Wafer And Spatial

Wafer map modes

The wafer map supports these modes:

• hard bin
• soft bin
• pass/fail
• parametric heat map for a selected test

Parametric mode shows the test value on the wafer, not just the bin or pass/fail result.

Orientation tick

A small tick outside the map marks the orientation reported by the file’s WCR. It is a direction reference, not a drawing of the physical flat or notch. Unknown orientation and sequential, non-spatial layouts have no tick.

The current maps show normalized native coordinates, with increasing Y drawn downward. A reported down-facing flat (D) therefore appears at the top of these maps; U appears at the bottom. The caption states the screen edge explicitly, for example Native / flat D / shown at top. The tooltip separates the file’s reported axes from the normalized display convention. An aligned acquisition does not mean the map has been rotated into the canonical frame.

The reference comes from the selected snapshot, including on comparisons and stacked views. It follows pan and zoom, disappears when its edge is offscreen, and is retained in canvas captures. Changing bin colors or brushing does not change orientation.

When to use parametric mode

Use it when you are looking for:

• center-to-edge drift
• quadrant asymmetry
• radial bands
• reticle or chuck signatures
• hotspot structure before it becomes yield loss

Spatial analysis card

The spatial card ranks explainable failure signatures before showing the lower-level evidence. Current labels cover edge/ring, local cluster, linear or scratch-like, directional gradient, repeating or reticle-like, and across-lot drift patterns.

Labels are calibrated against deterministic pass/fail permutations over the same wafer geometry. A label must clear both a sample-size-aware significance gate and a practical effect-size floor. It is evidence about a shape, not a diagnosis of a process or equipment cause. The card can explicitly report that no statistically credible signature was found. Approximate die counts and wafer lists are prefixed with ~ when they are projected from the bounded calibration sample.

The displayed q is adjusted across the six signature families, not a raw shuffle count. Hover it to see the raw permutation value, exceedance count, observed statistic, largest null statistic, and null model. Edge, gradient, and drift use global label shuffles. Cluster, scratch, and reticle preserve failure totals within fixed radial bands. Every shuffled label set repeats the full maximum search for that family before the cross-family adjustment is applied.

Hotspot evidence is conditioned on the radial fail-rate profile, so an edge ring is not presented again as a local cluster merely because edge die have a higher base failure rate. The ranked cluster automatically searches tight through broad scales; the Neighborhood Radius control changes only the exact hotspot evidence list and linked selection regions. Overlapping candidate neighborhoods are suppressed. Adjusted q gates the result; it is not part of the score. The displayed score ranks family-relative effect magnitude and is not a probability.

Correlated shapes are conditioned before they are shown. A peaked cluster does not also appear as a ring, gradient, or scratch when removing that cluster removes the secondary evidence. Smooth gradients and repeating residue patterns are likewise prevented from producing incidental line labels. Independent combined patterns can still produce more than one ranked signature.

In Lot Stack mode, progression uses the detected monotonic wafer tail when one is identifiable. Clicking the signature selects that tail rather than an arbitrary fixed fraction of the lot.

Lot Stack also screens a bounded, balanced sample wafer by wafer and groups surviving edge/ring, cluster, scratch, gradient, and reticle findings by family. A group selects the fitted region on each listed wafer; it does not reuse one region across the lot. The pooled lot result appears first, while the wafer counter shows grouping progress.

Ranked signatures choose their own spatial extent. An edge/ring signature searches a fixed radial grid and selects only the detected normalized radial range. Edge Width and Radial Bands change the exact evidence panels and manual linked selections, not the ranked signature. This keeps an evidence-display preference from changing the classification or over-selecting a detected ring.

The manual radial evidence bands are equal-width normalized-radius buckets and use all observed die. A ranked edge/ring selection is different: its inner and outer radii are fitted from the bounded classifier sample, so its projected count and boundary may differ from the exact evidence panel. Broad cluster selections are centered and sized from radial-adjusted excess failures inside the calibrated search window rather than from the window boundary itself. The ranked detector uses a 24 by 24 normalized search grid and requires at least eight sampled die in a window, so very small defects may appear only in the exact hotspot evidence. Reticle detection searches coordinate periods 2 through 12.

Evidence & Advanced Metrics remains visible below the ranked signatures and shows the exact edge/core, radial, quadrant, and hotspot measurements behind them.

The spatial card also turns detected patterns and evidence regions into selection controls.

You can select:

• edge vs core
• radial bands
• quadrants
• hotspots
• detected line, gradient, repeating, or lot-drift regions
• selected wafer or lot-stack scope

These selections flow through the same shared selection state as the other analysis surfaces. Spatial-card clicks add regions to the linked selection and remember each active control. Brushing, replacing, inverting, or undoing a selection turns a control off when any part of its region is no longer selected; clicking it adds the complete region again. All regions and the wafer-family groups use the same remembered toggle behavior, preserve unrelated manual selections, and combine their ranked targets in the current wafer or lot scope. Failing dies behaves the same way for all final failures in that scope; use it for failure-only inspection, not as a spatial-signature claim.

If coordinates are missing, spatial classification is unavailable. Partial geometry uses only rows with observed coordinates and reports the omitted count. When the coordinate frame is not aligned to canonical wafer orientation, absolute direction is withheld and locations are marked as frame-relative.

Best practice

Start with the wafer map if you suspect process structure.

Then:

1. confirm the pattern in spatial analysis
2. push the region as a selection
3. inspect that same population in histogram or scatter