Semiconductor Yield Screening

A production line discovered most defects at final test, after the value had already been added. Engineers could investigate a failed lot, but they could not see which in-process runs were drifting toward failure early enough to intervene. The default response — inspect everything, or inspect nothing until a customer complains — is either unaffordable or too late.

The screening decision

Each shift, the quality team had to decide which lots to hold for extra inspection and which process signals deserved investigation. The output that mattered was not a defect probability on its own. It was a short, ranked list sized to the inspection team's capacity, with the sensor groups behind each flag, so an engineer could start from evidence rather than from a hunch.

What we worked with

• Production runs: 1,567 records, each with a timestamp and a binary in-house test result.
• Process signals: 591 numeric measurements from the line, with no descriptive names supplied.
• Outcome: 104 failures (about 6.6%) against 1,463 passes — a rare-event problem.
• Data quality: about 4.5% of values missing, spread across every record rather than isolated to a few runs.

With far more signals than runs, the risk of finding patterns that do not repeat was high. The work therefore started with time alignment and signal grouping, not with a model.

The traps we designed around

• Leakage from later measurements. Some signals are recorded after the step that causes a defect, or after the test itself. Using them makes the model look excellent and useless in production. Every feature was checked against the moment the decision is made.
• Rare events and misleading accuracy. Predicting "pass" for every run scores about 93%. Evaluation used precision and recall at the inspection capacity, plus the cost of a missed defect versus a false hold.
• Missingness was not random. A blank sensor value can mean a skipped step, a tool state or a recording failure. Missing indicators were kept, and imputation was treated as an assumption to be tested rather than a cleaning step.
• Too many features, too few failures. With 104 positives, an unconstrained model can memorise the line. Signals were grouped by process stage, screened for stability across time, and regularised.
• Alert fatigue. A list longer than the team can inspect is not a screening system. Thresholds were set against capacity and reviewed with the quality engineers after the first weeks.

How the screening was built

1. Fix the decision point. We agreed where in the process the screening happens, what counts as a failure, and how far ahead the flag must arrive to change the outcome.
2. Audit time alignment. Signals were ordered by timestamp and tested for post-decision leakage. Ambiguous records were excluded from training rather than guessed.
3. Study the missing patterns. Missingness was compared between passes and failures before any imputation. Where a pattern was informative, it became a feature; where it was mechanical, it was documented.
4. Reduce and group signals. Correlated sensors were collapsed into process-stage groups, and stability selection removed signals whose relationship to failure changed between periods.
5. Compare a simple baseline with a flexible model. A regularised logistic model provided direction and sanity checks; a tree-based ensemble captured interactions. Both were validated on held-out periods, not random rows.
6. Set the threshold by cost and capacity. The operating point balanced the cost of an escaped defect against the cost of holding a healthy lot, subject to how many lots the team could inspect.
7. Close the loop. Whether a flagged lot actually failed, and what the inspection found, was recorded so the screening could be reviewed and recalibrated.

What this changes

• Inspection follows risk. The team works a ranked list sized to its capacity instead of applying the same checks to every lot.
• Flags come with evidence. Each alert points to the process-stage signals that drove it, so an engineer can confirm or reject it quickly.
• Yield conversations become specific. Quality and process engineering share one view of which signals are drifting and which lots are exposed.
• The model has an owner and a review cadence. Thresholds, miss rates and false-hold rates are part of the operating routine, not a one-off validation report.

What we delivered

• A risk-ranked screening list for incoming lots, with process-stage signal groups attached.
• A documented leakage review and a missingness map for the 591 signals.
• A cost-sensitive threshold tied to inspection capacity, with a simple rule for when to recalibrate.
• A quality dashboard showing flagged lots, outcomes and signal drift over time.
• A playbook for reviewing flags, recording findings and feeding results back into the model.

Where this applies

This fits semiconductor, electronics, precision components and any high-volume line where final-test results exist but in-process signals are plentiful. The minimum inputs are run identifiers, timestamped process measurements, a reliable pass/fail outcome and a team that can act on a short list. The first deliverable is usually the leakage and missingness audit — because a yield model built on post-decision signals is worse than no model at all: it looks accurate and cannot be used.

This is an illustrative case study built on an anonymized dataset; it does not report results from a specific client engagement. Screening quality depends on stable definitions and on the quality team's willingness to act on flags; both were treated as part of the work.