What Data Does Foundry AI Need? A Minimum Data Model for HPDC and LPDC

The short answer: foundry AI does not need every available sensor tag. It needs a traceable chain that connects part identity, machine and thermal state, the process curve, the inspected outcome, and the validated action. If that chain is broken, more PLC data creates volume—not learning.

This article presents the Castella Foundry AI Minimum Data Model, a practical framework for HPDC and LPDC teams. It is not a universal standard. It is a shop-floor starting point for building AI recommendations that engineers can audit, test and improve.

Why “more data” is the wrong target

A die-casting cell can produce thousands of values per cycle. That does not automatically create a useful training record. A pressure curve without the correct die, cavity, alloy batch and inspection result is only a waveform. A defect label without its cycle history is only a quality record.

The real objective is contextualized evidence. NIST’s advanced manufacturing work emphasizes data collection, transformation, traceability and interoperability as foundations for trusted analytics. The NIST AI Risk Management Framework likewise treats validity, reliability, transparency and ongoing measurement as core properties of trustworthy AI.

For a foundry, those principles translate into one practical question:

Can an engineer reconstruct why this recommendation was made, which evidence supported it, and whether the change improved the next production window?

If the answer is no, the system is not ready for operational decision support.

The six records every casting cycle should carry

1. Part and production identity

Every cycle needs a stable identity before it needs advanced analytics:

• plant, line, machine and cell;

• part number, die/tool ID, cavity and revision;

• alloy grade, melt or batch ID and metal treatment state;

• cycle timestamp, shift and production order;

• recipe version and relevant maintenance state.

This layer prevents the model from treating different parts, cavities or tool conditions as if they belonged to one process. It also makes later root-cause work reproducible.

2. Process trajectory—not only setpoints

A setpoint says what the machine was asked to do. A trajectory shows what actually happened.

For HPDC, the minimum useful trace commonly includes slow-shot velocity, fast-shot transition, switch-over position, filling time, intensification pressure, vacuum timing and relevant alarms. For LPDC, it includes the pressure-time curve across lift, filling, feeding/holding and release, together with pressure stability and interruptions.

Store the curve or meaningful engineered features, not just one average. Two cycles can share the same average pressure while having very different transitions. For LPDC teams, our pressure-curve engineering workflow explains why the stages must be interpreted separately.

3. Thermal and utility state

Casting quality depends on the state of the system before metal enters the cavity. The minimum thermal record should include:

• melt and holding-furnace temperature;

• die or mould temperatures by meaningful zone;

• cooling-water flow, temperature and timing;

• die spray or release-agent timing where relevant;

• cycle interruptions, warm-up status and restart conditions.

Thermal readings need location, unit, sampling time and sensor health. A value called “die temperature” is not enough if nobody knows which zone it represents. The same rule applies to flow, pressure and time units.

4. Material and tool condition

AI should not explain a material or tooling change as a machine-parameter problem. Record the variables that can change the operating window:

• alloy certificate or verified composition when available;

• melt treatment, degassing and transfer history;

• die coating, lubrication and cleaning condition;

• vent, vacuum and cooling-circuit maintenance;

• tool life, repair events and controlled changes.

These do not all need high-frequency sampling. They need correct effective dates so every cycle inherits the right condition.

5. Quality outcome with location and evidence

The most common failure in manufacturing AI is weak labelling. “Porosity” is rarely enough. A useful outcome record includes:

• accepted, scrap, rework or hold status;

• defect family and suspected mechanism;

• defect location, cavity and severity;

• inspection method: visual, X-ray, CT, leak test, machining or destructive section;

• inspection timestamp and, where possible, the image or report reference;

• disposition confirmed by quality or process engineering.

Gas and shrinkage porosity, for example, demand different corrective logic. Our HPDC diagnostic workflow shows why morphology, CT, shot-profile, vacuum and thermal evidence should be combined before parameters are changed.

6. Intervention, approval and validation result

A recommendation becomes organizational knowledge only when the change and its result are recorded:

• parameter changed, old value and new value;

• reason and supporting evidence;

• person who reviewed or approved the change;

• time and production window in which it was applied;

• before/after quality, stability and cycle-time result;

• model or rule version, confidence and stated limitations;

• decision to retain, revert or investigate further.

This closes the learning loop. It also prevents a temporary improvement from being mistaken for a causal result.

One data contract matters more than another dashboard

Foundries often connect machines successfully but still lack shared meaning. A tag called PRESSURE_1 can refer to a setpoint on one line, an actual value on another, and a filtered average on a third.

Every variable in the data contract should therefore carry:

• a stable name and definition;

• unit and expected range;

• source system and sensor location;

• sampling or aggregation rule;

• timestamp and clock source;

• missing-value and sensor-fault status;

• recipe, tool and part context.

Open industrial approaches such as OPC UA and MTConnect demonstrate the value of structured information models and semantic context. A foundry does not have to replace every legacy system to adopt the principle: values must retain meaning when they move from the machine to MES, quality systems and AI.

A practical evidence ladder for foundry AI

Not every plant should jump directly to automatic control. A safer maturity path is:

1. Describe: unify cycle, thermal and quality records.

2. Diagnose: identify repeatable relationships and competing root causes.

3. Recommend: propose bounded parameter changes with reasons and confidence.

4. Validate: run engineer-approved trials inside documented process limits.

5. Standardize: retain only changes that survive repeat production and quality review.

6. Monitor: detect drift and stop trusting the recommendation when the context leaves its validated range.

This is why Castella is designed as engineering decision support. Traditional casting simulation, live production data and human validation serve different roles; our comparison of simulation and real-time AI explains how they complement each other.

The 30-day minimum viable implementation

Do not begin by connecting every tag. Begin with one part, one defect family and one accountable team.

Week 1 — Define the question. Select a measurable problem such as gas porosity in one HPDC cavity or unstable feeding in one LPDC wheel. Freeze the defect definition and success metric.

Week 2 — Build the cycle record. Connect identity, process, thermal and material/tool context. Verify units, timestamps and missing values on the shop floor.

Week 3 — Join quality evidence. Link inspection results back to the correct cycle or production window. Review uncertain labels with process and quality engineers.

Week 4 — Test bounded recommendations. Compare a baseline window with controlled, engineer-approved changes. Record every intervention, result and reason to retain or reject it.

The deliverable is not a beautiful dashboard. It is a small, auditable evidence chain that can be repeated.

Frequently asked questions

How much historical data does a foundry AI project need?

There is no defensible universal cycle count. The requirement depends on defect frequency, process variation, number of parts and cavities, label quality and the intended decision. Clean, contextualized records from a focused process are often more useful than years of mixed, weakly labelled data.

Can we start without new sensors?

Often, yes. Begin with the PLC, machine curves, furnace records, existing thermal measurements and quality results already available. Add sensors only where a decision-critical variable is missing or unreliable.

Should AI change machine parameters automatically?

Not by default. Start with bounded recommendations, engineering approval, documented limits and a revert path. Automation should follow demonstrated validity and reliable monitoring—not precede them.

What is the single most important field?

There is no single field. The most important asset is the join key that connects the correct part or production window to its process history and inspected outcome.

Final principle

The competitive advantage is not owning the largest data lake. It is being able to turn one casting result into a traceable lesson that improves the next decision.

When identity, process, thermal state, material/tool condition, quality evidence and human-approved interventions remain connected, AI can support repeatable foundry engineering. When those records are separated, even a powerful model is forced to guess.

Serdar Perçin is the founder of Castella AI for Die Casting and a Metallurgical & Materials Engineer. Castella develops practical decision support for HPDC, LPDC and sand-casting teams, combining process physics, production evidence and human engineering validation.