AlN Circuit Surface Defects: Compare Counts per Inspected Area

Measurement purpose

Compare countable surface indications using a valid observable-area denominator.

Specimens and conditions

Surface domain

Material state, functional region and eligible area identified before counting.

Event definition

Stable inclusion threshold, grouping rule and unresolved-feature handling.

Equipment and records required

• Imaging and scale: Qualified detection conditions and calibrated object-space field dimensions.
• Spatial registration: Repeatable coordinates for overlapping fields and event deduplication.

Method sequence

1. Map

   Define eligible observable regions and exclusions.

   Record: Area masks and coverage gaps.

2. Count

   Locate qualifying events without double-counting overlap.

   Record: Event coordinates and original sightings.

3. Compare

   Report density, part status and spatial strata separately.

   Record: Count/area table and inference assumptions.

Decision and uncertainty

Use the denominator and model appropriate to the question; retain functional part disposition separately.

Unobserved area, clustering, small counts and changed detectability limit inference.

Inspection-method owner and drawing acceptance authority.

Traceable outputs

Define one countable event before totaling the images

State the feature being counted: a resolved ceramic surface pit, an unprinted island inside a required film region or another specifically defined indication. Do not combine unrelated mechanisms merely because the inspection software labels all of them defects. Record the dimensional or visual criterion that makes an event eligible, including how touching features are separated or grouped.

A long connected scratch can intersect several image tiles. Counting each appearance as a new event inflates the total; treating a cluster of separate pits as one event can reduce it. Preserve coordinates and a segmentation rule that can be applied consistently. If a feature cannot be classified reliably, retain its unresolved status rather than forcing it into the count used for a process comparison.

Calculate the area in which that event could actually be seen

Map the relevant region on the actual circuit face. For an exposed-film coverage question, the denominator might be the printed region required by the drawing, not the full ceramic rectangle. For a bare-substrate surface question, metallization, clamps and obscured portions may remove areas from the observable domain. State exclusions before comparing results and show them on the field map.

The area must also meet the method's detection conditions. A blurred or saturated image tile is not fully inspected simply because it was captured. Record unsuccessful fields and the resulting coverage gap separately. Excluding a difficult region changes the scope of the conclusion; it must not be described as evidence that the unobserved region is free of defects. Use calibrated object-space dimensions rather than pixel counts from differently scaled images.

Remove field overlap without deleting real neighboring events

Stitched scans commonly overlap to align adjacent fields. Build the union of their accepted observable areas, rather than adding each rectangular field's area independently. Use registered event coordinates to remove duplicated sightings of the same physical feature. Keep the original sightings available so a reviewer can distinguish deduplication from deletion of inconvenient observations.

If a feature crosses the boundary of the chosen analysis region, apply one declared rule consistently. A centroid inclusion rule can be useful for suitable countable features, while another method may be needed for irregular continuous damage. Do not move the region boundary after viewing the counts. Record incomplete edge features and whether they meet the chosen inclusion rule; their handling can matter greatly in a small field.

Compare event density and part status side by side

Consider two hypothetical inspection groups using the same event definition and qualified optics. Group A has five events in fifty square millimetres of observable area. Group B has eight events in two hundred square millimetres. Their observed densities are 0.10 and 0.04 events per square millimetre, respectively. Group B has the larger raw count but the smaller area-normalized density.

Now suppose the five A events occur on one of five parts, while the eight B events occur on eight of ten parts. If every counted event makes its part nonconforming under the stated drawing rule, the defective-part fractions are twenty percent and eighty percent. This reverses the apparent ordering. Neither metric is incorrect: one describes events per area, the other the distribution of affected parts under a specific acceptance rule.

Use a count model only when its assumptions fit the surface

A homogeneous Poisson model represents the expected event count as inspected area multiplied by event density. It assumes a suitable constant rate and independent event occurrence over the modeled domain. Under that restricted model, count variance equals expected count. It can provide a starting point for planning an area-based comparison, but it is not a physical law of ceramic defects.

A squeegee-related streak, edge-handling damage or clustered contamination can violate the constant-rate or independence assumption. Several marks from one debris particle are not necessarily independent events. Compare equal-area subregions and retain their locations before choosing a count-based uncertainty or test procedure. For small counts, avoid casually applying a symmetric normal approximation that can yield impossible negative densities.

Keep location and feature consequence in the result

Divide observations by functional region where the mechanism or consequence differs. A ceramic handling margin and a high-field separation region should not disappear into the same average. Likewise, bare AlN and a fired conductor surface have different inspection visibility and defect definitions. Report separate densities when pooling would obscure those differences.

Preserve each region's count and area, not just a mean of the reported densities. If compatible regions can legitimately be combined, divide total events by total eligible area. An unweighted average gives a tiny region the same influence as a large region and can change the answer. Even the correctly weighted total must remain accompanied by the local results when a concentrated defect pattern affects the functional decision.

Keep the observation opportunity stable across process comparisons

Use the same face, manufacturing state, detection threshold and field-selection method before and after a process change. A higher-resolution inspection can increase the count even when the product improves. Conversely, added glass coverage can hide a previously visible substrate feature without removing it. Treat a changed observation method as a separate factor rather than calling the numerical shift a process improvement.

Select fields through the agreed sampling plan, not by choosing the cleanest available areas. If a study intentionally targets suspect regions, label its scope as targeted diagnosis and do not extrapolate its density to the entire lot. Preserve unsuccessful inspections and missing parts so the apparent improvement cannot arise from a reduced opportunity to observe difficult material.

Issue an area-linked inspection comparison

Deliver event coordinates, classification rules, observable-area masks and the count summary together. Identify which parts and functional regions are affected, with their actual drawing disposition. Where clustering or insufficient visibility prevents a justified statistical comparison, state the additional observation needed rather than publishing a falsely precise improvement percentage.

For an AlN circuit review, ChipSimple needs the substrate and film layout, the defect question and the inspection conditions. The practical output is a reproducible comparison of a defined surface condition. Area-normalized counts do not establish thermal conductivity, insulation performance or remaining ceramic strength. Those properties require their own material- and application-specific evidence.