Uniformity is a statistic, and the statistic that gets quoted decides which problems the fab is able to see. The number almost everyone reports is a half-range: the difference between the thickest and thinnest measured site, divided by twice the mean. It is computed from two of the forty-nine sites and ignores the other forty-seven entirely. Two wafers can carry an identical one percent half-range and be in completely different conditions — one a smooth centre-to-edge bowl that a single gas or thermal knob will flatten, the other a random speckle that means the chamber is shedding particles or the measurement is failing. The half-range cannot distinguish them, because it is a function of extremes and carries no information about shape.

The standard-deviation form uses every site and is far more stable run to run, which is why it is the better control-chart statistic even though the half-range remains the customary one for specifications. But neither is sufficient on its own, because both collapse a two-dimensional map into a scalar. The practice that actually finds root cause is to decompose the map rather than summarise it. Fit and remove a radial component — thickness as a function of distance from wafer centre. Fit and remove an azimuthal component — thickness as a function of angle. What is left is the residual. Each of those three pieces points at a different part of the hardware, and their relative magnitudes are far more diagnostic than any single number computed from the raw map.

A dominant radial term is the ordinary case and it is the one the tool was designed to control: it comes from the balance between centre and edge gas delivery, from the thermal profile of a multi-zone heater, from the electrode gap in a plasma system, and from the way flow turns outward and exits at the wafer edge. A dominant azimuthal term is a hardware alarm rather than a tuning opportunity, because a properly assembled chamber with a rotating or symmetric geometry has no reason to produce one: it points at a partially blocked showerhead sector, an asymmetric pumping path, a tilted or warped susceptor, a lift-pin that is not seating the wafer flat, or an RF return path that is not symmetric. A dominant residual with no structure at all points at measurement noise, at particle contamination, or at a genuinely stochastic film. Reporting one percent tells the engineer nothing about which of these three worlds they are in; reporting that the one percent is ninety percent radial tells them where to go.

Temperature is the reason uniformity is hard in the surface-reaction-limited regime, and the sensitivity can be written down rather than asserted. When growth is limited by a thermally activated surface reaction, rate follows an Arrhenius form, so a small temperature difference across the wafer translates into a thickness difference with a gain set by the activation energy:

Put realistic numbers into that. For an activation energy near one and a half electron-volts at a deposition temperature around six hundred degrees Celsius, the prefactor works out near two percent per degree. A wafer that is one degree hotter at the centre than at the edge will therefore come out roughly two percent thicker at the centre, which is already at or beyond the specification for most films — from a temperature difference that a casual thermal design would not even notice. This single number explains why deposition chambers carry multi-zone heaters with individually trimmed setpoints, why susceptor flatness and wafer-to-susceptor contact are treated as critical, why backside gas pressure is a controlled parameter, and why a wafer that sits on three particles instead of flat on the chuck produces a thickness signature. It also explains the standard trade: a transport-limited process is less uniform in principle but far less sensitive to temperature, so moving a process deliberately toward transport limitation is sometimes the correct uniformity fix even though it sounds backwards.

The most consequential idea in wafer-level uniformity is that minimum non-uniformity is not the objective. What a device cares about is the result after every module has run, not after any one of them. If a deposition is systematically centre-thick and the etch that follows it is systematically centre-fast, the two signatures subtract and the finished structure is flatter than either step was. Fabs exploit this deliberately: a deposition profile is tuned not to be flat but to be the mirror image of a downstream signature that cannot itself be removed. Driving each step independently to its own minimum can make the integrated result worse, and a process engineer who improves a deposition from one and a half percent to half a percent without checking the downstream compensation can genuinely degrade the final critical dimension. Uniformity is a budget allocated across a flow, not a per-step virtue, and the correct question about any deposition signature is what it will be added to.

How the map is sampled matters more than most specifications admit. A forty-nine-point measurement on a three-hundred-millimetre wafer is not a fine grid; it is a coarse one, and where those points sit changes the answer. A cartesian grid under-samples the outer annulus badly, because area grows with radius: the outermost tenth of the radius on a three-hundred-millimetre wafer holds close to a fifth of the total area and a comparably large share of the die, yet a square grid places only a handful of sites there. A polar sampling plan with more sites at larger radius represents the wafer far better and will typically report a worse number for exactly the right reason. The edge exclusion setting is an even blunter lever on the reported result: moving exclusion from three millimetres to two, with no change whatever to the film, can move the reported non-uniformity by a large fraction of the specification, because the excluded ring is where the steepest gradient in the entire map lives. Any comparison of uniformity numbers between tools, fabs, or vendors that does not first reconcile site count, site placement and edge exclusion is comparing sampling conventions rather than films.

The metrology itself has to be audited before any of its output is trusted, and the audit is not optional for tight specifications. Spectroscopic ellipsometry infers thickness through an optical model, so a change in film composition, density or interface roughness across the wafer will present as an apparent thickness gradient that is partly or wholly an artifact. Four-point probe measures sheet resistance and converts through an assumed resistivity, so a resistivity gradient — extremely common in metal films, where grain structure varies with local thermal history — appears as a thickness gradient. When a uniformity signature does not respond to any recipe change that should affect it, the most likely explanation is that the property varying across the wafer is not the one being reported.

Wafer-level uniformity is only one of four variation terms, and the others are frequently larger. Within-wafer variation is what everything above concerns. Wafer-to-wafer variation within a lot exposes first-wafer effects, where the chamber state after an idle period or a clean differs from its state in steady flow, and it is the reason for dummy wafers, seasoning layers and warm-up sequences. Lot-to-lot variation exposes chamber drift between preventive maintenance events, consumable ageing, and the slow accumulation of deposits on chamber walls. Chamber-to-chamber variation across a matched set is often the largest single term in a high-volume fab and the least visible in development data, because development runs on one qualified tool while production runs on twelve. A film that is half a percent uniform within a wafer and three percent different between chamber four and chamber nine has a three percent problem, not a half percent one, and no amount of within-wafer optimisation touches it.

Reading a uniformity excursion follows from the decomposition. A radial signature that appeared suddenly points at flow — a mass flow controller drifting, a partially blocked line, an altered pumping speed, or a pressure control valve behaving differently. A radial signature that drifted slowly points at thermal, at consumable ageing, or at wall deposits changing the chamber's radiative environment. A new azimuthal term points at something physically disturbed, and it is worth checking that the chamber was reassembled correctly before touching a recipe. An edge-only change points at the edge ring, the susceptor pocket, or a change in the incoming wafer edge profile. A shift in the mean with the shape unchanged is a rate problem rather than a uniformity problem and belongs to a different investigation. And a signature that appears only on product and not on monitor wafers is loading, which is a layout interaction rather than a chamber condition and will not be found by any amount of blanket-wafer work.

A uniformity specification that will hold up therefore states considerably more than a percentage. It names the statistic, since half-range and standard deviation are not interchangeable and differ by roughly a factor of three for a typical map. It names the site count, the site placement convention, and the edge exclusion, because those three determine the number as strongly as the process does. It states the acceptable decomposition, not just the total, so that an azimuthal term is caught as a hardware fault even when the total is inside limits. It states the wafer-to-wafer, lot-to-lot and chamber-to-chamber budgets alongside the within-wafer one. It states the metrology and its known cross-sensitivities. And, most usefully and least commonly, it states the downstream signature this deposition is expected to compensate, so that a future engineer who finds a centre-thick profile understands that it is deliberate before helpfully removing it.

CVD uniformity qualification and excursion workflow

st=>start: Freeze wafer identity, film state, map recipe, edge exclusion, and statistic
gauge=>operation: Verify gauge repeatability, site registration, and measurement model
shape=>operation: Separate mean, radial, azimuthal, edge, layout, and residual components
class=>condition: Is the signature structured and repeatable?
hardware=>operation: Inspect flow, temperature, gap, pumping, rotation, seating, and wall state
noise=>operation: Challenge metrology, particles, handling, and unstable nucleation
scale=>operation: Compare monitor and product across wafer, lot, chamber, and maintenance timescales
integrate=>operation: Test downstream compensation and device-level response
release=>end: Release statistic, map convention, variance budget, limits, and reaction plan
st->gauge->shape->class
class(yes)->hardware->scale
class(no)->noise->scale
scale->integrate->release

Half-range and standard deviation answer different questions. Half-range is intuitive for a specification but is controlled by two extreme sites and grows more volatile as site count increases. Coefficient of variation uses every site and is better suited to control charts, while neither statistic preserves map shape.

A normalized statistic must retain its denominator definition. Dividing by wafer mean, target thickness, centre-site thickness, or a fitted surface produces different percentages. Preserve raw units, the normalization basis, and the unrounded calculation.

Map decomposition converts geometry into process evidence. Fit radial terms, angular harmonics, edge behavior, and known layout regions before interpreting residuals. Save both component maps and residuals so a good scalar cannot conceal a new signature.

Radial coefficients should be trended independently from the mean. Average thickness can hold while a bowl becomes a dome because centre and edge changes cancel. Curvature and edge-slope charts expose that drift.

Azimuthal structure is fundamentally a symmetry audit. A first harmonic suggests tilt, asymmetric exhaust, delivery imbalance, or seating. Higher harmonics can correspond to showerhead sectors, heater zones, lift pins, or RF return geometry.

Residual structure must be tested for spatial correlation. Random-looking points may cluster at a length scale associated with holes, die layout, scan order, or particles. Repeat maps and correlation checks separate noise from unresolved process structure.

Sampling density must match the shortest relevant length scale. Sparse maps resolve broad radial curvature but cannot prove the absence of local edge roll-off or pattern loading. Establish signatures with dense characterization maps before reducing production sites.

Edge exclusion is part of the measurement recipe. State whether distance is measured from the physical edge, nominal radius, or valid-die boundary. Lock registration and exclusion logic so software changes cannot create false excursions.

Gauge capability sets the narrowest credible process limit. Separate repeatability, repositioning, model fitting, tool-to-tool, and time components. A limit narrower than demonstrated gauge capability creates overcontrol rather than uniformity.

Thickness must be separated from correlated material properties. Ellipsometry couples thickness to index and roughness; sheet resistance couples thickness to resistivity; X-ray fits couple thickness to density. Confirm suspicious maps with an orthogonal technique.

Chamber state belongs in every uniformity model. Post-clean condition, seasoning count, idle time, accumulated deposition, source life, and consumable age change flow, emissivity, plasma impedance, and wall reactions.

Product loading can invalidate blanket-wafer conclusions. Dense patterns consume precursor, modify plasma current, alter temperature, and challenge optical models. Segment product results by pattern density and feature class.

Feature-scale uniformity is not wafer-scale uniformity. Field thickness, sidewall coverage, bottom thickness, seams, and selectivity can vary independently. Qualify the local metric controlling device performance at centre, mid-radius, and edge.

The time signature narrows the physical cause. A maintenance step change suggests assembly or seasoning; slow drift suggests coating, source, thermal, or consumable aging; a repeating first-wafer signature suggests idle recovery.

Chamber matching requires matching shapes and mechanisms. Equal half-ranges can describe opposite profiles. Compare mean, fitted coefficients, residual covariance, product response, and material properties at multiple setpoints.

Nested variance prevents tuning the wrong level. Partition within-wafer, wafer-to-wafer, lot-to-lot, chamber-to-chamber, and metrology contributions with a balanced plan. Pooled statistics routinely hide the dominant term.

Deliberate compensation must be documented as an integration requirement. Store the target component map and downstream sensitivity when deposition cancels etch, polish, lithography, or implant variation. Requalify the pair when either module changes.

Control limits should monitor both magnitude and shape. Use scalar charts for mean and dispersion, coefficient charts for spatial components, and residual alarms for new patterns. Keep every contributing signal interpretable.

An excursion response should preserve evidence before adjustment. Hold material, repeat the map when safe, retain raw spectra, check coordinates, compare sensors, and inspect event history before changing a recipe.

Production release requires an auditable uniformity contract. Name film state, statistic, units, normalization, sites, coordinates, edge exclusion, gauge model, sampling frequency, variance level, limits, compensation target, and reaction plan.

Statistic selection

Sampling and edge control

Nested variance and chamber matching

Read CVD uniformity through a statistic, spatial-decomposition, sampling, gauge-capability, variance-hierarchy, and integrated-process lens rather than a single percentage lens.