Semiconductor defect investigation connects several kinds of evidence: an inspection system locates candidates, review resolves what they look like, analysis tests what they contain or how they extend, and process experiments test the cause. Preserve wafer identity, coordinates and recipe context across that chain.
Give each step a defined output
Applied Materials' defect-control overview describes the tradeoffs among sensitivity, speed and sampling. Its SEMVision review documentation distinguishes automated review and classification and describes combining images with design context. These capabilities support an investigation; the resulting classification still needs a defined meaning in your process.
On a narrow screen, scroll the table horizontally. Keyboard users can focus it and use the arrow keys.
| Step | Retain | Decision it supports |
|---|---|---|
| Inspection | Recipe, inspected area, candidate coordinates and signal | Which sites need review? |
| Review | Relocated images, viewing conditions and unresolved sites | What morphology is visible? |
| Classification | Class definitions, reference images and uncertainty | Which recurring groups deserve investigation? |
| Further analysis | Composition or section evidence with method limits | Which hypotheses remain plausible? |
| Process confirmation | Controlled change and repeat measurements | Does the proposed mechanism explain the response? |
Make the review sample explainable
Record the inspected population and the rule for selecting review sites before looking at the most interesting images. A targeted sample is useful for investigating a cluster; a probability-based sample may serve a population-estimation question. Keep the selection probabilities or strata where those are required by the analysis. Record sites that cannot be relocated or resolved.
Fictional wafer: inspection flags 1,000 candidates. A targeted review selects 70 edge-cluster sites and 30 other sites. It finds 60 particles, 25 pattern defects and 15 unresolved sites. Particles are 60% of this selected sample. The edge-focused selection prevents treating that figure as an unbiased estimate for all 1,000 candidates. It is also a different denominator from electrical die yield.
Use separate counts for candidates, reviewed sites, classified defects and affected dies. Deduplicate where the inspection method can report one physical feature more than once. If a Pareto chart changes, check sampling and class definitions before attributing the change to the process.
Turn a plausible cause into a testable hypothesis
Join the defect pattern to process history: chamber, consumable batch, maintenance event, route and time. Compare suitable unaffected material and check confounding changes. An elemental signal can narrow the possibilities, but its interpretation depends on interaction volume, substrate and contamination; morphology or composition alone rarely identifies the complete generating mechanism.
Define a controlled confirmation with the process owner. State the predicted change, the observations that would contradict it and how repeatability will be checked. Keep instrument settings and detection criteria comparable. Destructive cross-section or focused-ion-beam work needs its own sample disposition and qualified method.
Close with retained evidence
Document the initial observation, review selection, rejected hypotheses, intervention and repeat result. Preserve unresolved classes rather than forcing them into a convenient category. Use the process-control guide when connecting a confirmed finding to a bounded production response; a review label alone should not become an unvalidated automatic process correction.
Use the working record
Download the editable semiconductor defect review. Save a copy, complete the actual requirements and evidence fields, and review open issues with the responsible owner. The plain-text record includes instructions, units where needed, the fictional example, and source references.