In a production inspection workflow, not every defect detection should trigger a human review. High-confidence detections of clearly classifiable defects can be dispositioned automatically — the inspection system's output is reliable enough that adding human review adds cost and latency without improving accuracy. Low-confidence detections, borderline cases, and defect types where the classifier's accuracy is known to be limited should be escalated for human confirmation before a yield disposition decision is made.
The confidence threshold that separates automatic-disposition detections from human-review escalations is not a parameter that can be set once and left unchanged. It is a calibration that should be reviewed as classifier performance changes, as defect mixes shift with process variation, and as the human review capacity available in the fab changes with staffing and volume.
What Confidence Score Calibration Means in Practice
A classification model outputs a confidence score for each defect detection — typically a probability value between 0 and 1 for the predicted class. The intuitive interpretation: a score of 0.95 means the model is "95% confident" in its classification. The actual meaning is more nuanced.
In practice, many classification models are not well-calibrated in the probabilistic sense. A model that outputs 0.95 for the correct class may actually be correct 78% of the time on that category, not 95% of the time. The score is a ranking signal (higher scores are more reliable than lower scores) but not an accurate probability estimate without explicit calibration. This matters because threshold-setting based on uncalibrated scores can produce a review escalation rate that is systematically different from what the threshold was intended to achieve.
Calibration testing is straightforward: take a held-out labeled validation set, run the model on it, bin the detections by confidence score (e.g., 0.90–0.95, 0.95–0.98, 0.98–1.00), and measure the actual accuracy in each bin. If the bin at 0.95–0.98 has actual accuracy of 82%, the model is overconfident at those scores and the threshold needs to account for this. Platt scaling or isotonic regression are established post-hoc calibration methods that can correct this overconfidence without retraining the classification model.
Setting Thresholds by Defect Class, Not Globally
A single global confidence threshold applied to all defect classes is rarely optimal. Different defect classes have different misclassification costs and different model accuracy profiles. An approach that works better: set per-class thresholds based on the misclassification cost structure for each class.
For defect classes where misclassification as "non-defect" has high cost (yield killers — metal bridge shorts at critical layers, gate oxide damage), the threshold should be set low enough that essentially all detections are reviewed. The human review rate for this class will be high, but the cost of an automated false-release is unacceptable. For nuisance classes (edge exclusion artifacts, thin-film interference patterns that appear consistently on a specific product) where the model accuracy is high and confirmed defect cost is low, a high threshold allows nearly all detections to auto-disposition.
This per-class threshold approach requires that the inspection system's output includes per-class confidence scores alongside the overall confidence score, and that the disposition logic supports per-class threshold configuration. Many modern inspection platforms support this; legacy systems may have only a single global threshold parameter. If threshold granularity is a workflow requirement, it should be specified in tool requirements before procurement.
The Review Capacity Constraint
Threshold-setting cannot be done purely from a quality perspective without considering the human review capacity available. A threshold that sends 30% of detections to human review in a facility with adequate QA staffing is operationally feasible. The same threshold in a facility with a staffing gap sends a review queue that grows faster than it can be processed, leading to WIP aging and eventually to operators clearing holds under time pressure — which, as discussed in other posts, mechanically increases false negative rate.
The practical approach: start by defining the maximum acceptable human review volume per day (based on realistic QA staffing capacity with reasonable workload per reviewer). Translate this into a maximum daily escalation rate by dividing by the total daily detection volume. Set the threshold to achieve this escalation rate, then evaluate the resulting false-release rate — the fraction of high-risk defects that will be auto-dispositioned rather than reviewed. If the false-release rate at this threshold is unacceptable, the review capacity constraint needs to be relaxed (more QA staffing, shift adjustments) or the detection volume needs to be reduced (better nuisance kill management, which reduces the denominator without changing the true defect capture rate).
We are not saying that review capacity should always limit quality outcomes — there are cases where higher QA staffing investment is clearly justified by the yield impact of increased review coverage. The point is that threshold-setting that ignores review capacity produces a threshold that cannot be sustained in practice, regardless of its theoretical quality properties. The sustainable threshold is the one where the theoretical quality outcomes are achievable given the actual operational constraints.
Threshold Drift and Monitoring
Once thresholds are set, they will drift from optimal performance over time for three reasons. First, classifier accuracy degrades as the process evolves and new defect subtypes appear that the model was not trained on. Detection of novel defect types at low confidence by a model that does not recognize them looks identical in the confidence score distribution to detections it would have classified correctly six months ago. Second, defect mix shifts — as process improves, the mix of defect types changes, and a threshold calibrated on one mix may produce different escalation rates on a different mix. Third, human review quality changes with operator rotation, training, and fatigue patterns.
Monitoring threshold performance requires tracking two metrics continuously: auto-disposition accuracy (the rate at which auto-dispositioned detections are subsequently confirmed as correctly classified via sampling review) and escalation rate (the fraction of detections escalating to human review as a function of time). Deviations in either metric from the target range should trigger a threshold recalibration review.
Quarterly threshold reviews, or reviews triggered by process changes, product changes, or operator team changes, are a practical cadence for facilities that have not automated the monitoring. Automated threshold health monitoring — alerting when the estimated auto-disposition accuracy drops below a configurable limit based on sampled confirmation data — is achievable in modern inspection platforms and eliminates the lag between performance degradation and detection.
Documentation and Change Control
Confidence thresholds for disposition decisions in a production quality system are part of the inspection process specification and should be under change control. A threshold change has the same yield and quality implications as a recipe change on a process tool — it directly affects which wafers proceed through the line and which are held for review.
Threshold change documentation should capture: the current threshold values, the calibration data supporting the new threshold values, the expected change in escalation rate and auto-disposition accuracy, the authorization for the change, and the effective date. Without this documentation, retrospective yield analysis that tries to correlate quality events with process changes will be unable to account for threshold changes as a causal variable. Yield investigations become harder because a category of production specification changes is invisible in the process history records.