First-Pass Yield Prediction

First-pass yield (FPY) prediction scores each unit for the probability it will pass final test, using process parameters recorded earlier on the line. The score is not a pass/fail decision. It is a risk signal Quality uses to decide which units deserve extra inspection before they reach the final gate.

Upstream signals typically include recipe setpoints and actuals, temperatures, pressures, speeds, dwell times, torque, fill levels, cure profiles, and station cycle times. Models trained on historical process-to-outcome pairs learn which combinations correlate with escapes and scrap. At run time, each unit (or lot, if that is the traceability grain) gets a pass probability and, optionally, a ranked set of contributing factors.

The operational payoff is selective re-inspection. Instead of spreading the same inspection effort across every unit, Quality concentrates capacity on units the model flags as statistically at risk, while the baseline sample plan continues to cover the rest of the population.

Vendors commonly used for this pattern include Sight Machine (plant analytics and predictive quality on process and quality data), Siemens Quality / Opcenter Quality (quality management and analytics tied to manufacturing execution), and InfinityQS (SPC and quality intelligence with predictive extensions). The same pattern can also be built on plant historians, MES traces, and an in-house ML stack; the workflow below is independent of the product choice.

Data capture starts at the process stations that influence final test. Parameters must be time-aligned and unit-keyed (serial, barcode, or lot) so each prediction is traceable to the same physical item that will later be tested. Missing, late, or mismatched keys produce empty predictions, not guesses.

Feature construction turns raw traces into model inputs: windowed averages, excursions from recipe limits, cumulative thermal dose, order of operations, and rework indicators. Labels come from final test and disposition history (pass, fail, rework, scrap). Training uses periods where both process data and outcomes are complete; deployment monitors for drift when recipes, materials, or test criteria change.

At inference, the model emits a pass probability (or risk score) per unit. A threshold or risk band maps scores to actions, for example:

• High confidence pass: continue under the standard sample plan.
• Elevated risk: add targeted re-inspection or secondary checks before final test.
• Very high risk or incomplete inputs: hold, escalate, or require full inspection for that unit.

Quality still owns the sample plan. Prediction only reallocates extra effort; it does not replace AQL, skip-lot rules, or regulated inspection frequencies. When process parameters for a unit are missing or incomplete, the system returns an empty prediction and that unit falls back to the plan's default treatment (often full or increased inspection), never a synthetic "likely pass."

Agree on the prediction grain: unit-level where serial traceability exists, lot-level where it does not. Mixed grains confuse thresholds and make escapes harder to audit.

Define the outcome label carefully. "First pass" means passed final test without rework after the process window used for features. Including reworked units as passes without a rework flag will train the model to ignore problems that only appear after salvage.

Set ownership of thresholds and actions. Process engineering may own feature definitions and recipe limits; Quality owns pass/fail criteria, sample plans, and when a risk score triggers hold or 100% inspection. Document that split so production cannot treat a green score as license to skip required checks.

Establish the empty-prediction rule in writing: no process parameters (or incomplete critical fields) → no score → sample-plan default. That rule prevents silent under-inspection when sensors drop out or MES tags fail to bind.

Decide how explainability will be used on the floor. Factor rankings help process engineers chase root causes, but operators should act on clear bands (re-inspect / continue / hold), not raw SHAP plots at the station.

Label noise is the most common failure mode. If final test criteria changed mid-history, or scrap codes are inconsistent, the model learns the wrong boundary. Retrain only after Quality validates that historical labels match current disposition rules.

Process–outcome lag and weak genealogy break unit-level scoring. If parameters cannot be bound to the serial that fails later, predictions become lot averages that are too blunt for selective re-inspection.

Overfitting to a single product family or supplier lot looks strong in a pilot and collapses at changeover. Monitor performance by product, line, and material revision; freeze or retrain when those strata shift.

Treating the score as a gate is a governance failure, not a model failure. If Production starts shipping on high pass probability without the sample plan, escapes rise and the program loses credibility. Keep the model advisory until Quality has evidence that selective re-inspection improves detection without thinning coverage elsewhere.

Sensor outages and recipe edits without model refresh produce either empty predictions (correct) or stale scores (dangerous). Prefer empty over stale: configure critical-feature checks so incomplete inputs suppress the score rather than fill with defaults.

Track first-pass yield at final test as the primary business metric, but pair it with inspection effectiveness metrics so yield gains are not bought by missing defects.

Useful paired measures:

• Escape rate or customer return rate for defects that final test should have caught.
• Hit rate of the risk band: share of flagged units that actually fail final test or secondary inspection.
• False-alarm rate: flagged units that pass all extra checks (drives capacity cost).
• Coverage of empty predictions: share of units with no score, and whether those units received the mandated fallback inspection.
• Sample-plan compliance: confirmation that baseline sampling was not reduced when the model was introduced.

Run a holdout or A/B design where possible: same line, same sample plan, with vs without selective add-on inspection driven by the score. Compare detected defects per inspection hour and escape rate, not only FPY. If FPY rises while escapes rise, the model is optimizing the wrong outcome.

Review thresholds on a fixed cadence with Quality and Process. Thresholds that maximized pilot hit rate often need retuning when mix changes. Keep a change log for model version, feature set, and action bands so audit trails stay intact.

Selective FPY scoring pairs naturally with richer inspection and process feedback loops:

• Inline Vision-Based Defect Detection for catching surface and assembly defects at station rate, which can supply labels or secondary checks for high-risk units.
• Inspection Escape Root Cause Correlation for linking escapes back to process and inspection gaps when predicted-pass units still fail downstream.
• Unlabeled Anomaly Detection when failure labels are scarce and you need early process deviation flags before a supervised FPY model is reliable.
• Process Parameter Optimization to close the loop: once at-risk combinations are known, adjust recipes and controls so fewer units enter the elevated-risk band.