AI manufacturability review: catch 3D print failures before you slice

By Spencer Rose

Most failed 3D prints are decided before the slicer ever runs — a wall too thin for the nozzle, an overhang the printer can't bridge, an internal corner that cracks the first time the part takes a load. An AI manufacturability review catches that class of problem while the part is still parametric in CAD, so you fix a number instead of reprinting an afternoon later.

Cadre runs one before you export an STL. It reads the model, flags the spots likely to give you trouble, and hands back fixes you can apply with one click. This post is what the reviewer actually checks, how it works, and — just as important — what it doesn't do. (For the manual version of the same rules, see design for 3D printing.)

How review-before-you-slice changes the loop

The normal loop is model → export → slice → print → inspect → go back and fix. Every trip costs filament and an afternoon.

The review happens one step earlier: model → review → fix → export. The check runs in CAD, where a fix is a parameter change, not a re-print. It doesn't replace looking at the part in your hand — it's a cheaper place to catch the problems that are obvious in the geometry but invisible until they fail.

What an AI manufacturability review checks

Every flag falls into one of four kinds, each tagged blocker, warning, or info so you can triage by what actually stops a print versus what merely makes it better.

Thin walls

Below the threshold for your process, a wall either won't print or won't hold. On FDM that's roughly a 0.4 mm floor (a single perimeter); under about 0.8 mm you're in warning territory. SLA goes thinner. A blocker here means "this won't print as drawn."

Sharp internal corners

A zero-radius inside corner is a stress riser — the first place a functional part cracks under load, and a spot FDM layers bond poorly. The fix is usually a small fillet.

Overhangs

Steep unsupported overhangs sag, string, or fail outright. Past roughly 45° from vertical you're relying on support; past 60° over any real span, the printer is unlikely to bridge it cleanly. The reviewer flags the ones your process won't handle unsupported.

Small features

Detail near the resolution limit of the process — a 0.2 mm boss, a hole narrower than the nozzle, embossed text too fine to resolve — disappears or prints as mush.

It calibrates to your process (FDM, SLA, CNC, sheet metal)

A wall that's a blocker for SLA can be fine for CNC. So the thresholds shift with the material preset you pick — FDM, SLA, machined, or sheet metal. Tell it you're milling aluminium and it stops warning about overhangs that only matter for printing and starts caring about internal corner radii a cutter can't physically reach. Same model, different review, depending on how you're making it.

What the AI actually reads

This is the part most "AI for CAD" pitches stay vague about, so I'll be specific.

The reviewer does not see a rendered image and does not run a slicer simulation. It reads your feature tree — the operations that built the part, with their parameters — alongside facts the geometry kernel measures directly:

  • Wall thickness is measured, by casting rays inward through the mesh to find the thinnest sections. (On dense, high-poly meshes that minimum is an approximation over a sampled subset, not every triangle.) A thin-wall flag has to cite the measured number — "measured min wall 0.42 mm, increase to 1.2 mm" — not a guess from proportions.
  • Overhangs are measured too: the kernel computes the overhanging area and the worst angle against the build direction.
  • Sharp internal corners and small features are inferred from the geometry — edge-convexity counts and face areas — rather than exhaustively measured.

So when it calls a wall thin, that's a measured millimetre, not a vibe. What the AI adds on top is the judgment: which measured number actually matters for your process, how severe it is, and the single parameter change that fixes it. It's tuned to speak up only when confident, and "nothing to flag" is a valid answer.

Suggestions you accept with one click

Alongside the flags, the review returns parameter suggestions — single-dimension changes tied to a specific feature. "Extrude distance 20 → 25 mm, because the current wall is below the FDM floor." Each names the feature, the current value, the suggested value, and the reason. Accept it and the parameter updates; reject it and nothing happens.

It keeps these to one parameter at a time and won't redraw a sketch or refactor your tree. Cadre has no constraint solver, so features don't know about each other's shared edges — and the reviewer is told not to suggest a change that would silently break an alignment you authored on purpose (widening a wall that's flush against another, say). When a change is risky it gives you the flag and a textual fix and leaves the parameter alone. A suggestion that quietly breaks your model is worse than no suggestion.

What it doesn't do

The gap between what a tool claims and what it does is where trust dies, so plainly:

  • It doesn't run a slicer. No layer simulation, no print-time estimate, no support generation.
  • It doesn't redesign the part. It tweaks parameters and points at problems; the design decisions stay yours.
  • It doesn't replace the print in your hand. Some failures only show up physically. It catches the class that's easier to see in CAD than after a failed print — that's the whole claim.
  • It can miss things, and occasionally flag a non-issue. It's tuned for fewer, higher-signal flags over a wall of noise.

FAQ

Does it measure wall thickness, or guess it?

It measures it — the kernel casts rays inward through the mesh to find the thinnest walls, and a thin-wall flag has to cite that measured number. The one caveat: on dense, high-poly meshes the minimum is an approximation over a sampled subset, so for safety-critical tolerances still verify in a dedicated tool. Sharp corners and tiny features, by contrast, are inferred from the geometry rather than measured.

Does it replace a slicer?

No — run it before the slicer. It catches design problems; the slicer handles toolpaths, supports, and print time.

Will it change my design without asking?

Never. Every suggestion is a proposal you accept or reject one at a time.

Try it before your next print

Model a part, pick your process, and run a review before you export. If it catches something you'd have caught yourself, it cost you ten seconds. If it catches something you wouldn't have, it saved you a print.

Cadre is free during the alpha — open it in your browser and run it on whatever you were about to slice next.