plant knows it: the nominal recipe is stable, but the real batch depends on things the recipe does not capture.

1. Raw material changes from supplier to supplier — bisphenol-A can arrive with a different share of the o,o'-isomer, epichlorohydrin picks up moisture depending on how it was shipped, and the catalyst performs differently between lot A and lot B from the same manufacturer. The incoming analysis is in the LIMS, but it rarely reaches process control in time.
2. The reactor is not always the same reactor — a heat exchanger slowly losing efficiency, a worn agitator, a dosing valve that meters less precisely than when it was new. Today's exotherm curve is not the one from six months ago, even if the SCADA setpoint has not changed.
3. The customer spec does not forgive — an EEW (epoxy equivalent weight) outside the ±10-unit window is a rejected batch. A viscosity outside the agreed Brookfield range means renegotiation. A MIL-PRF standard or an aerospace customer's TDS does not accept a letter of explanation: either it complies or it does not.

The lab measures at the end of the batch and returns the number hours later. By then the batch is made and, if you are lucky, it goes to rework; if you are not, to hazardous waste. The problem is not the recipe — it is that the classic system always arrives late, and every late batch is a cost that eats into the quarter.

How it fits the IRIS system

iLEAN does not change your reactor — it adds the intelligence that was missing on top of the data you already generate.

The problem in epoxy resin control is not a lack of information: it is information that lives on islands — the LIMS knows how the raw material came in, the SCADA knows what is happening inside the reactor right now, the ERP knows which customer is waiting for that batch and to which spec, but none of the three talks to the next one at the speed the batch demands. iLEAN acts as the putty that fills those gaps, without asking you to change your SCADA, your LIMS or the plant.

Edge sees viscosity and temperature in line. Connect reads the LIMS analysis and the customer spec wherever they live. The agent cross-references them against the expected curve and, if something does not add up, holds the batch before packing. The person signs — never the other way round.

The three iLEAN pieces applied to epoxy resin control:

• Edge — a terminal with machine vision (CNN) and in-line signal capture: in-line Brookfield viscometer, refractometer, reactor thermocouples. It reads the exotherm curve in real time and, if the slope drifts away from the trained model, it triggers the status light and notifies the operator before the batch is lost. It works with no network. If the plant loses WiFi, Edge keeps reading and holding batches — what is critical cannot depend on connectivity.
• Connect — captures the raw-material analysis whether it comes from a modern LIMS, from the lab's spreadsheet or from the sheet the R&D manager updates every Monday. And it also captures what arrives from outside (the aerospace customer's TDS, a change in the RoHS standard, an email from the supplier flagging a bisphenol lot with a different isomer profile) at second zero, with nobody having to forward anything.
• Agent — cross-references the batch recipe, the raw-material analysis, the real reactor curve and the customer spec. If there is a deviation, it does not send an email at 10 pm: it holds the batch and alerts the quality manager through whichever channel they use. It also generates the conformity dossier — which raw material came in, how the reactor behaved, what complied and what did not — ready for an auditor or a customer. The person validates and signs; the batch does not ship on its own.

See the full IRIS architecture →

Before and after

Manual control vs. cross-referenced control with iLEAN

Aspect	Manual control + LIMS at the end	With iLEAN Edge + Connect + Agent
Detecting a viscosity deviation	When the lab returns the number (hours)	In line, during the batch
Anticipating the exotherm	Reaction once the limit is crossed	Warning when the slope predicts the rise
Cross-check with raw-material analysis	The operator checks the LIMS if there is time	Connect brings it to the reactor panel automatically
Customer spec compliance (TDS/MIL-PRF)	Post-batch verification, rework if it fails	Real-time validation, hold before packing
Conformity dossier per batch	Rebuilt by hand, days	Generated automatically, ready to sign
Operation with no network	n/a	Edge keeps running on the panel's power

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.

• Mid-sized epoxy resin plant (liquid and solid), 2-3 batch reactors, a mix of standard resins and customer-spec grades (aerospace, electronics, composites).
• Edge pilot on one reactor (in-line viscometer/refractometer + thermocouple capture + integration with the LIMS and the customer spec). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often batches have been rejected or reworked in recent years and on the average cost of an out-of-spec batch in your product mix.
• The hard lever is a single rejected batch avoided: lost raw material, reactor energy, tank occupancy, delay to the customer. The expected reduction in in-line rejects is ≥ 30% over your current baseline.

And the technical director's reasonable doubt

“What if the AI gets an exotherm prediction wrong and stops a batch that was running fine?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares a real curve against a model trained on your own reactor, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN proposes and the person signs. The three safety rings exist precisely for this — the agent does not touch critical OT.

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