Aluminum casting control with AI — porosity is invisible at demolding; finding it at the leak test means paying twice.

The problem

The defect is born at injection — and discovered after machining.

The head of an aluminum foundry tells it the same way in every plant: the part comes out of the mold looking good, passes visual control, goes into machining. And there — or worse, at the leak test — up comes the porosity, the internal shrinkage, the leak. The part was already born defective at injection, but nobody could see it; and by the time the defect surfaces, the plant has given away machining, handling and testing to a dead part.

Three things happen at once in a real HPDC cell and almost nobody looks at them together:

1. The defect is not visible while it is still cheap — porosity and internal shrinkage cannot be seen on a freshly demolded part. They are discovered downstream, when machining opens up the porous zone or the leak test gives away the leak, with all the added value already lost.
2. The parameters that cause it drift slowly — melt temperature, die temperature, second phase velocity, vacuum level. None of them jumps out of range: they slide cycle by cycle, and by the time somebody notices, the whole batch has come out compromised.
3. Cycle data and defect live in separate systems — the cycle record stays in the die casting machine; the defect is written up in the machining quality report, days later. The cause-and-effect link gets lost along the way, and root cause turns into a meeting full of hypotheses.

The result is always the same: scrap is discovered late, the plant pays the cost multiplied, and the veteran setter goes back to correcting the machine by ear and by notebook. The knowledge that prevents the problem lives in his head; the day he retires, it walks out with him.

How it fits the IRIS system

iLEAN does not replace your die casting machine — it stitches together what today lives apart.

The foundry's problem is not a lack of data: the die casting machine records its cycle and quality writes down its defects. The problem is that the two never meet — the cycle stays in the machine and the defect appears in another system, days later, with no serial to join them. iLEAN acts as the putty that stitches the cycle record, each part's serial, the demolding image and the downstream quality verdict, without asking you to change the machine or the process.

Connect stitches the cycle to every part's serial. Edge sees the surface defect at demolding. Agents correlate defect and parameters and warn about drift before the batch comes out porous. The setter decides.

The three iLEAN pieces applied to casting control:

• Connect — captures the die casting machine's cycle record: through direct integration when the machine exposes the data, or with a camera over the controller panel on older machines, without touching the PLC. Every cycle — melt and die temperatures, phase velocities, intensification pressure, vacuum — is stitched to the serial of the part it produced. It works locally: if the plant loses its network, Connect keeps capturing and recording. What is critical does not depend on WiFi.
• Edge — vision at demolding: it detects the surface defect early (visible shrinkage, cold shuts, drag marks, misruns) on the freshly demolded part, while pulling it out still costs only the aluminum and not the machining. Every image stays linked to the serial and the cycle.
• Agents — correlate the defect confirmed downstream (machining, leak test) with the cycle parameters that produced each part, learn which combinations generate it, and give warning when drift in those parameters points at the defect again — before the whole batch comes out porous. The agent does not act on the machine by itself: it proposes, and the setter decides.

See the full IRIS architecture →

Before and after

Classic casting control vs. control with iLEAN

Aspect	HPDC cell with classic control	With iLEAN Connect + Edge + Agents
Where porosity is discovered	At machining or the leak test, value already spent	Early signal at demolding and through cycle drift
Cycle record per part	Stays in the machine, with no serial	Stitched to each part's serial, queryable
Drift in melt, die, second phase and vacuum	Noticed once the batch has already come out compromised	Trend warning with room to adjust
Root cause of a defect	A meeting full of hypotheses, days later	The exact cycle of the part, in minutes
Older machines with no data interface	Outside the system, data lost	Camera on the panel, cycle digitized all the same
Capturing the setter's knowledge	It lives in his head	A pattern learned by the agents, repeatable

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.

• Automotive HPDC cell, parts with leak-tightness requirements or structural parts, a mix of machines with and without a data interface, and scrap discovered after machining documented over recent months.
• Connect pilot on one or two machines + Edge vision at demolding + Agents learning the defect-parameter correlation for 4-6 weeks before they start anticipating. First value expected within a few weeks: simply stitching cycle to serial already turns every downstream defect into a query that takes minutes.
• Expected reduction in scrap discovered after machining of ≥30% in the first months, a defensible floor — estimate to be validated.
• Indicative payback between 5 and 12 months. The hard lever: every porous part pulled out at demolding, and every batch whose drift is corrected in time, is machining, handling and testing you do not give away to a dead part.
• A recurring benefit that does not enter the ROI but carries weight: the defect-parameter pattern learned from your cell stays as a permanent capability of the plant, not of the person who retires.

And the quality manager's reasonable doubt

“What if the AI correlates wrongly and triggers an unnecessary machine adjustment?” — the iLEAN agent does not act on the die casting machine by itself. It proposes; the setter decides; the correction is applied or it is not. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI cross-references the cycle record with the quality verdict by serial, the best models brought error below 1.5%^[1]. And even so, what is critical goes to the safety rings — the agent lives in the outer ring, proposes inward, and the adjustment on the machine is signed by a person. Never the other way round.

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