CD Uniformity Control
# CD Uniformity Control
Critical-dimension uniformity (CDU) control keeps the spatial and temporal distribution of a declared feature dimension within its manufacturing limits. CDU is not synonymous with line-edge roughness, and it is not adequately described by a wafer range or one pooled standard deviation. A useful control system preserves wafer coordinates, exposure-field position, layout context, chamber/module identity, time order, and metrology uncertainty so each signature can be assigned to a physical mechanism.
The specification must state the measured feature, material boundary, profile height, algorithm, sampling population, and statistical convention. “Across-wafer CDU,” “within-field CDU,” “local CDU,” wafer-to-wafer variation, tool matching, and long-term process capability answer different questions and must not be mixed without a hierarchical model.
## Uniformity levels and statistics
Across-wafer CDU describes a declared set of wafer sites after any explicitly documented exclusions or model removal. Within-field CDU compares locations inside exposure fields. Field-to-field variation measures changes among fields. Local CDU describes nearby nominally identical features under a declared layout and sampling rule. Wafer-to-wafer, lot-to-lot, chamber-to-chamber, and tool-to-tool terms describe manufacturing matching and drift.
Range is intuitive but unstable because it depends strongly on sample count and outliers. Standard deviation is useful only for a defined population and distribution. Robust spread measures can help diagnose contaminated data but do not replace the product specification. Capability indices require a stable process, justified distribution, independent observations or an appropriate time-series model, and trustworthy gauge uncertainty. Control limits describe process behavior; specification limits describe acceptable product. They are not interchangeable.
| Question | Preserve | Avoid |
|---|---|---|
| Is the wafer centered? | wafer mean by feature/layout | averaging different patterns together |
| Is there a spatial signature? | coordinates, field position, orientation | shuffled site lists |
| Are tools/chambers matched? | balanced reference wafers and timestamps | confounding product mix with chamber |
| Is local variation changing? | repeated feature populations and PSD/LWR context | estimating tails from a handful of sites |
| Is the gauge adequate? | repeats, operators/tools, reference and model uncertainty | declaring measurement error as process CDU |
## Sources and causal controls
Lithography signatures can originate in dose, focus, leveling, aberration, illumination, slit position, reticle CD/transmission, pellicle state, overlay interactions, resist thickness, bake, and develop. Etch signatures can originate in plasma distribution, wafer temperature, pressure and flow, chamber wall state, loading, microloading, passivation, endpoint, and mask erosion. Product stack, film stress, topography, pattern density, orientation, and incoming uniformity couple into both.
The correct actuator follows the mechanism. A stable radial resist-thickness or bake signature may justify a track correction after qualification. A repeatable within-field layout signature belongs in imaging/model/OPC analysis. Chamber-specific etch bias belongs in etch matching or run-to-run control. Local stochastic tails are not corrected by shifting the wafer mean. Cross-module compensation is risky because it can center one monitor while worsening profile, roughness, defects, or another layout.
## Metrology and sampling
Sampling must match the signature scale and decision risk. Wafer maps need enough radial, azimuthal, edge, and field coverage. Layout effects need repeated pattern classes. Local tails require many nominally identical features, not just more wafer sites. Chamber matching needs balanced routing and common reference material. Long-term control needs stable timestamps and maintenance context.
Gauge studies should include repeatability, reproducibility, tool matching, reference stability, image/model sensitivity, and sampling uncertainty. Optical and electron-beam methods have different strengths and model dependencies. Cross-sectional methods anchor profile interpretation at sparse locations. Measurement uncertainty should be propagated into control and capability decisions rather than handled through an arbitrary percentage rule.
## Qualification and production signoff
Qualification should exercise tool/chamber combinations, product stacks, reticles, layouts, film lots, track modules, exposure conditions, etch states, edge regions, and maintenance transitions. Confirm proposed corrections at ADI, AEI, defect, profile, and electrical levels. Include holdout wafers or independent monitors so a fitted correction is not judged only on the data used to build it.
Production APC needs guardrails: valid-input checks, recipe and product context, actuator saturation, delay and sampling logic, confidence or uncertainty, fallback behavior, fault detection, change control, and audit history. A controller that reduces one reported CDU metric while moving profile or another layout out of specification is not successful.
The practical signoff question is: Across wafers, fields, layouts, chambers, tools, and time, are all material CDU signatures separated from gauge uncertainty, linked to the correct causal module, and controlled without sacrificing final profile, defects, reliability, or electrical performance?