Wafer Anomaly History Analysis: From Commonality Screening to Root-Cause Verification

Key pointCommonality analysis narrows down the search space; it is not root-cause certification. The evidence graph organizes findings, while algorithms and engineering validation answer specific questions.

Commonality: Comparing excursion wafers with reference groups across shared processing history
Adder: Defects identified as newly detected between consecutive inspections
FDC Trace: Time-series sensor readings recorded during equipment processing
Evidence Graph: A graph connecting observations, results, hypotheses, and sources

Narrow Down the Suspect Process Window First

A routing sheet defines what was supposed to happen; the actual processing history shows what truly occurred. Investigations must track Wafer ID, actual process steps, Tool/Chamber, Recipe, and timestamps, preserving split lots, rework, re-entries, and maintenance events. Lot-level records are often insufficient to isolate single-wafer chamber anomalies.

Prior and current inspection steps provide spatial contrast, but materials and recipe sensitivities may change. 'No detection at the prior step' does not prove the defect did not exist earlier. While the initial window can be bounded between two observations, the possibility of earlier defects being revealed or amplified downstream must be preserved.

Process IllustrationExpanding Evidence Layer by Layer from Wafer Anomaly Images

Images & HistoryObservations & HypothesesReport / Next Verification

01
InputImages and Processing History Enter Investigation Together
Confirm excursion wafers, inspection coverage, and candidate process intervals.
02
CompareSelect Comparable Reference Groups to Screen for Shared Deviations
Examine both excursion and normal reference groups to avoid ranking solely by raw fault counts on affected wafers.
03
ExplainConnect Statistical Correlations to Candidate Physical Mechanisms
Verify processing timestamps, recipes, maintenance logs, and available sensor traces.
04
VerifyDeliver Localized Level and Prioritized Next Actions
Preserve supporting, opposing, and missing evidence; update engineering judgment as new results arrive.

Evidence relationship diagram; solid lines represent recorded relationships, dashed lines indicate candidate explanations awaiting verification.

Reference Groups Must Have Clear Engineering Justification

Reference	Purpose	Prerequisites to Verify
Stable lots prior to excursion	Observe baseline shifts over time	Product type, recipe, inspection sensitivity, and maintenance status
Parallel chambers in same time window	Compare parallel tool performance across the same period	Dispatching rules, incoming materials, product mix, and recipe differences
Consecutive inspections on same wafer	Locate newly detected adders or coordinate shifts	Coordinates, inspection coverage, sensitivity threshold, and defect visibility
Qualified reference baseline	Evaluate tool matching or inspection repeatability	Baseline scope, validity window, and measurement uncertainty

The unit of statistical analysis depends on the engineering question and experimental design. Multiple defect sites on a single wafer are often spatially correlated, and wafers within the same lot may share common bath or chamber exposure. Points cannot simply be treated as independent identically distributed (i.i.d.) observations; correlation must be modeled at the wafer, lot, or tool-run level.

Commonality: Both Sides Are Needed for a True Denominator

Illustrative example:Of 25 abnormal wafers, 22 passed through B; of 100 reference wafers, 5 passed through B. The remaining counts are 3 and 95. B is strongly associated with the anomaly, but if it processed only a particular product or time period, confounding may remain.

Comparison	Processed by B	Not Processed by B
Excursion Group	22	3
Reference Group	5	95

Advanced: Significance, Effect Size, and Multiple Testing

The sample odds ratio is (22 × 95) / (3 × 5) ≈ 139.3, but it is not a probability of root cause. Fisher's exact test is suitable for some small-sample contingency tables, with reasonable independence and sampling assumptions. Reports should state one- or two-sided definitions, effect sizes, and intervals. Screening many candidates also requires appropriate error control: BH controls the expected false-discovery proportion (FDR), while Bonferroni controls the probability of at least one false rejection across the family (FWER). Their objectives differ.

Sensor Trace Drilldown: Preserve the Time Axis Before Applying Alignment

Align sensor traces to wafer processing timestamps and recipe phases first. Verify engineering units, sampling frequency, clock drifts, missing data points, and calibration offsets. Compare means, peak values, durations, and transient responses, selecting features physically tied to the observed failure mode. Not every anomaly is hidden in high-frequency transients, and tools do not all log identical raw waveform signatures.

Dynamic Time Warping (DTW) allows comparison between sequences of varying speeds, but it can artificially warp away genuine process delays or abnormal phase durations. Always retain the raw physical time axis, constrain the allowable warping path, and justify why DTW is selected; it is not a universally mandatory algorithm for all industrial curves.

A Graph Organizes Three Types of Relationships: Keep Them Distinct

Facts:A wafer passing through a chamber is a processing-history record.
Observations and Calculations:An increase in regional density or a waveform difference must have references to its input, method, and result.
Candidate Explanations:State possible mechanisms, supporting and opposing evidence, and the next check.

An Evidence Graph is not a Causal Graph. Traceable links do not establish that a statement is true or that a root cause has been proven.

Discriminating Verification Matters More Than Simply Computing More

Verification may involve morphology review, sensor calibration, maintenance checks, grouped metrology, or splits/DOE. Specify expected observations, comparable conditions, stopping rules, and the responsible engineer. An SEM/EDS elemental peak alone cannot identify a faulty component.

Results can support or weaken an explanation, or still fail to distinguish alternatives. A non-significant result does not automatically falsify a hypothesis. When key data is missing, wait for evidence rather than inventing adjustment values.

Pre-Flight Checklist for Historical Investigations

Problem wafers, inspection stage, anomaly definition, and justification for reference selection.
Actual histories, recipe/inspection versions, maintenance, and material changes.
Coordinates, effective inspection area, time precision, and missing-data notes.
Analysis inputs, parameters, method versions, results, and recalculation materials.
Supporting and opposing evidence, explanations still open, and verification objectives.

WAFERPILOT / APPLICATION TO THIS TOPIC

Deliver Complete Investigation Materials and Engineering Recommendations

WaferPilot aims to organize these checks and generate a report when data connections and professional methods are ready: the level reached, the evidence, what is missing, and what to check next. Grounding handles planning and professional-judgment constraints; computation and result reporting follow. The scope of full cross-step investigation must be confirmed against actual data access and scenario validation.

Deliverables for Engineers:A report containing comparison objects, methods and parameters, results, candidate sources, and a verification plan, rather than only a relationship graph.

View the Current Demo and Capability Scope →