BlogExplainers
Why reverse engineering is hard: 6 problems to solve
Six problems make reverse engineering mechanical parts slow and risky: missing data, worn parts, scan blind spots, dumb solids, slow redraws and proof.
In short
- Scanning is quick. The hard part of reverse engineering is everything after it.
- Six problems slow it down: the design data is gone, the part is worn, the scan has blind spots, a mesh isn’t CAD, the redraw is slow, skilled work, and the model has to be proven right.
- A scan measures geometry only. It holds no tolerances, no material and no design intent.
- Converting a mesh to a solid copies the wear and leaves nothing to edit.
- AI can take on the slow redraw and keep every decision with its evidence, but an engineer still decides what the evidence can’t settle.
Scanning is the easy part
A modern 3D scanner captures a part’s surface quickly and in fine detail. So why does it still take so long to get a part back into CAD, and why do rebuilt parts still come back from the workshop and not fit?
Because the scan is only the start. Reverse engineering a mechanical part means recovering a design that was never handed to you, from a part that has changed since it was made. Six problems stand in the way. Each one below comes with what causes it and what helps. (For the steps themselves, see How reverse engineering works.)
1. The design data is gone
Most parts are reverse-engineered because their drawings are missing, out of date, or never existed. But a drawing holds far more than shape: tolerances, datums, material, heat treatment, surface finish, thread specs and notes. A scan recovers none of it. It measures geometry, and only geometry.
What helps:
- Collect everything that survives: old drawings, even superseded ones, the mating parts and their drawings, the machine’s manual, and the people who maintained it.
- If the material matters, test the part. Hardness tests and chemical analysis can tell you; a scan can’t.
- Write down every assumption, so nobody mistakes it for a fact later.
2. The part isn’t the design
The part in your hand was made within a tolerance, then used. Running faces have worn, edges have been knocked, thin walls may have moved. A scan measures all of it faithfully. Average it into the model and you’ve copied the damage.
Even a perfect, unused part only shows where its size landed inside the tolerance, not the nominal value the designer chose.
What helps: measure where the part doesn’t wear, prefer standard sizes and fits, use the mating parts as evidence, and flag what the evidence can’t settle. More in Design intent: why a worn part isn’t the drawing.
3. The scan doesn’t see everything
An optical scanner has to see a surface to measure it. Shiny, black and transparent surfaces scatter or swallow the light. Deep bores, narrow slots and undercuts hide from the camera. Internal passages are invisible. Threads, small fillets and sharp edges sit close to the scanner’s resolution, so they come out rounded or noisy.
What helps:
- A thin coat of scanning spray on difficult surfaces, and more setups and angles.
- Measure small or critical features by hand or on a CMM.
- CT scanning for internal geometry.
- Treat a gap in the scan as unknown, not as a surface to smooth over.
4. A mesh isn’t CAD
A scan becomes a mesh: thousands of triangles that record the shape and nothing else. There are no planes, no cylinders and no dimensions. Convert it to a solid and each triangle becomes a face. The file opens in CAD, but you can’t change a diameter, CAM has no clean surfaces to work from, and a simulation has to mesh every facet.
What helps: rebuild the part as features instead of converting it. Ask for a STEP file for the solid and a native CAD file for the feature tree, because STEP doesn’t carry the tree. More in Scan to CAD, explained and STEP or IGES?
5. The redraw is slow, skilled work
Turning a mesh into features is where the time goes. Someone has to cut sections, sketch profiles over them, decide which face is the datum, choose the order of the features and pick every nominal size, one decision after another. That takes a person who reads both a scan and a drawing, and specialist software that costs money and takes training.
Because the work is manual, it’s also hard to review. The decisions live in the engineer’s head, not in the model.
What helps: record each decision and its evidence as you go, keep the feature tree, and automate the repetitive interpretation (see below).
6. Proving the model is right
A model can overlay the scan neatly and still be wrong where it counts: a bearing seat a fraction of a millimetre off, holes on the wrong pitch circle. And a model that matches the scan everywhere has usually copied the wear. Either way, the mistake shows up at the machine or at assembly.
What helps: align the model to the scan on its datums rather than by best fit, read the deviation map by category, confirm each key dimension with independent evidence, and change a parameter to check the model follows. More in Is your rebuilt CAD right? Check it before you cut metal.
A checklist before you start
Six questions, one for each problem, to answer before the first scan:
| Question | Why it matters |
|---|---|
| What survives: drawings, mating parts, manuals? | Each one is evidence the scan can’t give. |
| Which surfaces are worn, and where doesn’t it wear? | Tells you where to measure the design, not the damage. |
| Which surfaces will the scanner struggle with? | Plan spray, extra setups, a CMM or CT before you start. |
| Do you need a feature tree, or only a solid? | Decides whether you rebuild the part or convert the mesh. |
| Who decides the dimensions the evidence can’t settle? | Someone has to own each decision, and record it. |
| What tolerance does each functional feature need? | Sets how you check the model, and what close enough means. |
What AI changes, and what it doesn’t
AI helps most with problems 4 and 5, the slow and repetitive part: reading the mesh, finding the datums, fitting features, proposing standard sizes and building the feature tree. It helps with problem 6 too, by keeping every decision with its evidence.
It can’t solve problems 1 to 3 on its own. No model can recover a tolerance or a material that was never measured, or see inside a bore the scanner missed. That’s why Oscar, our AI copilot for reverse engineering, flags each dimension it can’t prove and shows the evidence. Oscar proposes; the engineer decides.
Try it in your browser at app.originshape.design, or tell us about a part you need back in CAD.
Sources and further reading
- 3D scanning. Wikipedia, on how optical scanners work and where they struggle.
- Industrial computed tomography. Wikipedia, on scanning internal features.
- Engineering tolerance. Wikipedia, on why every made part differs from its nominal size.
- Polygon mesh. Wikipedia, on what a mesh of triangles stores, and what it doesn’t.
- Geometric dimensioning and tolerancing. Wikipedia, on datums and tolerances (ASME Y14.5, ISO 1101).
- How reverse engineering works: a part, step by step. Our post on the seven steps, from the brief to delivered CAD.
Questions
Why is reverse engineering so difficult?
Because a scan only measures the part as it is today. The tolerances, material and design intent are missing, the part is worn, some surfaces are hard to scan, and turning a mesh into editable features is slow, skilled work.
What are the limitations of 3D scanning for reverse engineering?
Optical scanners struggle with shiny, black and transparent surfaces, deep bores and undercuts, and anything internal. Small features such as threads and fillets come out rounded or noisy. And no scan records tolerances, material or design intent.
Why can’t I just convert the STL to a solid?
You can, but the solid has one face per triangle and no features or dimensions. It copies the wear and the noise, and you can’t change a diameter. Rebuilding the part as features gives you a model you can edit.
Can a 3D scan tell you a part’s material?
No. A scan measures shape only. Material, hardness and heat treatment need tests on the part, or records such as an old drawing or the supplier’s documents.
Can AI reverse-engineer a part on its own?
Not fully. AI can read the mesh, find datums, fit features and propose nominal sizes. It can’t recover data that was never measured, so an engineer still confirms the uncertain dimensions. Oscar flags those with their evidence.
