HPDC Process Windows: How AI Should Recommend Parameter Changes Safely

Short answer: Foundry AI should not search for one “perfect” HPDC setting. It should define a bounded, part-specific process window, recognize interactions between variables, and recommend only small changes that remain inside validated safety and quality limits.

An HPDC machine may display hundreds of values, but a casting is not produced by isolated setpoints. Fast-shot velocity changes the way air is displaced. The switch-over position changes the melt front at the instant velocity rises. Vacuum performance changes the tolerance to the resulting flow pattern. Die thermal state changes filling, solidification and soldering risk. Intensification pressure acts only after the cavity, gates and local metal are in the condition to transmit it.

That is why “increase pressure” or “raise velocity” is not yet an engineering recommendation. It is a hypothesis. A safe recommendation must also state the operating context, expected response, evidence, limits and stop condition.

What is an HPDC process window?

A process window is the validated region in which a combination of inputs can repeatedly meet defined quality and production requirements. It is not a list of independent minimum and maximum values.

Imagine that fast-shot velocity is acceptable from A to B and switch-over position from C to D. The four corners of that rectangle are not automatically safe. A late switch-over combined with a high velocity may create a very different filling condition from either change alone. The usable region may be curved, narrow or split by die temperature, alloy condition, lubrication or vacuum state.

HPDC research has long treated parameter selection as a multivariable problem. Studies have examined combinations of first- and second-stage piston velocity, metal temperature, filling time and pressure, while integrated work links die temperature, cooling and injection conditions to casting quality. That evidence supports a practical conclusion: local interactions matter, and a single universal recipe does not exist for every part and machine.

Why single-setpoint optimisation fails

Three recurring errors make apparently “optimised” settings fragile:

• The response is vague. Density, X-ray class, leak rate, dimensional stability, soldering and cycle time are not interchangeable objectives.

• The context is mixed. Start-up shots, steady production, die insert changes, alloy lots and abnormal vacuum cycles are pooled as if they came from one process.

• Correlation is treated as causation. A velocity change may appear to reduce rejects because it was made after the die reached thermal equilibrium, not because velocity caused the improvement.

This is also where AI can mislead. A model can predict accurately inside historical data yet recommend an unsafe combination when asked to extrapolate beyond it. Prediction confidence is not the same as process safety.

A six-step method for a safe AI recommendation envelope

1. Define the decision and quality response

Start with one decision: for example, whether to adjust fast-shot velocity after a repeated gas-porosity signal. Define the response with a measurement method and time horizon: CT indication by zone, leak-test result, reject code or a validated proxy. Do not use “quality improved” as a label.

2. Build the physics-based context

Record the variables that change the meaning of the proposed action: alloy and melt condition, shot weight, sleeve fill, slow-shot profile, switch-over, cavity fill time, vacuum curve, intensification, die temperatures, cooling state, spray history and time since interruption. Our earlier minimum data model for foundry AI explains how to preserve this traceability.

3. Separate operating states

Do not train one undifferentiated model across start-up, stable production and recovery after downtime. Create explicit state gates. If the die thermal state or vacuum integrity is outside the validated condition, the AI should diagnose the state first, not compensate by moving an unrelated setpoint.

4. Identify interactions locally

Use designed trials, simulation and production evidence to test the small region around the current process. Include interaction terms or models capable of representing them, but retain engineering review. Historical data alone often under-samples settings operators correctly avoided.

5. Convert the model into constraints

The recommendation layer needs more than a predicted defect probability. It should enforce:

• hard equipment and safety limits;

• part-, alloy- and die-specific validated boundaries;

• maximum change per cycle or trial block;

• forbidden combinations;

• evidence thresholds and data-quality checks;

• an automatic “no recommendation” state outside model coverage.

6. Require verification and monitor drift

The engineer approves the trial, its sample size and its rollback criterion. The result is then checked against both the target defect and counter-metrics such as flash, soldering, dimensions, cycle time and tool stress. A process window must be revalidated after meaningful changes to die, machine, alloy, sensors or control logic.

This approach is consistent with the NIST AI Risk Management Framework: deployed AI should be valid and reliable, remain within acceptable risk, be monitored, and fail safely beyond its knowledge limits. In a foundry, “fail safely” often means refusing to recommend a parameter change and asking for engineering evidence.

A modelled HPDC example

Suppose reject data suggests gas porosity rises when fast-shot velocity is high. A naive model recommends reducing velocity. A guarded workflow asks four more questions:

1. Did the vacuum curve reach the expected condition before filling?

2. Was the switch-over repeatable relative to actual metal quantity and sleeve fill?

3. Was the die in the same thermal state across good and bad shots?

4. Does lower velocity increase fill time enough to introduce cold-flow or incomplete-fill risk?

The result may be a conditional recommendation: test a small velocity reduction only when vacuum, switch-over repeatability and die thermal state are inside their validated gates; otherwise correct the abnormal state first. This is a modelled example, not a universal machine recipe.

What should remain under human authority?

AI may rank hypotheses, identify interactions, detect drift and propose a bounded trial. The process engineer remains accountable for whether the evidence is relevant to this alloy, part, die and machine; whether the trial is safe; and whether the observed improvement is metallurgically credible.

For defect diagnosis, morphology still matters. Before optimising settings around a porosity label, distinguish entrapped gas from feeding-related shrinkage using the gas-versus-shrinkage diagnostic workflow.

Castella’s role is therefore not to replace the engineer with an opaque number. It is to help Serdar Perçin and the foundry team connect process physics, traceable evidence and bounded AI assistance into a decision that can be reviewed and reversed.

Frequently asked questions

Is a process window the same as machine tolerance limits?

No. Machine limits protect equipment and basic operation. A validated process window is narrower and specific to the part, alloy, die condition, quality response and measurement system.

Can AI find a process window from historical data alone?

It can estimate the region represented in the data, but historical production rarely covers all safe combinations and cannot by itself prove causality. Designed trials, simulation, engineering knowledge and controlled verification are still needed.

How many parameters should be changed at once?

Enough to test the stated interaction, but no more than the trial can identify and control. In production troubleshooting, bounded staged changes are usually easier to interpret and reverse than a large multi-parameter jump.

When should the AI refuse to recommend?

When required sensors are missing, the operating state is abnormal, the proposed point lies outside validated coverage, conflicting quality objectives are unresolved, or the expected benefit is smaller than the uncertainty.

Primary references

• NIST Artificial Intelligence Risk Management Framework 1.0

• Kong et al., Integrated optimization system for high pressure die casting processes

• Syrcos, Die casting process optimization using Taguchi methods

• Dou et al., Complete CAE modelling and optimization of the HPDC process