Aerospace heat treatment with AI — quality cannot be verified afterwards: you have to prove the curve was THE curve.

The problem

The part comes out of the furnace perfect on the outside. The doubt is not about the part — it is about the paperwork.

In an aerospace heat treatment shop everyone knows how to run the cycle. The furnaces are validated, the thermocouples calibrated, the TUS and the SAT up to date. The problem is not thermal: it is documentary, and it always shows up at the same point — when somebody has to prove, months later, that that load followed that curve.

1. The curve lives outside the system — it comes out on a strip chart recorder, on a thermal printer or in software from the furnace manufacturer that nobody else opens. It is a chart in a folder, not a data point you can cross-reference. If the recorder jams or the ink runs out, the evidence of the cycle simply does not exist.
2. The load is linked by hand — the operator notes on a sheet which batches went into the basket and at what time. That sheet, the chart and the work order are three separate pieces of paper that only a person with a good memory can put back together. When that person is away, the link is made by deduction.
3. The audit rebuilds it — Nadcap arrives, the auditor picks ten batches at random and the hunt begins: which furnace, which zone, which thermocouple, which calibration was current, which operator loaded it. Weeks in the office rebuilding something that happened correctly but was never tied together.

And there is the worst case, the one nobody logs in the KPI: the load lost to a doubtful record. The part is fine, the cycle went fine, but the evidence is not conclusive and the only safe way out is to redo the treatment or scrap it. It is not that the shop treats badly — it is that the proof it treated well is kept worse than the treatment itself.

How it fits the IRIS system

iLEAN touches neither the furnace nor the control loop — it seals the crack between the curve and the dossier.

Here the rule is non-negotiable: the furnace is not touched. It is validated, it has its TUS, its SAT and its history, and any intervention on the control would force a revalidation. iLEAN acts as putty between what the furnace already shows and what the audit needs to read, without connecting to the control loop and without writing a single setpoint.

Edge reads the curve off the panel without going into the equipment. Connect ties the load to the batches and to the parts at loading time. The agent assembles the Nadcap dossier per load batch with the real curve — the certified person signs by role.

The three iLEAN pieces applied to an aerospace heat treatment shop:

• iLEAN Edge — a terminal with computer vision aimed at the recorder or controller panel: it reads the curve, the setpoint, the temperature by zone and the status of the cycle exactly as they appear on screen. Non-invasive capture: no new wiring, no touching the cabinet, no impact on the furnace's validation. When the equipment already exposes data, the reading is done through direct integration. It works with no network: if the plant goes offline, Edge keeps recording and syncs when the link comes back.
• iLEAN Connect — at loading time, the operator confirms on the terminal which batches and part numbers are going into the basket or the rack, by reading the label, the DMC or the work order. From then on there is a load identifier the curve attaches itself to. It also captures thermocouple calibration certificates and TUS/SAT reports whether they come by email, as a PDF or as a photo of the paper, and leaves them where the dossier is going to need them.
• iLEAN Agents — the agent cross-references part ↔ load ↔ curve ↔ thermocouples ↔ furnace ↔ operator and assembles the Nadcap/AMS dossier per load batch: real curve against the specification limits, times at temperature, ramps, current calibrations and deviations. It does not sign: it prepares it and hands it to the heat treatment or quality manager. The human signature is sacred; the three safety rings are there precisely to protect it.

See the full IRIS architecture →

Before and after

Heat treatment documented by hand vs. the curve linked with iLEAN

Aspect	Paper chart recorder + load sheet	With iLEAN Edge + Connect + Agents
Cycle curve	Chart in a folder, impossible to cross-reference	Real curve captured and linked to the load
Load ↔ part link	Handwritten sheet and the operator's memory	Load identifier at second zero
Thermocouple calibration / TUS / SAT	PDF in a quality folder	Captured and linked to the load batch
Nadcap dossier per batch	Manual reconstruction, weeks	Prepared by the agent, ready for signature
Load with a doubtful record	Redo the treatment or scrap it	Conclusive evidence from the start
Intervention on the furnace	n/a	None — non-invasive reading, no revalidation
Manager's signature	On a file rebuilt by hand	By role, on a dossier prepared by the agent

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 shop. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Aerospace heat treatment shop with three to six furnaces (vacuum and/or atmosphere), at least one with a paper chart recorder, and Nadcap AC7102 approval in force or in progress.
• Edge pilot on one furnace with non-invasive panel capture and load linking on the terminal. First value expected within a few weeks: the first load batch with its real curve linked and searchable in seconds, without having touched the furnace.
• Expected reduction in the time to prepare the audit dossier ≥50% and a target of zero loads redone or scrapped over a doubtful record. Indicative payback between 5 and 10 months, depending on the number of furnaces, audit frequency and the unit value of the parts that today get redone for lack of evidence.
• The hard lever is the quality manager's hours that today go into rebuilding Nadcap files and the loads that stop being lost because the curve was already tied to the part.

And the aerospace CAIO's reasonable doubt

"What if the system misreads a value off the panel and an incorrect curve ends up in a Nadcap dossier?" — hallucination is a problem of free generation, not of anchored tasks. Reading a numeric value and a temperature trace on a known panel and checking them against the specification limits is an anchored task: the best models brought error below 1.5% ^[1]. And even so, the system does not decide — it proposes, flags, records, and marks any low-confidence reading for a person to confirm. The heat treatment or quality manager's signature stays exactly where it has to be. The three safety rings are there precisely for this: conformity of a special process is not closed out by an AI, it is closed out by a certified person.

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

Related: Nadcap aerospace part traceability · Penetrant and ultrasonic NDT with AI · Aerospace painting and coating with AI