- Scalar
- A Summary Number, Such as Total Defect Count
- Spatial Pattern
- Arrangement of anomaly coordinates across the wafer surface
- Repeater
- Repeating patterns at identical relative die/field coordinates
- Hypothesis
- Candidate mechanism requiring empirical verification
Aggregation Discards Spatial Information, but Analytics Can Leverage Space
Wafers with identical total defect counts can display fundamentally different failure modes: radial scratches vs concentric rings. Spatial features can be quantified into radial profiles, zone densities, or periodicity indices for statistical modeling. Visual and numerical approaches complement rather than oppose each other.
- 01CountScalar Totals Conceal Critical Distinctions
Both wafers report similar defect totals, but spatial patterns differ completely.
- 02PatternOne Linear Scratch, One Outer Edge Ring
Spatial continuity, radius, and orientation establish priority check lists.
- 03HypothesesFormulate Competing Root-Cause Hypotheses
Handling, processing, materials, or inspection conditions may all contribute.
- 04ValidationSelect Discriminative Verification Checks
Cross-check pre/post-step inspections, SEM reviews, and tool histories to update engineering judgment.
Treating Spatial Patterns as Diagnostic Clues
| Visual Observation | Priority Verification Check | What Cannot Be Asserted Blindly |
|---|---|---|
| Annular Ring / Edge Concentration | Inspection Regions, Edge Processes, Thermal Fields, Materials, and Handling Conditions | That a specific focus ring or heater jacket is physically damaged |
| Linear Scratches | Mechanical Contact, Polishing, Handling, and Reviewed Morphology | That CMP slurry delivery line contamination is the sole root cause |
| Exposure Field Periodicity | Masks, Layout, Process Sensitivity, and Inspection Algorithms | That a reticle particle contamination event is 100% responsible |
| Localized Cluster | Local Materials, Contamination, Processes, and Inspection Background | That a specific gas delivery showerhead orifice is clogged |
Spatial Statistics Require Rigorous Applicability Constraints
Clustering algorithms, radial distributions, Hough transforms, and spatial autocorrelation quantify pattern geometry. However, methods must account for valid care areas, coordinate resolution, and inspection sensitivity without relying on rigid static thresholds to diagnose root cause.
Advanced: How Moran's I Should Be Evaluated
Moran's I evaluates spatial autocorrelation across spatial units with neighborhood spatial weight matrices. For spatial point processes, engineers must define grid binning or utilize point process methods (e.g. Ripley's K). Values depend on spatial weighting, sample size, and null hypotheses; near-zero does not prove benign noise, and significant clustering does not prove equipment failure.
Leveraging Mature Image Analytics in Real Fab Workflows
Manual review, heuristic rules, spatial statistics, and machine learning are established tools in fab defect analysis. CNN classifiers provide pattern tags and feature embeddings that integrate with temporal sensor models. An intelligent co-pilot makes these analytical outputs directly accessible within auditable engineering workflows.
WAFERPILOT / APPLICATION TO THIS TOPIC
Transforming Visual Observations into Actionable Recommendations
WaferPilot inspects images and existing analyses in ImageLens, recording anomaly locations, referenced material, and limitations. By further combining available history, metrology, and equipment conditions, it aims to deliver candidate causes and verification recommendations, beyond a morphology label.
Deliverables for Engineers:Spatial anomaly characterization, cross-wafer pattern comparisons, and representative defect reviews paired with discriminative verification steps.
View the Current Demo and Capability Scope →Related Reading & References
Below are verifiable industry references; specific processes, permissions, and acceptance criteria are governed by fab site specifications.