Every panel that reaches a placement machine is judged in the first seconds of the program, and not by the parts or the paste. The camera either finds three marks it recognises or it does not. When it does not, the machine falls back to the panel edge, placement drifts by a tenth of a millimetre, and the failure shows up at AOI on the smallest bodies first.
Fiducial mark requirements are short, and they are still the most common reason a new panel goes back for revision before the first build, because the marks are added last and treated as a formality.
What the camera is actually looking at
A fiducial is a target of known size at a known position. The machine compares where the marks are with where the panel data says they should be, and corrects offset, rotation and stretch across the whole panel. Three marks give scale in two directions plus rotation. Two marks fix position and rotation but say nothing about scale, which is why a long panel with two marks drifts at the far end.
The machine does not read a mark the way an eye does. It looks for high contrast with a defined edge and calculates a centre from that edge. Anything that softens the edge moves the centre.
Size, shape and the clearance around the dot
The dot is usually 1 mm across on the copper with a mask opening of 2 mm, leaving a clear ring of bare laminate. A 0.5 mm dot can work on a clean panel and fails on a matte finish. Round is the safe shape. A square mark has four candidate centres when one edge is partly covered.
Clearance matters as much as size. If the mask opening is barely larger than the dot, mask residue narrows the ring and the mark becomes unreliable. Anything inside the search window, a via annulus, a silkscreen line or a test pad, will pull the calculated centre off.
Contrast is a material decision
On ENIG the pad is bright and the laminate is dark, the easiest combination for a camera. On bare copper with a mask opening over OSP, the copper darkens and the edge softens within days of packing. On lead free HASL the surface is not flat, and a domed mark throws focus off.
No finish here is wrong. The mark has to be checked on the finish that will ship, not on the drawing. A panel that reads perfectly on an ENIG sample can fail on the production finish six weeks later.
Count and position on the panel
Three marks per panel, and three per array where the array is large enough, all outside the area that will be depanelised. A mark on a rail that is later routed away disappears between the first build and the second.
They should sit as far apart as the panel allows, non collinear, clear of the tooling holes and the clamping zone. Three marks along one edge is the same as having two.
Global marks against local marks
Global marks correct the panel. Local marks correct a single fine pitch device, and matter where the part is 0.4 mm pitch or the body is large enough that panel level correction is not. A local pair sits diagonally outside the pad field, on the same copper layer as the pads.
Local marks are needed where one placement error is expensive, which is a decision per part.
Where panels most often fail
| What the first build shows | Cause | What to change |
|---|---|---|
| The machine aligns to the panel edge and never uses the marks | Mark too small, or mask opening barely larger than the dot | Open the mask to 2 mm over a 1 mm dot and keep the ring free |
| Placement is good at one end of the panel and off at the other | Only two marks, so scale is not corrected | Add a third mark, non collinear, as far apart as the panel allows |
| Alignment passes on the sample and fails on the production lot | Finish changed and edge contrast dropped | Check the mark on the shipping finish, not on the sample |
| The mark is found on the first build and missing on the second | Mark sits on a rail that is routed away | Move the marks onto the panel body or a retained rail |
| A fine pitch part is offset while the rest of the panel is good | Panel level correction only, no local marks | Add a diagonal local pair outside the pad field |
Panels that arrive without usable marks
A panel with no marks can still run. The result is what the customer sees later: a wider placement distribution, more rework on the fine pitch positions, and a first article that passes because it was measured against itself. Three marks cost nothing next to a week of rework.
How SUPERB PCBA handles it
Every new build starts with a first article check on the machine that will run the order, where the placement program is confirmed against the customer data on a measured panel, and the alignment result is recorded rather than assumed. Seven SMT lines run that check before production, so a weak mark is found on one panel instead of one thousand.
Where a panel arrives from a customer drawing without usable marks, the point is raised before release and the fix agreed with the customer. The same review covers rails, tooling holes and mask clearances, because a good mark on a panel with no clamping area still will not run. That side is covered under tooling and stencil control, the panel construction under panelization, V-score and tab routing, and builds that mix reflow with through hole parts under mixed technology assembly.
Questions buyers ask
Can a test pad be used as a fiducial?
Yes, if it is round, exposed, isolated inside the search window and never covered with solder. In practice pads get probed, fluxed and tinned, and the reflection changes.
Is a mark needed on every board in the panel?
Usually not. Panel level marks correct the array and the boards inherit that correction. Local marks are for parts that cannot tolerate the residual error.
Do V-cut panels need different marks?
The marks have to survive depanelisation, which means keeping them off the cut line and off the rails that are removed. The cut does not damage a mark that sits clear of it.
Next step
Send the panel drawing with the Gerber set and the BOM. The review checks the marks against the finish, the array and the depanelisation method, and comes back with the specific change if one is needed, before a panel is built. If the order is already running, the first article record shows what the machine actually aligned to.