Pharma lyophilization with AI predictive control — a 48-hour cycle, and a deviation takes 5 minutes to destroy the batch.

The problem

The most expensive batch in your plant spends 48 hours looking at itself. And nobody is looking the whole time.

Lyophilization is the pharma process with the highest concentration of value per minute of cycle and, at the same time, the most opaque to sustained human attention. The reality for a biotech production manager is this:

1. The cycle lasts between 24 and 48 hours. Nobody watches the SCADA closely for 48 hours straight. Shifts change, screens get muted, alarms pile up, and the corners of the process (condenser profile, residual pressure at stoppering) become invisible.
2. The deviation is obvious once it is already too late. A vacuum micro-leak at the end of primary sublimation, a shelf temperature running 0.5 °C low for two hours, a miscalibrated product thermocouple — these are things a veteran operator feels when looking at the curve, but by the time they come over to look, the curve has already collapsed the cake.
3. The veteran's knowledge is in no system at all. The judgment of the operator who has spent 30 years watching that specific machine is worth more than the SCADA, and it is documented nowhere. The day they retire, part of the freeze dryer's performance walks out with them.

The financial consequence is severe: a single freeze-dried batch of a high-value SKU (vaccine, monoclonal antibody, biologic) is worth on the order of hundreds of thousands to millions of euros. The classic system works almost every time — and that “almost” is what puts the annual lost-batch figure into the committee's report.

How it fits the IRIS system

iLEAN captures the veteran's eye and holds it without blinking for 48 hours.

The lyophilization problem is not a lack of sensors: the freeze dryer's SCADA already has 50 capture points. The problem is that nobody cross-references those 50 points against the historical twin of the same SKU on the same freeze dryer, and nobody projects them forward to see 30 minutes ahead. iLEAN acts as the putty that covers that dead zone, without asking you to change the SCADA, the freeze dryer or the quality system.

Edge captures the cycle in real time. Brain compares it against the historical twin. The agent anticipates the deviation and warns before it collapses the cake. The person signs; the cycle does not restart on its own.

The three iLEAN pieces applied to pharma lyophilization:

• iLEAN Edge — a terminal on the floor that ingests every freeze dryer signal in real time (thermocouples, Pirani/capacitance, condenser, valves, profile). It works with no network: if the plant loses WiFi, Edge keeps capturing and holding the cycle locally; when the network returns, it syncs. What is critical cannot depend on connectivity.
• iLEAN Brain — the multi-agent brain. It ingests the full batch history for the same SKU and builds a digital twin of expected behavior. It detects when the running cycle starts to separate from that twin, early enough to correct. What the veteran operator did on intuition, sustained by the machine without a break.
• iLEAN Agent — cross-references the detected deviation with the rest of the context (raw material quality, primary vial batch, pending maintenance on the freeze dryer, history of micro-leaks) and proposes an action. If the deviation is critical, it holds the cycle at an acceptable phase and alerts the responsible person. The person validates and signs; the cycle does not restart on its own.

See the full IRIS architecture →

Before and after

Classic lyophilization vs. lyophilization with iLEAN Edge + Brain + Agent

Aspect	Classic lyophilization (SCADA + one operator per shift)	With iLEAN Edge + Brain + Agent
Cycle attention	An operator who comes and goes, alarms muted during the night peak	Continuous, without blinking, for the full 48 h
Micro-leak detection	When the residual vacuum fails to close at the end — already too late	On the vacuum rise curve, with room to correct
Miscalibrated thermocouple	Discovered during the post-mortem of the ruined batch	Cross-referenced against the other thermocouples in real time
The veteran's knowledge	Lives in the head of the operator who retires	Captured as a model, stays in the plant
Operation with no network	Local SCADA with no remote access	Edge keeps capturing, syncs when the network returns
21 CFR Part 11 dossier per batch	Export from the SCADA + cross-reference with quality by hand	Generated automatically, with a complete audit trail

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.

• 2 GEA/Telstar/IMA freeze dryers running a mid-to-high value SKU (vaccine or biologic), typical cycle 36 h, average batch 50-200 thousand vials.
• Deployment of iLEAN Edge over the SCADA + Brain with the history of previous batches. First value expected within a few weeks: detection of a micro-deviation against the historical twin.
• Indicative payback between 4 and 9 months, depending on the documented frequency of lost batches over recent years and the unit value of the SKU.
• Expected reduction of batches lost to cycle deviation of ≥ 30% in the first year, with enough lead time to correct.
• The hard lever is a single batch saved per year. For a high-value biologic SKU, one intervention pays for the pilot and leaves margin to scale to the rest of the freeze dryers.

And the reasonable doubt of the technical director / Qualified Person

“What if the AI tells me to change the recipe and gets it wrong?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares signals against a historical twin (what Brain does with every cycle), the best models brought error below 1.5% ^[1]. And even so, the iLEAN architecture does not decide alone: the agent proposes, holds at an acceptable phase, and the technical director signs. Digitalization done right in capital-intensive industries is what separates those who grow from those who fall behind — the most digitalized sectors raised productivity by up to 40% ^[2].

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