BlogGuides
Is your rebuilt CAD right? Check it before you cut metal
How to check reverse-engineered CAD before the part is made: functional surfaces first, a deviation map read the right way, and every key dimension confirmed.
7 min read
By OriginShape
In short
- Check a rebuilt model in order: a valid solid, the right datums, the functional surfaces, the deviation map, a parameter change, the fit with mating parts, then sign-off.
- Check functional surfaces first: the diameters, bores and faces that locate or mate. A model can be right everywhere else and still fail there.
- Read a deviation map by category: within tolerance, outside but explained (wear, damage, scan gaps), or outside with no explanation. Only the last means the model is wrong.
- Don’t chase zero deviation. A model that matches a worn seat exactly has copied the wear. And align on datums: a best fit spreads one error across the whole part.
- Every key dimension needs its value, how it was confirmed, and who confirmed it. Anything unconfirmed stays flagged until an engineer decides.
Looking right isn’t a check
A rebuilt model can look perfect on screen, overlay the scan neatly, and still be wrong where it matters: a bearing seat a fraction of a millimetre off, holes on the wrong pitch circle, a face square to the wrong axis. Those are the errors that only show up when the part is made and doesn’t fit.
The opposite trap is just as common. A model that matches the scan everywhere has usually copied the wear, dents and scan noise along with the design. So checking a reverse-engineered model isn’t about making the difference zero. It’s about knowing what every difference means.
Check in this order
Work from the basics to the details, so an early mistake doesn’t hide behind a later check:
- It’s a valid solid. Closed, with no gaps, overlaps or self-intersections. Everything else depends on this.
- The datums are right. The model’s axes and reference faces match how the part is located in use: the bearing axis, the mounting face, the bolt circle.
- Functional surfaces first. Check the diameters, bores, seats and faces that mate with other parts before anything cosmetic. A part can be right everywhere else and still fail here.
- Read the deviation map. Compare the model with the scan, and give every area outside the tolerance a reason (see below).
- Change a parameter. Edit a key dimension and check that the model updates the way the design intends: the taper follows the diameter, the holes stay on their pitch circle.
- Check the fit. Against the mating parts, their CAD, or their drawings, wherever you have them.
- Record and sign off. Every key dimension with its value, how it was confirmed, and who confirmed it.
How to read a deviation map
A deviation map colours the model’s surface by the distance between the model and the scan, after the two are aligned. It’s the fastest way to see where they disagree. Read it with two things in mind.
How the two were aligned. A best-fit alignment moves the scan until the overall distance is as small as possible. That spreads one real error across the whole part, where it looks like a little noise everywhere. For a functional check, align on the datums instead: the bearing axis, the mounting face. Then a wrong diameter shows up as exactly that.
What each colour means. Read the map by category, not by colour:
- Within tolerance: nothing to do.
- Outside, and explained: wear on a running surface, a dent, a gap or noise in the scan on a shiny or deep surface. The model is right; the part has changed, or the scan couldn’t see it.
- Outside, with no explanation: this is the one that matters. The model is probably wrong. Fix it, or measure the real part again.
- Not decided yet: a dimension the evidence couldn’t settle. It stays flagged until an engineer confirms it.
On the example shaft, the worn seat on the journal sits outside the tolerance, and that’s expected: the model holds the nominal Ø24.00 mm and the part has worn. A model that followed the scan into that seat would pass the deviation map and be wrong.
Confirm every key dimension, and say how
A deviation map shows where the model and the scan disagree. It doesn’t say which dimensions you can trust. For that, list every key dimension with how it was established:
| Dimension | Value | How it was confirmed | Status |
|---|---|---|---|
| Journal diameter | Ø24.00 mm | Measured where the journal doesn’t wear | Confirmed |
| Collar diameter | Ø36.00 mm | Cross-section at z −10.00 | Confirmed |
| Taper | 21.80° | Follows from Ø36 to Ø28 mm over 10 mm | Derived |
| Shoulder position | z −35.00 mm | Proposed by Oscar | Waiting for the engineer |
Good ways to confirm a dimension: an old drawing, the mating part, a standard or catalogue size, or a direct measurement of the real part with a caliper, micrometer or CMM. “It looked right on the scan” isn’t one of them. Anything you can’t confirm stays flagged, with the evidence you have, until an engineer decides.
How close is close enough?
There’s no single number. The tolerance comes from what the part does: a bearing seat needs a fit with its bearing, a cover plate needs to line up with its holes. Take it from the drawing if one exists, from the mating parts, or from the standard fits for holes and shafts in the ISO 286 system.
Two limits apply whatever the tolerance. A scan can’t confirm a dimension more closely than the scan itself was measured, so measure the tightest dimensions on the real part. And a worn part can’t tell you where its size sat inside the original tolerance; that’s a design-intent decision (see Design intent: why a worn part isn’t the drawing).
Where Oscar helps
Oscar, OriginShape’s AI copilot for reverse engineering, runs the first checks with the rebuild and keeps the open questions in view. In the example on our home page, its checks read:
- ✓ The model is a valid solid.
- ✓ A change to each parameter updates the model.
- ! Confirm the shoulder position. Oscar proposed z −35.00 mm.
Each dimension carries where it came from, so the check record above starts filled in. The engineer confirms what’s open, and Oscar doesn’t commit a proposed value until they do. (Example values from the illustrative part, not a screenshot of the app.)
Have a rebuilt model you’re not sure about, or a part you need back in CAD? Tell us about it.
Sources and further reading
- Iterative closest point. Wikipedia, on the best-fit alignment of a scan to a model.
- Engineering fit. Wikipedia, on hole and shaft fits and the ISO 286 system.
- Geometric dimensioning and tolerancing. Wikipedia, on datums and tolerances (ASME Y14.5, ISO 1101).
- Coordinate-measuring machine. Wikipedia, on measuring the dimensions that need it most.
- Measurement uncertainty. Wikipedia, on why every measurement has a range.
- Design intent: why a worn part isn’t the drawing. Our post on recovering nominal dimensions from a worn part.
Questions
How do you check a reverse-engineered CAD model?
In order: confirm it is a valid solid, check the datums, check the functional surfaces, read the deviation map against the scan, change a key parameter to see the model update as intended, check the fit with the mating parts, then record how each key dimension was confirmed.
What is a deviation map?
A colour map of the distance between a CAD model and the scan of the part, after the two are aligned. Read it by category: within tolerance, outside but explained by wear, damage or scan gaps, and outside with no explanation, which usually means the model is wrong.
Should a reverse-engineered model match the scan exactly?
No. Worn, dented or poorly scanned areas should differ from the model. A model that matches the scan everywhere has usually copied the wear and noise along with the design.
What deviation is acceptable in reverse engineering?
There is no single number. The tolerance comes from what the part does: take it from a drawing, the mating parts, or standard fits such as ISO 286. Measure the tightest dimensions on the real part, because a scan can’t confirm a dimension more closely than it was measured.
Should I use a best-fit or a datum alignment to check a model?
Use a datum alignment for functional checks. A best-fit alignment spreads one real error across the whole part, where it looks like noise; aligning on the datums shows it where it is.
