Foundry AI Model Drift: When Should You Retrain—or Refuse a Recommendation?

Short answer: Retrain a foundry AI model when verified process or data drift causes a persistent loss of decision quality—not merely because a calendar date arrived. If critical inputs are missing, the operating state is outside validated coverage, or uncertainty exceeds an approved limit, the model should refuse to recommend a parameter change.

A production model is a snapshot of its development conditions. The foundry does not stand still. A die insert is repaired, a cooling channel is cleaned, an alloy supplier changes, a thermocouple is replaced, a new inspection rule is introduced, or the machine reaches a different wear state. The same sensor pattern can then carry a different metallurgical meaning.

This change is often called model drift or concept drift. It is not proof that the AI is bad. It is evidence that the relationship between data, process physics and quality must be checked again.

What does model drift mean in a foundry?

Four different changes are easily confused:

• Input or data drift: the distribution of measured inputs changes. Examples include a new die-temperature range, different vacuum levels or a shifted cycle-time profile.

• Concept drift: the relationship between inputs and the target changes. The same fast-shot velocity and die temperature may no longer imply the same porosity risk after a gate repair or cooling modification.

• Measurement drift: the signal changes because of calibration, sensor position, emissivity, sampling logic or inspection criteria rather than the physical process.

• Label drift: defect codes, CT acceptance zones, leak-test thresholds or operator classification practices change.

These cases require different responses. Retraining on a faulty temperature signal can teach the model to trust the fault. Changing the sensor without recording the event can make a healthy process look abnormal. The first question is therefore not “Should we retrain?” but “What changed in the process, measurement system or quality definition?”

How often should a foundry AI model be retrained?

There is no defensible universal interval. Weekly or monthly retraining may create the appearance of maintenance while silently learning from unverified labels, temporary abnormalities or process faults.

A safer rule is evidence-triggered retraining. Begin a retraining assessment when one or more of these conditions persist:

1. The input distribution moves beyond the validated operating envelope.

2. Prediction residuals or defect-ranking performance deteriorate on newly labelled production.

3. Uncertainty or abstention rates rise in an otherwise stable operating state.

4. A controlled change affects the causal chain: die, runner, cooling, alloy, machine, sensor, inspection method or control software.

5. The model encounters a part, alloy, die or machine configuration that was not represented during validation.

NIST’s AI Risk Management Framework emphasizes documented test, evaluation, verification and validation, regular safety evaluation, real-time monitoring and safe failure beyond knowledge limits. Manufacturing research likewise shows that static training data captures only part of future operating conditions and that environmental, equipment and operating changes can degrade model performance.

Drift detection is not defect detection

A drift alarm does not automatically mean the castings are defective. It means the current data no longer looks or behaves like the model’s validated reference.

For example, a lower die-temperature distribution after cooling-channel cleaning may be an intended improvement. The drift detector should flag it because model coverage changed; the engineer should classify the change as expected, verify the new thermal balance and decide whether the old model remains valid.

A six-step drift-control workflow for HPDC and LPDC

1. Maintain a process change log

Record die maintenance, cooling and spray changes, alloy lots, furnace practice, machine software, sensor replacement, calibration and inspection-rule revisions. Time-align each event with shot or batch data. Without this history, drift diagnosis becomes guesswork.

2. Monitor by operating state

Separate start-up, steady production, post-stop recovery, maintenance trials and known abnormal cycles. Compare like with like. A single global baseline can confuse normal state transitions with harmful drift.

3. Use three evidence lanes

Monitor:

• data: distributions, missingness, sensor health and coverage;

• model: uncertainty, confidence calibration, residuals and abstention;

• process outcome: verified defect, leak, dimensional or metallurgical results.

An alert becomes actionable when the lanes agree or when a hard safety/data rule is violated.

4. Diagnose before adapting

Ask whether the change is physical, measurement-related or administrative. Confirm sensor calibration and data lineage. Review the minimum data model for HPDC and LPDC and the safe HPDC process-window method before allowing the model to adapt.

5. Validate a candidate model in shadow mode

Freeze the production model. Train a candidate only on reviewed data, then compare it against the current model on time-separated production periods and relevant operating states. Check the target metric together with counter-metrics and failure modes. The candidate should not influence machine settings during this evaluation.

6. Approve, release gradually and retain rollback

The process engineer approves the intended coverage, change limits and rejection conditions. Release the candidate to a bounded part/machine scope, monitor it, and preserve the previous model and its data contract for rollback. Version the model together with feature definitions, training period, validation evidence and responsible approver.

When should the AI refuse to recommend?

A trustworthy system needs an explicit no-recommendation state. It should abstain when:

• a critical sensor is missing, stale or implausible;

• the current part, alloy, die or machine identity is unknown;

• the operating state is outside the validated process window;

• input combinations are materially outside training coverage;

• model uncertainty or conflicting evidence exceeds approved limits;

• inspection labels are delayed, changed or unreliable;

• a recent engineering change has not yet been revalidated.

Refusal is not model failure. In a safety- and quality-relevant process, it is controlled behavior.

A modelled foundry example

Assume an HPDC porosity model was validated before a die insert repair. After the repair, the cavity fills with a slightly different local flow pattern. The overall fast-shot velocity and intensification pressure remain within their historical ranges, but the relationship between vacuum timing, switch-over and porosity changes.

A naive system continues recommending within the old window. A drift-aware system detects rising residuals in the repaired cavity zone, links the timing to the maintenance record, blocks automatic parameter recommendations for that die state and requests a controlled validation set. A candidate model is tested in shadow mode before the recommendation envelope is reopened.

This is a modelled scenario, not a universal parameter recipe.

What should Serdar Perçin and the foundry team require?

Every recommendation should identify its model version, validated scope, uncertainty, applicable state and stop condition. Retraining must leave an auditable comparison. Castella helps the engineering team detect meaningful change and update decision support without losing traceability or metallurgical judgement.

Frequently asked questions

Does input drift always require retraining?

No. It requires investigation. An intended and understood process improvement may only require revalidation or a revised operating-state baseline. Retraining is justified when the model’s decision relationship or validated coverage has materially changed.

Can an AI model retrain itself automatically?

Technically it can, but production release should remain gated by data-quality checks, time-separated validation, engineering review and rollback. Automatic ingestion of unverified production labels can reinforce faults.

What is the difference between drift and normal process variation?

Normal variation remains inside a validated state and preserves the expected relationship between inputs and outcomes. Drift changes the distribution, the relationship or the measurement system enough to challenge that validation.

What is the safest action when drift is uncertain?

Keep the current validated model frozen, restrict or suspend recommendations for the affected state, investigate the process and measurement system, and collect reviewed outcome data.

Primary references

• NIST AI Risk Management Framework Core

• Jourdan et al., Handling concept drift in deep learning applications for process monitoring

• Bachinger et al., Continuous improvement and adaptation of predictive models in smart manufacturing

• Holzinger et al., Machine learning based concept drift detection for predictive maintenance