Spray drying with AI — process drift shows up half an hour early, not after the batch.

The problem

The SCADA warns you once you are already out. By then it is too late.

A pharmaceutical spray dryer is a process where dozens of variables move at once — inlet air temperature, outlet temperature, feed flow rate, atomization speed, relative humidity of the process air, actual viscosity of the suspension, cyclone pressure. And the operator has a panel with numbers and thresholds.

As long as the curve stays inside the thresholds, everything is “fine”. By the time a threshold trips, you are already out of specification — residual moisture too high, build-up in the chamber, batch compromised. The SCADA says “this happened”, it does not say “this is going to happen in 30 minutes”.

The three pieces of data you would need to cross-check to anticipate the drift live on three islands:

1. The correct batch curve — the process signature of a batch that came out perfect last week. In the data historian, untagged, never cross-checked with the raw material.
2. Today's real curve — the CPPs in the equipment's SCADA, second by second, watched by the operator one at a time.
3. What changed between the two batches — a different API supplier, higher ambient humidity in the plant because the weather turned, a filter purge done at a different point. Every change in a spreadsheet, an email, a maintenance report.

Three islands, none of them talking to the others. The deviation lives in the joint between the three, and nobody was there listening in time.

How it fits the IRIS system

iLEAN does not add another SCADA — it puts a brain on top that compares curves.

Spray drying control is not a problem of missing measurement. The modern plant already has a SCADA, a historian and the suspension with its spec sheet. The problem is that nobody compares today's real curve with the signature of the correct batch at the moment they start to diverge, and nobody cross-checks that deviation with what changed between the two batches. iLEAN is the putty that seals that crack.

Edge reads local pressure and temperature without going through the SCADA. Connect captures the entire SCADA stream and the supplier emails. The agent compares against the good batch, anticipates the drift and proposes the adjustment. The person signs — never the other way round.

The three iLEAN pieces applied to spray drying:

• Edge — a terminal with vision over the cyclone or the bag filter to detect incipient build-up, and with its own temperature/pressure sensors if the SCADA does not arrive in time. It works with no network.
• Connect — hooks into the spray dryer's SCADA (OPC UA if it is modern, PLC reads if it is old, a photo of the panel if it is very old) and captures the whole trace second by second. And it also captures the external channels: the email from the API supplier flagging a change in the new batch, the outside relative humidity, the maintenance reports. All of it at second zero.
• Agent — trains on the curve of the batches that came out right and compares today's real curve against that signature. If it detects that in 30 minutes the outlet temperature is going to fall 4 °C and residual moisture will come out of spec, it does not wait for the SCADA: it notifies the operator with the proposed adjustment (raise the air flow rate or lower the feed rate). The person validates; the plant does not adjust itself.

See the full IRIS architecture →

Before and after

Classic spray drying vs. spray drying anticipated with iLEAN

Aspect	SCADA + classic thresholds	With iLEAN Edge + Connect + Agent
Drift detection	When the threshold trips (already out)	30+ minutes earlier, against the good batch signature
Build-up in the chamber	Visual inspection between batches	Edge sees it during the process
Raw material change	Assumed “the same as the last batch”	Connect captures the supplier's email
Ambient humidity in the plant	Not recorded with the batch	Cross-checked by the agent with the curve
File for inspection	Rebuild from 4 systems	Automatic dossier per batch, with a signed curve
Operation with no network	n/a	Edge keeps capturing on the panel's own 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 pharmaceutical spray dryer, 1-3 t/h, multi-product with weekly recipe changes, equipment SCADA + a separate historian.
• Connect + Agent pilot capturing from the SCADA + the MES recipe. First value expected within a few weeks: anticipation of residual moisture deviations in ≥ 30% of the documented incidents.
• Indicative payback between 4 and 9 months, depending on the average cost of a deviated or reprocessed batch and the frequency of incidents.
• The hard lever: a single batch anticipated per quarter pays for the pilot. The second batch anticipated capitalizes it.

And the process director's reasonable doubt

“What if the AI gets it wrong and proposes lowering the flow rate at the wrong moment?” — hallucination is a problem of free generation, not of anchored tasks. Here the AI is anchored to the good batch curve and to the recorded CPPs: in tasks like that, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes, the person signs with their credential — the three safety rings guarantee that the live process (ring 1, the OT network) only accepts what has been validated and signed. And the aggregate effect in more digitalized plants is striking: the sectors that invested in capturing reality in full raised productivity by as much as 40% compared with the least digitalized ones ^[2].

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks. ^[2] Fundación BBVA / Ivie — productivity through digitalization, 2000-2021.