What Makes a CAD Model "Manufacturable"? Our Validator Explained

A generated part can render perfectly in a browser-based 3D viewer and still be geometrically unusable — a shell with a gap a viewer's shading hides, a wall thinner than any tool can cut, a feature smaller than the radius of the cutter meant to make it. "It looks like a part" and "a machine shop can produce this" are different claims, and our pipeline is explicitly scoped to CNC machining and sheet-metal cutting, not general-purpose 3D printing or organic forms — so the validator checks for the specific ways geometry fails in that domain, not just whether a mesh is closed.

Watertight and manifold: the prerequisite, not the goal

The first checks are the most basic and the most non-negotiable: is the solid watertight (no gaps or holes in its boundary) and 2-manifold (every edge borders exactly two faces, with no ambiguous or self-overlapping surfaces)? These are run against a fine tessellation of the generated B-Rep. This isn't a manufacturability check so much as a solidity check — a shape that isn't watertight and manifold isn't a valid solid at all, and nothing downstream (STEP export, DXF flattening, machining) is meaningful until this passes. Alongside it, a self-intersection check confirms the geometry doesn't fold back on itself in ways that would confuse a boolean operation or a slicer. Both are hard fails: if either one trips, the part doesn't ship — it goes back through the repair loop.

Wall thickness: sampling rays inward

A shape can be perfectly watertight and still have a wall too thin for a mill or a laser to cut without warping, breaking, or simply not existing at the intended thickness. The validator checks minimum wall thickness by inward ray sampling: casting rays inward from points across the surface and measuring how far they travel before hitting the opposite side of the material. Thin spots show up as short ray distances. This check is a soft fail — a part with a marginal wall doesn't automatically get rejected and re-generated, since sometimes the design is intentionally thin and workable with process changes, but it does ship with a warning in the report so a human reviewer or downstream automation can flag it before cutting.

Planar-dominant faces and sheet-metal feasibility

Some parts are meant to be laser or plasma cut from flat stock rather than milled from a block, and the difference matters for which export the pipeline should even produce. The validator checks for a planar dominant face — one flat face covering more than 60% of the part's surface area — as the signal that a DXF flat-pattern export makes sense at all. If no such face exists, DXF emission is skipped rather than producing a meaningless flattening of a fundamentally 3D shape. When a dominant face is present, a further sheet-metal feasibility check attempts a flat unfold using OpenCascade's offset/unfold machinery, another soft fail that annotates the report rather than blocking the part outright.

Feature size versus tool radius

The last check ties geometry back to the physical reality of cutting it: comparing the smallest internal radius in the part against the radius of the tool that would actually cut it (a configurable value, defaulting to 1mm). A tool cannot cut an internal corner sharper than its own radius — physically, the cutter traces an arc no smaller than itself. A part with, say, a 0.3mm internal fillet specified for a 1mm end mill is asking for something no real tool in that setup can produce. Like wall thickness, this is a soft fail: informative rather than blocking, since it depends on tooling choices the pipeline doesn't always know in advance.

Hard fails versus soft fails

The distinction running through all of this is deliberate. Hard fails — failed STEP export, non-watertight geometry, self-intersection — mean the part isn't a valid solid, so there's nothing to ship; these trigger the repair loop, which feeds the traceback or failure reason back to the model for a corrective retry. Soft fails — thin walls, tooling mismatches, sheet-metal infeasibility — mean the part is a valid solid but has a manufacturing caveat; these ship anyway, with the caveat surfaced as a warning annotation in the report rather than silently hidden or used to block a result that might still be useful. That split keeps the validator honest about the difference between "this isn't a real shape" and "this is a real shape you should look at before you cut it."