From 2D Drawings to 3D Models: The Limits of Automating the Conversion

What you can extract from a drawing is the part's current state: this length, this hole diameter, this bend angle. What you cannot extract is why those values are what they are.

A concrete example: the drawing shows a 120 mm distance. Is that distance

• a fixed assembly dimension,
• half of the overall width,
• or 120 because it must sit 40 mm from the adjacent hole?

All three produce the same line. But the moment you want the part to be parametric, all three behave completely differently. Change the width and the first stays put, the second re-centres, the third follows the hole.

The model an automatic conversion produces usually picks the first assumption — that is, a dead solid. The picture is correct, but change a parameter and the model falls apart.

None of this means the conversion cannot be automated. Large parts of the chain automate very well:

• Cleanup. Broken lines, overlapping entities, unused layers. Tedious and rule-shaped — the part best suited to automation.
• Recognition. Distinguishing closed profiles, hole groups, and bend lines.
• Classification. Determining which product family a drawing belongs to. The rule layer is selected from that.
• Draft solid. Producing the first solid with the correct profile and thickness.

Together those four take most of the mechanical load. What remains — deciding what the parameters attach to — is a design decision.

AI makes two genuine contributions here, and one that is overstated.

Genuine contribution 1 — recognition. The image-processing side beats hand-written rules comfortably. For "is this a sheet metal bend line", pattern recognition is strong.

Genuine contribution 2 — classification. Grouping thousands of drawings by product family is a job that takes weeks by hand.

The overstated part — intent inference. The claim that "AI understands design intent" is too optimistic today. A model can guess which dimension is the driver; accept that guess without verifying it and you have buried the error inside a parametric model. And it is exactly the silently-wrong-output kind of error — it surfaces months later, while producing a different variant.

1. Start narrow. One product family, one variant. Run it end to end.
2. Define intent by hand. Which dimension drives that family is decided once, then applied across it.
3. Automate verification. Rebuild the produced model at several parameter values; if it falls apart, the intent is wrong.
4. Then widen. Move to the next family only after one works reliably.

This is the four-step method from the first post in this series, applied to this specific problem.

What automates in the 2D-to-3D move is cleanup, recognition, and draft generation. What does not automate is how the dimensions relate to each other — because that information was never in the drawing.

The productised version of this approach is described in the ParametriX case study.