Chemical recipe optimization with AI — the raw material changes, so the recipe should change with it.

The problem

The process manager can see it coming — but never gets there in time.

In a batch chemical plant, raw material never enters the reactor in the same condition as last time. The supplier changes, the lot from the same supplier changes, the storage humidity changes, the marginal purity changes — and the DCS recipe is still the one that was tuned months ago during a commissioning exercise.

1. The process manager knows it — for years they have watched how, when a batch from supplier X comes in, it pays to raise 3 degrees or extend by 12 minutes. They carry it in their head, note it in a notebook, sometimes in a spreadsheet running alongside the MES.
2. The DCS operator does not have that intuition — and even if they did, changing the recipe mid-campaign without a justification is risky. So the plant runs on the average recipe, and yield drops whenever the raw material moves away from the center.
3. The lab confirms the problem too late — the product analysis comes out of the LIMS several hours after the batch has finished. What you see is the consequence, not the cause.

The result is yield variability that the plant pays for in margin, and a critical piece of knowledge (how to tune the recipe to each raw material) that lives in the heads of two or three veterans. The day one of them retires, a slice of the bottom line leaves with them.

How it fits the IRIS system

iLEAN does not touch the DCS — it captures the veteran's eye and hands it back to the operator on shift.

The problem is not a lack of algorithm: it is information living on islands — the LIMS on one side, the DCS on another, the veteran's notebook on a third. iLEAN acts as the putty that seals those cracks and gives the operator on shift the same intuition the veteran has, on their screen, with the rationale right next to it.

Brain learns the golden batch per product. Connect reads the LIMS, the DCS and the veteran's spreadsheet. The agent proposes the adjustment batch by batch. The person signs — the DCS executes it.

The iLEAN pieces applied to chemical recipe optimization:

• Connect — capture in its three modes: OPC UA integration with a modern DCS, driver-based reading of the LIMS, an intermediate mode for the old lab system that only has an isolated PC, and manual capture by photographing the panel when needed. It also captures what arrives from outside (the supplier's email with the certificate of analysis of the new lot) at second zero.
• Brain — on top of all that history it builds the golden batch per product: the signature of a good batch (temperature curve, dosing, agitation, pH window) and its correlation with the quality of the incoming raw material. When a new raw material arrives, Brain looks for the most similar batch in the past and extracts the adjustment that worked.
• Agent — it proposes the adjustment on the process manager's screen along with the rationale: “raw material from supplier X, lot 2025-1138, purity 97.2% (vs. average 98.5%) → I propose raising 4 °C in stage 2 and extending by 8 min, based on 7 similar batches with yield ≥ 95%”. The manager approves or rejects; the approved change is cryptographically signed, passes through the validation chamber of the intermediate ring and only then reaches the DCS mailbox.

See the full IRIS architecture →

Before and after

Fixed recipe + the veteran's intuition vs. a golden batch with a batch-by-batch proposal

Aspect	Fixed DCS recipe + the veteran's notebook	With iLEAN Brain + Agent
Where the adjustment comes from	Intuition in the heads of 2-3 people	Golden batch per product and campaign, proposed on screen
Reaction to variable raw material	The plant runs on the average recipe	Adjustment proposed before the reactor is charged
Starting data	DCS, LIMS and spreadsheet living on islands	All three sources cross-referenced at second zero
Approval of the change	Verbal, at the shift handover	Manager's signature; cryptographic traceability to the DCS
The veteran's knowledge	Leaves with them on the day they retire	Captured as a permanent capability of the plant
Time to first value	Months of a classic APC project	First value in a few weeks — low-hanging fruit first

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.

• Specialty chemicals plant with 2-4 multi-product batch reactors, an existing Honeywell/Yokogawa/ABB DCS, and a LIMS with at least 12 months of raw material and product quality history.
• Brain pilot on one reactor and one product family (Pareto: the one that drives 80% of the cost). First value expected within a few weeks.
• Reduction of yield variability of around ≥30% on the worst quartile of batches — you already have the good batch; what the agent fixes is the odd one out.
• Indicative payback between 4 and 9 months, depending on the unit margin of the product and the cost of an out-of-spec batch (rework, downgrade to a lower grade, destruction).

And the process manager's reasonable doubt

“What if the AI proposes an adjustment and gets it wrong?” — hallucination is a problem of free generation, not of anchored tasks. When the AI limits itself to recontextualizing a piece of data from one system into another (reading the supplier's certificate of analysis, finding the most similar batch in the history, proposing the same adjustment), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the proposal reaches your screen with the rationale, you sign, and only then does it pass to the DCS. The three safety rings are designed precisely for this — the SIS remains untouchable.

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