ORIGINAL COMPANY-CONTEXT PRACTICE
Prevent leakage in synthetic wafer-defect modeling
Original practice scenario: A semiconductor simulator creates many die measurements per wafer and several wafers per lot. How should a defect-prediction dataset be split?
Written for independent practice using company context; this is not a verified question from a company interview.
A REASONED APPROACH
Model answer.
Respect the hierarchy that creates correlated observations. Random die-level splitting can place near-identical process conditions in training and testing, overstating generalization to future lots.
Work through it
- Identify lot, wafer, tool and time dependencies, and define whether deployment targets future lots, new tools or both.
- Construct grouped temporal holdouts matching that target, fit preprocessing only on training data and exclude measurements unavailable at prediction time.
- Report performance by tool, lot and defect category, compare simple baselines and assess uncertainty using independent production units rather than individual correlated dies.
Tradeoffs to discuss
- Holding out whole lots creates a harder and smaller test but better reflects real generalization than a large contaminated random split.
Common pitfalls
- Do not use final inspection results as early-process features.
- High aggregate accuracy can conceal failure on rare but important defect categories.