DEV Community: Fabdose

Do You Need to Hire a DFM Engineer Before You Commit to a Mold?

Fabdose — Thu, 17 Sep 2026 12:48:07 +0000

Two undercut faces. Eleven faces at zero draft. Three sections over the material maximum. One STEP file, one pass, a little over two seconds on my laptop, and every one of those findings named by its face number.

That is the narrow question about a part, closed. It is worth separating from the wide question, because of what three different people were each about to spend in order to get the narrow one answered.

Three unrelated threads, one week. One had a 3D printed product already selling and was about to buy an injection molding add-on for their CAD package, specifically so they could check draft angles themselves before committing to a mold. One had a finished industrial design and STEP files, and was hiring a product engineer specifically to de-risk manufacturability before manufacturing. One had STEP files and a BOM finished, and was shopping for a full-service manufacturing partner partly because that partner would catch fit and assembly problems before commit.

Different budgets, different stages, same underlying question: is there something in this geometry that is going to blow up when I commit to a tool?

That question has a narrow half and a wide half, and they have very different prices. The wide half is genuine engineering judgment and it is worth paying for. The narrow half, whether any face in the model breaks a rule that every toolmaker applies anyway, is a measurement. Measurements do not require a hire, a software purchase, or a vendor relationship.

What part is being tested?

A barrel-style housing body, 100mm long, 26mm wide, 14mm tall, 16 faces. I built it as a test part, so this is a demonstration and not somebody's product. It is deliberately unkind: it has a cylindrical internal pocket that curls back under itself, and nothing on it was drafted.

Run through the Fabdose engine as ABS, injection molding.

What does an undercut look like in a report?

Two faces came back critical:

Critical undercut, requires complete redesign or advanced mold technique detected at face 0 (angle: 90.0°)
Critical undercut, requires complete redesign or advanced mold technique detected at face 4 (angle: 90.0°)

Face 0 is 817 square mm. Face 4 is 2,704 square mm. Both are cylindrical, both sit at minus 90 degrees to the pull direction, which is the geometric definition of the problem: the mold cannot come apart in a straight line without tearing that surface. Undercuts are also the finding most sensitive to which way you decided the tool opens, which is a choice worth making deliberately rather than inheriting from how you happened to model the part. There is a separate teardown on which way the mold should open, and another on the standard ways to design an undercut out.

The tooling model then does the thing that makes this concrete rather than academic:

Slide core   Undercut at face 0    $4,000
Slide core   Undercut at face 4    $4,000
Base mold (moderate complexity)   $10,000
Total mold cost                   $18,000

That is a model estimate with a region and complexity tier baked in, not a quote from anyone. Treat the 18,000 as arithmetic on assumptions, not a price. What is not an assumption is the shape of the arithmetic: two undercut faces, two moving sub-assemblies in the tool, and the base cost is no longer the cost.

This is where the industry numbers line up with the estimate. A single side-action pull adds roughly 50 to 80 hours of skilled toolmaking labor, an internal lifter 100 to 200 hours, a threaded unscrewing device 200 to 300 hours. Side-pulls need clearance from the cavity plate edges, normally 7.5cm to 15cm added to plate length and width, which pushes you into a larger and pricier mold base. The moving parts also restrict where cooling channels can run near the undercut, which makes hot spots, which extends cooling time, and they add metal-on-metal wear points that need maintenance and eventually flash.

None of that is exotic knowledge. It is the reason the first thing a toolmaker looks for is whether your part pulls straight.

How many faces failed the draft check?

Eleven of sixteen, each named individually:

Draft angle 0.0° at face 2 is below minimum 2.0°
Draft angle 0.0° at face 3 is below minimum 2.0°
Draft angle 0.0° at face 5 is below minimum 2.0°
Draft angle 0.0° at face 6 is below minimum 2.0°
Draft angle 0.0° at face 7 is below minimum 2.0°
Draft angle 0.0° at face 8 is below minimum 2.0°
Draft angle 0.0° at face 9 is below minimum 2.0°
Draft angle 0.0° at face 11 is below minimum 2.0°
Draft angle 0.0° at face 12 is below minimum 2.0°
Draft angle 0.0° at face 13 is below minimum 2.0°
Draft angle 0.0° at face 14 is below minimum 2.0°

Worth being precise about the threshold, because it is the kind of number that gets overstated. The established absolute minimum for a smooth, untextured ABS wall is 0.5 degrees per side, with 1.0 to 2.0 degrees preferred for clean ejection. Fabdose checks against 2.0, the conservative end of that band. If your part sat at 0.8 degrees, reasonable people could argue about whether that is a finding.

Nobody argues about zero. A wall at 0.0 degrees fails the strict threshold, the preferred range, and the most permissive baseline in the industry, all at once.

What goes wrong physically is worth knowing, because it explains why this is not a cosmetic nit. Cooling resin shrinks onto the steel it surrounds. A drafted wall loses contact across its entire surface the moment the mold begins to open. A zero-draft wall stays pressed against the steel for the full length of the ejection stroke, so you get drag marks, scuffing and scratches down the face, and the force needed to shove the part off the core shows up as ejector pin indentations, stress whitening, warpage or cracking. At the bad end the part die-locks and seizes in the cavity.

This one also cannot be pushed onto the molder, and the reason is a fit problem rather than a manufacturing one. Draft lives in the CAD master data, and adding a degree moves real material: on a typical draw depth it shifts the wall by roughly 0.017 inches per inch of depth per degree, which is more than enough to break a fit with the part sitting next to it. So if steel gets cut to add draft without the model being updated, nominal walls drift from the engineering definition and mating tolerances go with them. Toolpaths and CMM inspection routines are both generated from the model. The model is what has to change, which means it comes back to the designer regardless of who spots it.

Where are the thick sections?

Wall thickness 13.00mm at face 0 exceeds the 3.56mm maximum for ABS
Wall thickness 11.00mm at face 1 exceeds the 3.56mm maximum for ABS
Wall thickness 11.00mm at face 4 exceeds the 3.56mm maximum for ABS

Three sections, each located. The run ties them forward to a consequence rather than leaving them as a list: sink mark risk came back high, with the root cause given as wall thickness and the affected areas given as faces 0, 1 and 4. The same three faces. The cycle time estimate follows the same physics, with cooling at 201.7 seconds out of a 204.7 second cycle, which is about 17.6 parts per hour. Thick sections are slow sections. That is what makes them expensive twice, once in the defect rate and once in the cycle. For the mechanism behind sink marks, and what coring out a thick section does to the report, there is a separate teardown on where the thick section actually is on a part.

Does a different material fix any of this?

This is the test that separates a geometry problem from a material opinion, so I ran the same unchanged STEP file again as PP.

Finding	ABS	PP
Undercut faces	2 (faces 0, 4)	2 (faces 0, 4)
Faces below draft minimum	11	11
Slide cores in tooling estimate	2	2
Total tooling estimate	18,000 USD	18,000 USD
Wall thickness maximum	3.56mm	3.81mm
Sections over that maximum	3	3
Extra note	none	2.00 percent shrinkage, compensate dimensionally

Changing material changed the numbers the part is judged against. It did not change a single thing about whether the part comes out of the tool.

So what were those three purchases actually for?

Here is the honest split, because this is the part it would be easy to lie about.

What the geometry check answers. Does any face undercut the pull. Does any wall stand below the draft minimum. Does any section exceed the material maximum. Where each of those is, by face, with area and coordinates. Roughly what the first two do to the tool. That question is closed in seconds, locally, from a STEP file, before anyone is committed to anything.

What it does not answer, at all. Whether this is the right design. Whether the wall you drafted still fits the part it mates with. Whether your tolerance stack survives assembly across a production run. Whether this vendor is the right vendor. Whether the tool should be a single cavity or a family tool. What to do when two constraints genuinely conflict and someone has to make a call and own it.

The second list is what an engineer is for, and it is not a list a geometry checker is ever going to take over. The distinction is not tool versus human. It is measurement versus judgment.

The reason the split matters commercially is sequencing. If you buy the add-on, make the hire, or pick the full-service vendor and the first thing that comes back is eleven zero-draft faces and two undercuts, you have spent the budget to receive a result that was sitting in your own file the whole time. Close the measurement first, cheaply. Then spend the engineering hours on the questions that actually need an engineer, with a model that is not wasting their first afternoon.

An experienced engineer will spot these findings too, of course. They will just spot them on your clock.

How do you run this on your own file?

Fabdose is a desktop app. The STEP file is analyzed on your own computer and is never uploaded, which matters at exactly this stage, when you are still evaluating vendors you have not signed with. Load the file, pick a material and a process, and read the per-face list.

If the result is clean, you have removed a category of risk for the cost of a couple of minutes. If it is not clean, you now have a named list of faces, which is a far better thing to hand an engineer than a file and a worry.

Check your design with Fabdose

Frequently asked questions

Do you need to hire a DFM engineer before committing a part to a mold?

Not to answer the narrow question of whether the geometry has an obvious problem. That part is a measurement: does any face undercut the pull direction, does any wall stand at zero draft, does any section exceed the material maximum. A geometry check answers that from the STEP file. You still need an engineer for the questions that are judgment rather than measurement, including whether the part is the right design, whether tolerances survive assembly, which vendor to use, and who owns the decision to cut steel.

What does an undercut cost you in an injection mold?

An undercut is any feature that blocks straight-pull separation along the mold opening direction, so the toolmaker has to put a moving sub-assembly in the mold: a side-action slide, an internal lifter, or an unscrewing device. Industry figures put a single side-pull at roughly 50 to 80 hours of skilled toolmaking labor, an internal lifter at 100 to 200 hours, and a threaded unscrewing device at 200 to 300 hours. Side-pulls also need clearance from the cavity plate edges, typically adding 7.5cm to 15cm to plate length and width, which forces a larger and more expensive mold base. On a 16-face test housing with two 90-degree undercut faces, the tooling model put two slide cores at 4,000 USD each on top of a 10,000 USD base, for an 18,000 USD model estimate. That is not a quote: it assumes a region and a complexity tier rather than reflecting a real toolmaker's bid.

What is the minimum draft angle for a smooth ABS surface?

The established baseline absolute minimum for a smooth, untextured ABS wall is 0.5 degrees per side, with 1.0 to 2.0 degrees per side preferred for clean ejection. Fabdose checks ABS against 2.0 degrees, which is the conservative end of that accepted range. The distinction stops mattering at zero: a wall at 0.0 degrees fails every one of those thresholds, including the most permissive one.

What actually goes wrong when a wall has zero draft?

As the resin cools it shrinks onto the steel it is wrapped around. A drafted wall breaks contact across its whole surface the instant the mold starts to open. A zero-draft wall stays in high-friction contact with the steel for the entire ejection stroke, which produces drag marks, scuffing and scratches, and the extra force needed to push the part off the core produces ejector pin indentations, stress whitening, warpage or cracking. In bad cases the part die-locks and seizes in the cavity.

Can the molder just add draft on their end?

No, not on their own. Draft lives in the CAD master data, and adding it changes real geometry: a degree of draft shifts the wall by roughly 0.017 inches for every inch of draw depth, which is easily enough to break a fit with the part next to it. If a molder cuts draft into the steel without the model being updated, nominal walls drift away from the engineering definition and mating tolerances can be destroyed. Toolpaths and CMM inspection routines are both generated from the CAD model, so the model is what has to change.

Does switching material fix a geometry problem?

It changes the thresholds, not the geometry. The same unchanged test part was run twice, once in ABS and once in PP. Both undercut findings survived, all eleven zero-draft findings survived, the same two slide cores stayed in the tooling estimate, and the total stayed at 18,000 USD. The only things that moved were the wall thickness limit the sections were compared against, 3.56mm for ABS versus 3.81mm for PP, and one added note about PP's 2.00 percent shrinkage. A material swap is not a fix for a face that undercuts the pull.

Originally published at fabdose.app/blog/do-you-need-a-dfm-engineer-before-committing-to-a-mold.

How I Detect Undercuts in a STEP File Without Uploading the Geometry

Fabdose — Mon, 14 Sep 2026 03:22:40 +0000

Fabdose is a desktop tool I built that checks whether a CAD part can actually be injection molded, before you send it to a shop for a quote. Drop in a STEP file, pick a process, get a report back in about a minute: which faces have undercuts, which have insufficient draft angle, where the walls are too thin or too thick.

The part I want to write about here is the one requirement I set for myself early on and then had to keep re-earning: the CAD geometry never leaves the computer. The report text is generated by a cloud model, but the actual geometry analysis, the ray casting, the mesh work, the angle math, all of it runs locally. That constraint shaped the undercut detector more than anything else, so this is a post about what that detector actually does.

What an undercut even is, geometrically

In injection molding, the mold has to open in a straight line (the pull direction) and the part has to slide out. An undercut is any face whose geometry blocks that. Picture a part shaped like a hook, or a boss with a lip on the underside: pull the mold straight up and something on the part is now in the way of the mold wall. That's an undercut, and it's exactly the kind of thing that's easy to miss looking at a 3D view on screen and expensive to discover after a mold is already cut.

Faces, not pixels

A STEP file isn't a mesh, it's a boundary representation, a mathematical description of surfaces and edges. Before I can shoot rays at anything, I have to triangulate it, turn each face into a mesh dense enough to sample without missing detail on curved surfaces but not so dense that a part with hundreds of faces takes forever to process.

Once a face has a mesh, the check for that face is conceptually simple: pick sample points across the surface, and for each one, cast a ray back along the pull direction. If the ray is blocked by another part of the same solid before it reaches "outside," that point is shadowed, meaning the mold can't reach it in a straight pull. Enough shadowed points on a face and the whole face gets flagged as an undercut, with a severity that depends on how much of the face is affected and how steep the blocking angle is.

The bug that taught me the most: rays hitting their own face

The failure mode I didn't expect going in was a ray hitting the same face it started from. It sounds like it shouldn't happen, the ray starts just off the surface and travels away from it, but on curved or nearly-tangent faces, floating point rounding from the STEP export can put the ray's origin points close enough to the surface that the very first intersection it finds is itself. The result was faces getting flagged as blocked when they weren't blocked by anything, they were just tripping over their own geometry.

The fix was a self-hit filter: when a ray's first intersection is suspiciously close to its own origin face, on the same solid, within a small tolerance, discard that hit and look at the next one. It's an unglamorous fix and it's the kind of thing that never comes up in a spec, but it moved a real chunk of parts from "false undercut" to "correctly clear."

Why this stays honest about what it doesn't do

Injection molding gets the deepest version of this analysis, face-level undercuts plus draft angle. CNC and sheet metal are included for comparison but aren't run through the same face-level pipeline yet, and I'd rather say that plainly than let the report imply otherwise. A DFM check that quietly overstates its own coverage is worse than useless, it's a check you can't trust the one time it matters.

Why local geometry processing was worth the extra work

It would have been simpler to upload the STEP file to a server, run the same ray casting there, and ship the result back. I didn't do that because the people I built this for are the same people who get nervous about sending an unreleased product's CAD file to a random web service, and they're right to be. Keeping the geometry on the machine means the file never has to leave in the first place. It also means the code doing the ray casting has to be fast enough to run on a laptop in about a minute, which is its own constraint, but it's one I'd rather have than the alternative.

Fabdose is free for the first 5 analyses, no card required, if you want to see what it flags on a part of your own: fabdose.app.

Where Do Product Designers Go to Ask "Can This Be Made?"

Fabdose — Tue, 08 Sep 2026 06:04:52 +0000

You finish a part. It looks right. The proportions are clean, the CAD model is tight, and it solves the problem you were designing for. Then comes the question that stalls every designer at least once: is this actually manufacturable?

The honest answer is that there is no great place to get a fast, reliable, private answer to that question — and the cost of not having one shows up at the worst possible moment.

Where do product designers check if a design can be manufactured?

Product designers check manufacturability through four main routes: sending the file to a factory for a quote with DFM feedback, hiring a DFM consultant or manufacturing engineer, asking an AI assistant about design rules, and running a local geometry-analysis tool. Each option has a different speed, cost, and accuracy tradeoff. None is perfect for every stage of the design cycle — which is why the gap between fast iteration and accurate feedback is a real problem.

The real options, and their real gaps

Ask the manufacturer

The most accurate answer comes from the people who will actually make the part. But getting that answer has a catch: most factories will not perform a detailed DFM review until you request a formal quote. That round-trip, from sending the file to receiving markup on what needs to change, typically takes two to five business days. If you are iterating weekly or faster, waiting for factory feedback on every revision is not realistic.

There is also the exposure problem. Sending a STEP file to a factory you are not yet committed to means your geometry is on someone else's server. For early-stage work or designs with any IP sensitivity, that matters.

And the feedback format varies. Some shops mark up a PDF. Some send an email. Very few give you a structured report you can act on immediately in CAD.

A DFM consultant or manufacturing engineer

The most rigorous option is a person who does this professionally. An experienced manufacturing engineer or DFM consultant will catch things that software misses — material-specific edge cases, tooling constraints that vary by region or supplier, subtle geometric conditions that only matter at certain production volumes.

The gap here is not quality. It is availability. A consultant is not on-call for every design iteration. The engagement model does not fit fast-moving design work. And at typical consulting rates, running a DFM check on every version of a part before it is anywhere near final is simply not economical.

Ask ChatGPT or Claude

AI assistants know a lot about design for manufacturability. Ask about minimum wall thickness for injection molding and you will get a reasonable answer. Ask about draft angles for a given material and you will get useful rules of thumb. For DFM education, they are genuinely good.

But there is a structural limitation that does not get stated clearly enough: these models cannot open your STEP file and analyze the actual geometry.

A STEP file is a precise geometric database. Evaluating manufacturability requires estimating wall thickness from the part's shell geometry, measuring face normals to detect draft, analyzing topology for undercuts relative to a pull direction. These are geometric computations. A language model processes text. When you paste a STEP file's contents into a chat window, the model reads the entity definitions as text and responds with plausible-sounding feedback — but it has never seen your part. It has no idea what your actual wall thicknesses are, where your undercuts are, or which faces are missing draft.

The risk is not that AI gives you bad DFM rules. The risk is that it gives you confidence without geometric grounding. You submit a part thinking it has been checked, and the shop comes back with a list of specific locations where the geometry does not work.

Insight: ChatGPT is a good place to learn DFM principles. It is not a substitute for actual geometric analysis of your specific file, because it has no access to your file's geometry.

Design forums, Reddit, Discord

Community knowledge is real and often generous. Post a render or a drawing in the right forum and experienced engineers and machinists will sometimes catch issues a tool would miss.

But two problems limit this for regular use. First, the response is slow and variable — you might hear back in two hours or two days, and the quality depends entirely on who happens to see the post. Second, you are posting your design publicly. For anything with IP sensitivity, that is a non-starter.

The cost of not having a fast first-pass answer

Most DFM problems are not discovered during design. They surface at two particularly expensive moments.

The first is the quote call. You send a file to a shop, the shop comes back not with a price but with a list of manufacturing concerns. You now have to make revisions, re-send, and wait again. The delay compounds, especially if the shop and your timezone do not overlap.

The second is later: after a part has been through internal review, after stakeholders have seen it, sometimes after it has been communicated to a customer. Learning then that the geometry cannot be made as designed creates rework, schedule pressure, and sometimes project credibility damage.

The earlier a manufacturability issue is caught, the cheaper it is to fix. That is not a new principle. What is missing is a practical way to catch issues early across every design iteration, not just when a deadline forces a formal review.

Warning: The worst time to learn a part cannot be made is in a quote call or after you have presented it to a stakeholder. Early self-check is cheap. Late discovery is not.

Where a local first-pass self-check fits

None of the options above are wrong. A manufacturing engineer is the most thorough option; a factory quote is the most realistic; AI assistants are genuinely useful for learning. The gap is not in any of these tools individually. The gap is in what a product designer can do on their own, quickly, on every iteration, before involving any of them.

What would that look like? Something that runs on your computer without uploading your CAD files. Something that can compare your geometry against multiple process options — injection molding, CNC, die casting, 3D printing — and flag the specific issues your part has in each context. Something that gives you a structured report with locations, not just general advice. And something fast enough that running it on every version is not a burden.

That kind of first-pass self-check does not replace a manufacturing engineer. It does not replace a factory review. What it does is raise the quality of the part that reaches those reviews — so that the feedback you get is about things that actually matter, not about wall thicknesses you could have caught yourself in twenty minutes.

The designer-engineer split in DFM tooling

Most DFM tools on the market were built for engineers. They assume you work in an engineering organization, have access to manufacturing process data, and are comfortable with CAD-integrated analysis that surfaces in the same environment as your FEA and tolerance stack-up work.

Product designers — especially those working in smaller studios, early-stage startups, or consumer product contexts — often work differently. They move faster, they iterate more, they own the design but not always the manufacturing relationship. The tools that exist were not designed with that workflow in mind.

That is the gap Fabdose was built for. It is a desktop app for product designers, not engineers, that runs STEP and STP files locally on your Mac. In one to three minutes, it checks your part against a manufacturing process, flags draft angle, undercut, and process-fit issues with a 3D viewer that pins those at the specific locations in your model, reads wall thickness across the part as a whole for a variation-based warp-risk flag rather than a pinned location, and gives you a rough cost range. The free tier checks one process per analysis; the auto-comparison that runs your part across multiple processes side by side is a Pro feature.

Your CAD files are never uploaded — they stay on your computer. The geometry engine runs locally, which is also what makes the analysis fast; only the defect findings are processed by AI and synced to your account.

A local self-check: draft and undercut issues pinned on the geometry, wall thickness read as a part-wide flag, with fixes -- and your CAD file is never uploaded.

Fabdose is not a replacement for a manufacturing engineer. There are things a person with real manufacturing experience will catch that no first-pass tool will. The honest position is that Fabdose is the step before that — the self-check you run before you involve anyone else, so that when you do reach out to a shop or a consultant, you are bringing them a part that has already been through a first filter.

Tip: Your CAD files are never uploaded — your STEP and STP files stay on your computer throughout the analysis. The defect findings are processed by AI and synced to your account.

Summary

Option	What it gives you	Honest gap
Factory / manufacturer	Most accurate real-world feedback	2-5 day round-trip, requires formal quote request, file exposure
DFM consultant	Expert judgment with material and regional context	Not on-call for every iteration, cost per engagement
ChatGPT / Claude	Good DFM rules and education	Cannot open or analyze STEP geometry, gives confidence without grounding
Design forums / Discord	Community knowledge, sometimes catches edge cases	Slow, public, inconsistent
Local first-pass self-check	Fast, private, iteration-friendly	Not a substitute for expert review, rules-based not judgment-based

The right answer for most product designers is all of these at the right stage. The self-check step belongs at the beginning of the chain, before the slow or expensive options, not as a replacement for them.

If you want to run that first-pass check on your own Mac before involving anyone else, Fabdose is free for five analyses. No credit card required.

Frequently asked questions

How long does it take to get DFM feedback from a factory?

Getting DFM feedback through a factory quote round-trip typically takes two to five business days. Some factories offer faster informal feedback, but a formal quote with a markup list of manufacturing concerns is usually a multi-day process. This timeline does not fit design iteration cycles shorter than a week.

Can AI assistants check if a STEP file is manufacturable?

No. AI language models — ChatGPT, Claude, and similar tools — cannot analyze STEP file geometry. They read the file's text content as entity definitions but cannot reconstruct the 3D shape, measure wall thickness, detect undercuts, or evaluate draft angles from that text. They provide DFM education and rules of thumb, not geometric analysis of your specific part.

Is it safe to send a STEP file to a factory before you are committed to working with them?

Sending a STEP file to a factory before a committed relationship means your geometry is on someone else's server. For early-stage designs or anything with IP sensitivity, this is a real exposure risk. Local geometry-analysis tools that do not upload your file address this by running analysis entirely on your computer.

What is a first-pass manufacturability check?

A first-pass manufacturability check is a geometry analysis of your STEP file that runs on your computer in one to three minutes and flags the most common manufacturing constraint violations — wall thickness under threshold, insufficient draft angle, undercut locations, and process-fit issues. It is not a substitute for a manufacturing engineer's judgment, but it catches rule-based geometry problems before the part reaches a shop or consultant.

Do I need CAD software installed to run a manufacturability check?

No. Tools like Fabdose analyze STEP files directly without requiring SolidWorks, Fusion 360, or another CAD application. You export a STEP file from whatever CAD tool you use and load it into the analysis app. The analysis does not depend on the originating CAD environment.

Check your STEP file on your own computer, before you send it anywhere. Five free analyses, no credit card.
Download Fabdose Free

Originally published at fabdose.app/blog/where-to-check-if-your-design-can-be-made.

Injection Molding vs CNC vs 3D Printing: Which Process for Your Part?

Fabdose — Tue, 08 Sep 2026 06:03:50 +0000

The most common process selection mistake product designers make is choosing based on what they already know. If you have always used CNC, you quote the part in metal. If the previous product used injection molding, the new one gets the same process. The result is a part that works but costs three times what it should, or one that hits a geometry wall late in development.

Here is the practical answer: no single process wins. The right choice depends on four variables — how many units you need, what geometry your design actually contains, what material properties the part must have, and how much tooling cost you can absorb up front. This guide walks through each major process on those axes so you can make the call early, when it still affects the design rather than just the budget.

What is the difference between injection molding, CNC machining, and 3D printing?

Injection molding, CNC machining, and 3D printing each produce physical parts through fundamentally different mechanisms. Injection molding forces molten material into a mold — high volume, low per-part cost, strict geometry rules. CNC machining removes material from a solid block — flexible geometry, high precision, no tooling cost, higher per-part cost. 3D printing builds parts layer by layer — no tooling, almost no geometry constraints, weaker material properties at volume. The right choice depends on production volume, geometry, material, and tooling budget.

The four processes at a glance

Injection molding forces molten plastic into a steel or aluminum tool under high pressure. It produces high volumes at very low per-part cost, but the tooling investment is significant and the geometry must conform to strict rules: uniform wall thickness, adequate draft on all pulled faces, and no features that trap the mold as it opens.

CNC machining removes material from a solid block using rotating cutters. It reaches tight tolerances on almost any geometry a cutter can physically reach, handles metals and engineering plastics equally well, and requires no tooling investment. The cost per part is higher and scales linearly with complexity. Material waste can be substantial.

3D printing (including FDM, SLA, and SLS) builds parts layer by layer with no tooling cost and almost no geometry constraints. It excels at low volumes and complex internal or external geometry. The tradeoffs are weaker and more anisotropic material properties versus the same material machined or molded, coarser surfaces in most cases, and per-part costs that do not drop significantly at volume.

Die casting is the metal equivalent of injection molding: molten aluminum or zinc is forced into a steel die at high pressure. It produces strong, dimensionally consistent metal parts at high volume. Tooling cost is even higher than injection molding, and the geometry rules are similar — draft, uniform wall, controlled undercuts.

The comparison table

	Injection Molding	CNC Machining	3D Printing (SLA/SLS/FDM)	Die Casting
Volume sweet spot	10,000+ units	1 to ~5,000 units	1 to ~500 units	10,000+ units
Tooling cost	$3,000 to $80,000+	None	None	$10,000 to $150,000+
Per-part cost at volume	Very low ($0.50 to $5)	Medium to high	Medium ($5 to $100+)	Very low (similar to IM)
Tolerance	+/- 0.1 to 0.25 mm typical	+/- 0.01 to 0.05 mm	+/- 0.1 to 0.5 mm typical	+/- 0.1 to 0.2 mm typical
Surface finish	Excellent (tool finish)	Excellent (machined)	Varies — SLA best, FDM visible layers	Excellent with secondary ops
Primary materials	Engineering thermoplastics	Metals, plastics, composites	Resins, nylons, photopolymers	Aluminum, zinc, magnesium
Geometry freedom	Constrained (draft, wall, undercuts)	High (tool access limits)	Very high	Constrained (similar to IM)
Lead time (first part)	4 to 12 weeks (tool build)	Days to weeks	Hours to days	6 to 16 weeks (die build)

Where each process breaks down

Injection molding geometry rules

Injection molded parts must satisfy three geometric conditions that are non-negotiable if you want consistent, defect-free parts.

Draft angles — every face that is parallel or close-parallel to the mold opening direction needs at least 1 to 3 degrees of taper. Without draft, the part grips the tool on ejection, causing surface damage and ejection failures. Textured surfaces need more draft, often 3 to 5 degrees, because the texture mechanically interlocks with the tool steel.

Uniform wall thickness — significant thickness variation causes differential cooling rates, which creates sink marks on visible surfaces, warpage of the overall part, and internal voids. The standard target is walls within 25% of each other. Ribbing instead of thick sections is the typical fix.

Undercuts — any feature that prevents the tool from opening in a straight pull direction requires a side-action (a moving tool element that slides out before the mold opens), which adds significant tooling cost and complexity. Internal undercuts are even harder to address.

CNC geometry rules

CNC can reach almost any geometry, but access matters. Deep pockets with small radii require long, thin tools that deflect and break. A 10 mm deep pocket with a 0.5 mm corner radius is technically possible on a high-end machine with special tooling, but it is expensive and slow. The practical rule is to keep floor radii to at least 1 mm and avoid aspect ratios (depth-to-width) above 4:1 in pockets unless you have verified your shop can hold it.

Five-axis machining expands access significantly, but also expands cost. If your part needs undercutting or compound curves, CNC can often handle it — at a price.

CNC also wastes material. A complex aluminum part machined from billet can produce 80% or more scrap by weight. For expensive materials this matters.

3D printing limitations

The layer-by-layer build process introduces anisotropy: printed parts are weaker in the Z-axis (build direction) than in the XY plane. For structural parts under load this matters — the part may pass a static strength test in one orientation and fail in service under a different load direction.

SLA (resin) produces smooth surfaces and fine features but parts are brittle and UV-sensitive unless post-cured and coated. SLS (nylon) produces stronger, more isotropic parts and handles complex geometry including internal channels. FDM is cheapest but leaves visible layer lines and has the most pronounced anisotropy.

At low volume — prototypes, jigs, custom housings, one-off functional parts — these limitations are acceptable. At production volume they are not.

Insight: The geometry rules for injection molding and die casting are almost identical in concept: both need draft, both punish undercuts, both reward uniform wall thickness. The difference is material — plastic vs. metal — and the cost and complexity of the tooling.

How to decide in practice

Cost per part by production volume: 3D printing for low volume, CNC in the middle, injection molding at scale.

Start with volume. If you need fewer than 500 units, the tooling cost of injection molding almost never amortizes. CNC or 3D printing is the answer for prototypes and initial production. If you need more than 10,000 units of a plastic part, injection molding's per-part cost advantage eventually outweighs the tooling investment — usually somewhere between 2,000 and 5,000 units depending on part complexity and material.

Then check your geometry. If your design has features that violate draft or uniform-wall rules for injection molding, you have three choices: redesign the part, switch to CNC, or accept side-actions in the tool. The choice depends on whether those features are essential to the product function.

Material is often the deciding factor for metal parts. If you need structural aluminum or zinc, you are choosing between CNC (low volume, high tolerance) and die casting (high volume, moderate tolerance). For plastic structural parts, glass-filled nylons and engineering resins available in injection molding generally outperform what is achievable with 3D printing.

Tip: Lead time cuts both ways. CNC and 3D printing get you first parts in days. Injection molding and die casting require 4 to 16 weeks of tool build before you see a single part. If your design is still changing, starting a tool build is expensive timing risk.

The geometry check you can run before committing

The geometry rules above — draft, wall thickness, undercuts — are checkable before you send drawings to a shop. You do not need a manufacturing engineer to flag most of them. You need software that can open your STEP file and measure those conditions directly from the geometry.

The same STEP file can be checked against injection molding, CNC, die casting, and 3D printing constraints to see which processes your geometry actually fits today. Checking one process per analysis is part of the free tier; the side-by-side auto-comparison across processes at once is a Pro feature. Either way, you are not guessing which rules you might be violating, you are seeing the specific faces and features that would cause problems.

Fabdose is a local-first desktop app for Mac that does this. You load a STEP or STP file, the analysis runs in one to three minutes on your computer, and you get a process-by-process report: draft violations and undercut locations pinned face by face in a 3D viewer, a part-wide wall-thickness read (a variation-based warp-risk flag, not a per-face pin), and a cost-range estimate per process. Your CAD files are never uploaded — they stay on your computer, and only the defect findings are processed by AI and synced to your account. The first five analyses are free, no credit card required.

It does not replace a manufacturing engineer and it does not generate supplier quotes. What it does is tell you, before you talk to anyone, which of these four processes your current geometry is compatible with.

Frequently asked questions

When does injection molding become cheaper than CNC per part?

Injection molding becomes cheaper than CNC on a per-part basis somewhere between 2,000 and 5,000 units for most plastic parts, depending on part complexity and material. Below that range, the tooling cost — typically $3,000 to $80,000 for a mold — does not amortize. Above it, injection molding's per-part cost of $0.50 to $5 makes CNC's linear per-part cost uncompetitive.

What geometry issues disqualify a design from injection molding?

Three issues are most common: insufficient draft angle (faces parallel to the mold opening direction need at least 1 to 3 degrees of taper), significant wall thickness variation (walls should be within 25% of each other to prevent sink marks and warpage), and undercuts (features that block the mold from opening in a straight pull). Each of these can be addressed through design changes or tooling additions, but they add cost and complexity if not caught early.

Can the same STEP file be used for both CNC and injection molding quotes?

Yes. A STEP file is a neutral geometry exchange format that most manufacturing services accept. The same file can be sent to a CNC shop and an injection molding supplier. The manufacturability constraints differ by process, so the feedback will differ — which is also why checking a STEP file against multiple process rules before quoting is useful.

How does die casting compare to injection molding?

Die casting uses the same general approach — forcing molten material into a die under pressure — but with metals (aluminum, zinc, magnesium) instead of plastics. The geometry rules are similar: draft angles, uniform wall thickness, and controlled undercuts apply to both. Tooling cost for die casting is higher than injection molding, typically $10,000 to $150,000 or more. Die casting is the right choice when you need metal parts at high volume with good dimensional consistency.

Is 3D printing suitable for production parts?

3D printing is suitable for production parts at low volume, typically up to a few hundred units, for parts where the structural limitations are acceptable. SLS (nylon) produces stronger, more isotropic parts than FDM and is used in some production applications. At higher volumes, the per-part cost does not drop significantly and material properties remain weaker than equivalent injection-molded or machined parts. The primary production use cases are custom, complex, or low-volume parts where tooling cost cannot be justified.

Analyze your STEP file on your computer. Your CAD file is never uploaded.
Download Fabdose Free

Originally published at fabdose.app/blog/injection-molding-vs-cnc-vs-3d-printing.

How to Know If Your CAD Design Can Actually Be Manufactured

Fabdose — Tue, 08 Sep 2026 06:01:47 +0000

Sending a CAD file to a factory without knowing whether it can be manufactured is one of the most expensive mistakes in product development. The quote comes back with a list of issues, or it does not come back at all. Either way, you are now weeks behind and facing a redesign with tight timeline pressure.

The question every product designer should be able to answer before sending a file is: can this geometry actually be made with the process I have in mind, at the tolerance I need, without the factory having to rework or reject it?

That answer is not always obvious, and the ways most designers check today all have real gaps.

What does "manufacturable" mean for a CAD design?

Manufacturability is a CAD design's compatibility with a specific manufacturing process — meaning its geometry satisfies the dimensional and shape constraints that process requires. A part is manufacturable when it can be produced at the intended volume and tolerance without redesign, tooling rework, or excessive manual intervention. The same geometry may be manufacturable by CNC machining but not injection molding, because each process imposes a different set of constraints.

How to check if your design is manufacturable: step-by-step

Checking your design for manufacturability before sending it to a shop takes four steps.

Choose the target process. Manufacturability is always relative to a process — injection molding, CNC, die casting, or 3D printing. Each has different wall thickness minimums, draft requirements, and geometric constraints. Decide which process you are targeting before checking.
Run a geometry analysis on the STEP file. Load your STEP or STP file into a tool that can open the file and compute geometric properties — not a language model that reads the file as text, but software with a geometry kernel that can measure wall thickness, face normals, and surface topology.
Review the flagged issues by severity. A useful analysis ranks findings: which issues will cause manufacturing failure, which will affect cost or quality, and which are minor. Address the high-severity items first.
Iterate and re-run before sending for quotes. The goal is to arrive at the factory or the DFM consultant with a part that has already been through a first filter — so the external feedback focuses on judgment-level decisions, not geometry basics.

A multi-process manufacturability report: each process scored, with issue counts and a cost signal.

What "manufacturable" actually means for a physical part

Manufacturability is not a single property. It is a combination of geometry properties that interact with a specific manufacturing process. A part that machines perfectly in CNC may be completely impossible to injection-mold. A part designed for die casting may have features that require expensive hand-finishing if made by SLA printing.

The key geometric properties that determine whether a part is manufacturable depend on the process, but the common ones are:

Wall thickness. Every process has a minimum wall thickness below which material either will not flow (injection molding, die casting), will crack under machining forces (CNC), or will be too fragile to handle (SLA). For injection molding in standard ABS, the practical minimum is around 1.2 mm. Walls below that threshold risk incomplete fill, sink marks, or warpage.

Draft angles. Parts that are pulled from a mold -- injection molding and die casting especially -- need tapered faces so the part can release without tearing. The standard rule is 1 to 3 degrees of draft per side. Vertical faces with zero draft will stick to the mold and either tear the part or damage tooling.

Undercuts. An undercut is any feature that is trapped by the mold geometry in the pull direction. A snap-fit tab on the inside of a housing, a side hole perpendicular to the pull direction, a recessed groove -- these all require side-actions or collapsible cores, which add tooling cost and complexity. If you have undercuts you did not design intentionally, you need to know about them early.

Tolerances relative to the process. CNC machining can hold tolerances down to a few microns on a good day. Injection molding typically holds +/- 0.1 to 0.3 mm depending on part size and material. If your design has a tight-tolerance fit that only works within +/- 0.05 mm, and you are quoting injection molding, you have a mismatch that will only surface when you receive parts.

Process fit in general. Beyond the individual checks, the overall geometry needs to match what the process is good at. Deep pockets with very high aspect ratios are expensive or impossible in some CNC setups. Thin features that extend far from the base are risky in injection molding due to differential cooling. Long, thin pins or shafts in die casting may not fill completely.

Insight: The most common source of late-stage redesign is not a single catastrophic issue -- it is a set of smaller geometry mismatches that each individually seem minor, but together make the part difficult to quote, expensive to tool, or unreliable in production.

How designers check today, and why each has a gap

Ask the factory

The obvious approach: send the file, ask what they think, get a quote with feedback.

The problem is that most factories will not do free design feedback before a quote, and even when they do, their incentive is to get your business, not to be thorough about problems that would push work back to you. A factory quoting on a part with manageable issues may simply price in the extra work and ship you parts that have consistent sink marks.

If you send the same part to three factories, you may get three different feedback sheets, none of them complete. And you have now spent two to three weeks waiting.

Hire a DFM consultant or have an engineer review it

Accurate. A good manufacturing engineer who looks at your STEP file will catch real issues and explain the root cause. They can tell you which process fits the part and why, and give you concrete changes to make.

The gap here is speed and cost. A proper DFM review from a consultant or contract engineer takes days to schedule, costs several hundred to a few thousand dollars depending on part complexity, and is hard to justify early in an iteration cycle when the design is still changing.

You end up either skipping it to save budget, or doing it once at the end when it is most expensive to act on the findings.

Ask ChatGPT or Claude

AI assistants know DFM rules. Ask about draft angles and you will get a reasonable answer covering typical ranges for injection molding. Ask about minimum wall thickness for CNC in aluminum and you will get a defensible rule of thumb.

The gap is that AI language models cannot see your actual geometry. When you paste a STEP file into ChatGPT or Claude, what it receives is text -- AP203 or AP214 entity definitions. The model has no way to reconstruct the 3D shape from those definitions, estimate wall thickness from the part's shell geometry, detect which faces have insufficient draft, or identify undercut features relative to a pull direction.

You get a response that sounds DFM-aware but is not grounded in your actual part. You might get a response that says "looks reasonable" when your part has six locations with 0.4 mm walls. The model cannot measure anything from the file because it never rebuilt the geometry.

This matters most when you have already done the design work and need a specific answer about this part, not a general education about DFM rules.

Warning: Using an AI chat tool to review your STEP file for manufacturability feels like a check, but it is not. A confident-sounding response about a file the model cannot parse is false assurance at the worst time in your cycle.

What a fast first-pass geometry check looks like

A useful first-pass check is one that runs on your actual geometry, covers the key failure modes for the processes you are considering, and gives you specific findings rather than general advice.

That means software that loads the STEP file, builds a geometric representation of the part, and then computes: where the wall thickness falls below the process threshold, which faces lack adequate draft relative to a pull direction, where undercuts are located, and how the overall shape fits the process constraints. The output should be specific -- per-face, per-feature, with severity ranked so you know what to fix first.

A fast first-pass geometry check runs on your computer in 1-3 minutes.

This kind of check does not replace a manufacturing engineer. It is a self-check step you can run before the part leaves your desk. The goal is to find the obvious geometry issues early, while changing them is cheap and does not require coordination with anyone.

Where it fits in the workflow: after you finish a design iteration and before you send it for quoting, DFM review, or internal sign-off. Not as a gate, but as a sanity check that catches the issues you did not catch during modeling.

A note on cost estimates by process

One thing that is useful alongside a geometry check is a rough cost signal by process. If your part has geometry that works for both CNC and injection molding, knowing that the tooling amortized cost for injection molding at your expected volume is roughly 40 percent lower than CNC unit cost changes which direction you take the design.

This is not a supplier quote. It is an order-of-magnitude signal that helps you make better design decisions earlier, before you have a tooling vendor or a volume commitment. The accuracy depends on part complexity and region, but even a rough range is more useful than no signal at all.

Trying a first-pass self-check

Fabdose is a local-first DFM analysis app for product designers on Mac. You give it a STEP or STP file and in roughly one to three minutes it runs geometry checks for the process you pick -- injection molding, CNC, die casting, or SLA/3D printing. It flags draft issues, undercuts, and process-fit problems -- severity ranked, with a 3D viewer that pins those issues on the actual geometry. Wall thickness gets a part-wide read too: a variation-based warp-risk flag, not a face-by-face pin -- it's a nominal estimate, not a full shell analysis. It also gives a cost-range estimate per process as a rough planning signal. The free tier covers one process per analysis; the auto-comparison that runs your part across all four processes at once is a Pro feature.

Your CAD files are never uploaded — the file stays on your computer throughout. The defect findings are processed by AI and synced to your account.

It is not a replacement for a manufacturing engineer or factory DFM review. It is the step before those -- a way to find the obvious geometry issues yourself, fix them, and arrive at the factory or the engineer with a cleaner part.

The free tier includes five analyses, with an account required (Google or email) and no credit card needed.

Frequently asked questions

What are the most common reasons a CAD design fails manufacturability review?

The most common failures are insufficient draft angles on molded parts (faces parallel to the pull direction that will tear on ejection), wall sections below the minimum thickness for the process (typically under 1.2 mm for injection molding in ABS), and undercuts that require additional tooling to address. These three issues appear in the majority of first-pass DFM rejections because they are easy to miss during the design phase.

Can I check manufacturability without sending my file to a factory or consultant?

Yes. Geometry-analysis software can check your STEP file against process constraints locally — measuring wall thickness, detecting undercuts, evaluating draft angles — without requiring a human reviewer or factory quote. This type of first-pass check is not a substitute for expert review, but it catches rule-based geometry issues you can fix before involving anyone else.

How long does a first-pass manufacturability check take?

A local geometry analysis of a typical STEP file runs in one to three minutes on a modern laptop. A factory DFM review attached to a quote request takes two to five business days. A DFM consultant engagement takes days to schedule and complete. The difference in timing is why a first-pass check is worth running on every design iteration.

What file format do I need for a manufacturability check?

STEP (Standard for the Exchange of Product model data) files — with extensions .step or .stp — are the standard neutral format that geometry-analysis tools work with. Most CAD applications can export to STEP directly. SolidWorks, Fusion 360, Rhino, and Onshape all support STEP export.

Does a manufacturability check replace a manufacturing engineer?

No. A geometry-based check flags rule violations — wall thickness under threshold, draft angle insufficient, undercut present. It does not apply material-specific judgment, regional tooling knowledge, or production volume context the way an experienced manufacturing engineer does. The right model is: run the geometry check yourself first, then bring a cleaner part to the engineer so their time focuses on decisions that require judgment.

Check your STEP file on your Mac before it reaches the factory. Five analyses free, no credit card.
Download Fabdose Free

Originally published at fabdose.app/blog/how-to-check-if-your-design-is-manufacturable.

Can ChatGPT Analyze Your STEP File for Manufacturability?

Fabdose — Tue, 08 Sep 2026 05:59:37 +0000

Product designers working on physical parts face a question that comes up early in every review cycle: is this geometry manufacturable? And more and more, the first place people turn is ChatGPT.

The honest answer is that ChatGPT is genuinely useful for DFM education and can help you think through design rules. But it cannot analyze your actual STEP file. Understanding why that distinction matters will save you from getting false confidence at the worst possible time.

What is a STEP file analysis for manufacturability?

A STEP file analysis for manufacturability is a geometric computation that reads a CAD file's B-rep topology, measures dimensional properties — wall thickness, face normals, fillet radii — and checks those values against process-specific constraints. It requires a geometry kernel that can reconstruct and traverse the actual 3D shape, not a text-processing model.

What ChatGPT can do

If you describe a part feature in text, ChatGPT will give you solid DFM feedback. Ask about minimum wall thickness for CNC milling, and it will cite reasonable rules of thumb (typically 0.8 mm for aluminum, 1.5 mm for plastics). Ask about undercuts and it will explain why they require side-actions or EDM. Ask about draft angles and it will give you the 1-3 degree range that most injection molding processes expect.

This is genuinely valuable. ChatGPT has absorbed a large body of manufacturing knowledge from technical documents, engineering forums, and textbooks. For someone learning DFM or needing a quick sanity check on a design rule, it is a useful starting point.

You can also upload an image of your part, and ChatGPT will visually describe what it sees. If you share a rendering or a photo, it can sometimes identify obvious features and comment on them.

What ChatGPT cannot do

ChatGPT cannot open a STEP file and analyze the actual 3D geometry.

This is not a product limitation that might change in a future release. It is structural. ChatGPT is a language model. It processes text and images. A STEP file is a geometric database containing precise coordinate data, B-rep topology, surface normals, and tolerance information. Analyzing it for manufacturability requires a geometry kernel that can traverse that topology, compute measurements, and evaluate geometric conditions against process constraints.

When you paste a STEP file's text content into ChatGPT, you are feeding it AP203 or AP214 entity definitions. ChatGPT will try to interpret the text, but it has no ability to reconstruct the 3D geometry from those definitions, estimate wall thickness from the part's shell, detect undercuts relative to a pull direction, or measure fillet radii from the parametric surface data.

What you get back is a plausible-sounding response based on the text patterns in the file, not a geometric analysis.

The gap in practice

Here is what the difference looks like on a real part.

Suppose you have a CNC-machined aluminum bracket with a pocket that has a small flat floor radius. In text, you might describe it as "a pocket with a 0.5 mm corner radius at the bottom." ChatGPT will correctly flag this as a potential issue because standard end mills have a minimum radius floor clearance.

But if you have that bracket as a STEP file and want to know whether any floor radius falls under 0.8 mm, you need geometric computation. You need software that can read the B-rep data, measure the actual fillet radii, compare them against a threshold, and flag the part with a visual marker on the geometry.

ChatGPT cannot do any of that. It cannot measure anything from the file geometry because it never reconstructed the geometry in the first place.

The same gap applies to:

Wall thickness analysis: Estimating the minimum wall thickness of a part requires geometric computation over the actual mesh or B-rep. ChatGPT cannot compute this.
Draft angle detection: Determining whether faces have adequate draft for injection molding requires computing face normals relative to a pull direction from the actual surface data.
Undercut detection: Identifying features that require side-actions requires analyzing the part topology relative to a mold parting direction.
Thin feature detection: Finding ribs or bosses that are structurally undersized requires actual dimensional measurement from the model.

Why this matters at review time

The problem with using ChatGPT for STEP file review is not that it gives bad advice in general. The problem is that it gives you confidence without geometric grounding.

You submit a part to a shop and get a quote. The shop comes back with a list of issues: undersized fillets in four locations, insufficient draft on two faces, a wall section that is 0.6 mm and likely to warp on demold. These are specific geometric findings. You need to trace them back to your model and decide what to change.

If you had asked ChatGPT before submitting, it might have given you a response that sounded thorough. It might have mentioned draft angles and wall thickness as things to watch. But it would not have told you about those four specific fillet locations, because it never looked at the fillet locations. It looked at the text of a STEP file and gave you a general response.

The difference between a general DFM discussion and a geometric analysis of your actual part is the difference between studying for an exam and taking it.

Insight: A quote rejection or an ECO late in development costs real time and money. The value of early DFM analysis is catching geometry issues before they reach the shop, not after.

What a local geometry analysis does differently

Fabdose runs analysis directly on the STEP file on your computer. Your CAD files are never uploaded; only the defect findings are processed by AI and synced to your account.

The analysis engine loads the geometry, builds a representation of the part topology, and runs checks against configurable process constraints. Wall thickness is estimated from the part's shell geometry, and the part is flagged when its minimum wall falls below the process threshold. Draft angle is computed per face against a specified pull direction. Undercuts are detected by analyzing which faces would be trapped by the mold geometry.

The output is specific: locations, measurements, and which check triggered the flag. Not general advice about what to watch out for, but a report of what your actual part has.

This is the distinction that matters for a product designer on a real schedule. You are not looking for an education in DFM principles. You are looking for a fast answer to a specific question: does this version of the part have any geometry that will cause a manufacturing problem?

Tip: Your CAD files are never uploaded — your STEP and STP files stay on your computer throughout the analysis. The defect findings are processed by AI and synced to your account.

The honest comparison

What you want to know	ChatGPT	Fabdose
What are typical draft angle requirements?	Yes	Reference docs
Is my minimum wall thickness within the threshold?	No - cannot measure from STEP	Estimated from shell geometry, flagged if below
Which specific faces have insufficient draft?	No - no geometry access	Yes - per-face report
Where are my undercuts relative to pull direction?	No - cannot analyze topology	Yes - with pull direction config
What DFM rules apply to injection molding?	Yes	Yes
Can I keep my STEP file on my computer?	Upload not parsed anyway	Yes - local-first

ChatGPT is a good starting point for learning DFM. It is not a substitute for actual geometric analysis when you need to know whether your specific part is ready to send to a shop.

Frequently asked questions

Can ChatGPT read or open a STEP file?

No. ChatGPT is a language model that processes text and images. A STEP file is a geometric database containing precise coordinate data, B-rep topology, and surface normals. When you paste a STEP file's text content into ChatGPT, it reads the AP203 or AP214 entity definitions as raw text and cannot reconstruct the 3D shape from those definitions.

Can I upload a STEP file to ChatGPT's file upload feature?

Yes, you can upload the file. But ChatGPT still cannot perform geometric computations on it. It can acknowledge the file's content as text and offer general observations about entity types, but it cannot measure wall thickness, detect undercuts relative to a pull direction, or evaluate draft angles from the actual surface data.

What is the difference between DFM advice and a DFM analysis?

DFM advice provides general manufacturing rules and heuristics — for example, that injection-molded parts typically need 1 to 3 degrees of draft. A DFM analysis applies those rules to the specific geometry of your part, flagging which faces fail the draft check and at what angle. ChatGPT provides advice. Geometry-analysis software provides analysis.

Is there a desktop tool that analyzes STEP files for manufacturability locally?

Yes. Fabdose is a local-first desktop app that loads STEP and STP files directly on your computer. Your files are never uploaded — only the defect findings are synced. The analysis runs in one to three minutes and produces a report with per-face findings, severity ranking, and a 3D viewer with issue pins.

How accurate is a geometry-based DFM analysis compared to factory feedback?

A geometry-based first-pass analysis catches rule-based issues — draft violations, wall thickness under threshold, undercut locations — reliably. It does not replace a manufacturing engineer's judgment on material-specific edge cases, regional tooling constraints, or production volume tradeoffs. The goal is to find the obvious geometry problems yourself before the part reaches a shop or consultant.

Analyze your STEP file on your computer. Your CAD file is never uploaded.
Download Fabdose Free

Originally published at fabdose.app/blog/can-chatgpt-analyze-step-file.

Where Is the Thick Section on This Part?

Fabdose — Mon, 07 Sep 2026 13:49:33 +0000

The whole-part wall thickness number for the part below reads 12.157mm.

There is no 12.157mm wall on this part. There is no 12mm anything on this part. It is a 2mm shell.

That number is what a single figure for a whole part gets you when the thing you are actually looking for is local. The thick section on this part is real, it is 5.40mm, and it lives in one specific place. Here is the check naming it:

Wall thickness 5.40mm at face 0 exceeds the 3.56mm maximum for ABS
Wall thickness 5.40mm at face 28 exceeds the 3.56mm maximum for ABS

Face 0 is the 4900 square mm outer bottom, the flat surface a person actually looks at. Face 28 is the 576 square mm top of the coil plinth at (0.00, 0.00, 5.40). They are the two sides of the same wall.

Where is the thick section on this part?

Under the feature you added on purpose. A coil pillar, a magnet pocket, a boss pad, a raised logo. It stacks on top of the nominal wall, the outer surface stays perfectly flat, and nothing in your CAD viewport tells you the section behind it just went to two and a half times nominal.

The part below is a synthetic, representative model built for this post, not a real customer file. It is the lower housing of an electric toothbrush charging base, 70 by 70 by 18mm in ABS, modelled as an open-top box: 2.0mm floor, 2.0mm side walls drafted 2 degrees per side, two screw bosses with blind cored holes, a recessed label pocket in one side wall, and an inductive coil plinth 24 by 24mm and 3.4mm tall standing on the inner floor. Thirty-three faces. Every run in this post used the same command:

python cli.py analyze <part>.step ABS --process injection_molding --wall-thickness 2.0

The floor is 2.0mm. The plinth adds 3.4mm on top of it. Nobody drew a 5.4mm wall. It is the sum of two decisions that were each individually fine.

What does the report actually point at?

Two faces, both at 5.40mm, both over the 3.56mm ABS maximum. The reason the same wall shows up twice is worth sitting with, because it is the clearest picture of what a thick section is.

Face	What it is	Area	Position	Measured
0	outer bottom, the visible flat face	4900.0 square mm	(0.00, 0.00, 0.00)	5.40mm
28	top of the coil plinth, inside	576.0 square mm	(0.00, 0.00, 5.40)	5.40mm

The check measures thickness by casting a ray inward from a face and finding what is on the other side. From face 0 it goes up and hits the plinth top 5.40mm later. From face 28 it goes down and hits the outer bottom 5.40mm later. Same wall, read from both ends, and both ends flagged.

That second row is the one that matters for what happens on the shop floor. The feature you added is on the inside. The defect it produces will appear on the outside, on face 0, the cosmetic face nobody was worried about.

Why does a thick section cause a sink mark?

Because the middle of it is still liquid when the skin has already frozen. As the core cools it shrinks, and there is nothing to feed it, so it pulls the nearest surface inward. Thicker section, longer the core stays hot, deeper the dimple. Sink marks are one of the defects that do not show up in your CAD model at all, which is the reason a geometry check has to reason about them from the section that causes them rather than from anything visible in the viewport.

The report does not leave that as an abstraction. The sink mark risk in this run carries the root cause and the location:

defect_type: sink_mark
severity: high
message: "High risk of sink marks due to 2 thickness violations"
root_causes: ["wall_too_thick", "wall_too_thick"]
affected_areas: [0, 28]

Those are the same two face IDs. The risk is not floating free of the geometry, it is attached to the two faces that caused it.

For context on the numbers: fabdose's own material limits put the workable ABS range at roughly 1.14 to 3.56mm. The 5.40mm spot here is 170 percent over this model's own dominant wall of 2.00mm, which is what the uniformity finding reports separately.

What does the whole-part average tell you?

Very little, and this part is a clean demonstration of why.

Figure	Value	Is it a wall on this part?
whole-part average	12.157mm	no
whole-part maximum	12.157mm	no
whole-part minimum	4.717mm	no
dominant wall, per-face	2.00mm	yes
thick spot, per-face	5.40mm	yes, faces 0 and 28
thin spot, per-face	0.77mm	yes, face 14

The first three come from whole-part measures that a bounding geometry can produce without ever looking at a wall. The last three come from measuring each face against whatever is opposite it. Only the second set has anything you can act on, because only the second set comes with a face ID and a coordinate you can open in your CAD tool.

This is also why the part has a thin problem you would never have found by averaging. Face 14 is the wall behind the recessed label pocket, and it measures 0.77mm against the 1.14mm ABS minimum:

Wall thickness 0.77mm at face 14 is below the 1.14mm minimum for ABS

A 0.95mm pocket cut into a wall that was already thinning as it drafted upward. Another sum of two reasonable decisions.

Does switching material fix it?

It is the first thing people try, and on this part it does something quietly misleading.

Running the identical geometry as PP instead of ABS:

Finding	ABS	PP
face 0, 5.40mm	over the 3.56mm max	over the 3.81mm max
face 28, 5.40mm	over the 3.56mm max	over the 3.81mm max
face 14, 0.77mm	under the 1.14mm min	clears the 0.64mm min, no finding
high-severity count	5	4

The report got shorter. The thick section did not move. What disappeared was the thin wall, because PP tolerates a thinner section than ABS does, and if you were skimming counts rather than reading faces you would have recorded that as progress.

What happens if you core it out?

This is the fix a molder would ask for, so it is worth running rather than asserting. The plinth is cored from the open side of the part, which keeps it mouldable, so the local section returns to the 2.0mm nominal.

	as designed	cored out
`wall_too_thick` findings	2 (faces 0 and 28)	0
sink mark risks	2, top one high severity	0
warpage risk	high	low
total defect risks	10	7
high-severity issues	5	2
production readiness	difficult, Design Revision Required	concerns, Optimization Recommended
tooling estimate	$14,000	$14,000

Both thickness findings are gone. Both sink mark risks are gone, not reduced. The warpage risk drops from high to low, and the cycle time note that read Cycle time >373s: solid mass without core-out is no longer in the readiness details.

The tooling estimate did not move. It was $14,000 before and $14,000 after. Coring a section changes the defect picture and the cycle time, not the price of cutting the tool, and it would be easy and wrong to present this as a saving. The saving, if there is one, is in the parts that come out of the tool, and that is not something this estimate speaks to.

The thin wall at face 14 is still there in the cored run, correctly, because coring the plinth has nothing to do with the label pocket. Fixing one thing did not silently launder the other.

What could the check not measure?

This is the part most tools do not tell you, so it is worth stating plainly. Of the 33 faces on this part, 15 were measured. The other 18 carry no thickness verdict at all, and each says why:

Reason	Faces	What it means
measured	15	a ray from this face found an opposing surface
`slender_face`	16	too narrow to read as a wall, which is the gate that stops rib and rim edges being reported as thin walls
`no_opposing_face`	2	the ray found nothing on the other side

Those 18 faces are not silently passing. They are unreported, and the difference matters when you are deciding how much of your part this check actually covered. On this part that is 45 percent of faces measured, which is typical for a shape with drafted walls and small edge faces, and it is the number the check hands you rather than one you have to infer.

What doesn't this check do?

It is a geometry check, so it locates what geometry determines. Wall thickness, draft and undercuts are measured per face and come back with face IDs and coordinates. Weld lines, air traps, jetting and burn marks depend on gate placement, melt flow and cooling, and they come back as risk flags without a per-face location, because a geometry check cannot honestly place them. Those belong to a molder's DFM review or a flow simulation.

One correction to an earlier post in this series, since it is now out of date. The orientation teardown said sink marks arrive without a structured face reference. For the thickness-driven case that is no longer true: the sink mark risk in the run above lists faces 0 and 28. The rest of that sentence still holds for weld lines, air traps and jetting, which remain unlocated.

Wall thickness is measured per face for injection molding. Other processes still use a whole-part figure, so if you are running CNC or sheet metal, the per-face table described here is not what you get.

What is the takeaway?

The thick section on a part is almost never a wall somebody drew. It is a local feature landing on top of a nominal wall, invisible from outside, and it announces itself later as a dimple on the flat cosmetic face across from it. A single average for the whole part cannot find it. On this part the average was 12.157mm, a number matching nothing, while the thing that mattered was 5.40mm at two named faces with coordinates.

Fabdose reads your STEP or STP file on your own computer, on Windows or Mac, and measures wall thickness face by face against the material's limits, alongside draft and undercuts, before the file goes to anyone who quotes it. Your CAD geometry is processed on your computer; the defect findings and their written explanations are processed by AI. It covers what geometry determines, and it tells you which faces it could not measure instead of passing them silently.

If you have a housing with a screw boss, a magnet pocket or a coil pillar on the inside of a cosmetic face, run it once and look at what sits opposite that feature. That is where the dimple goes.

Check your design with Fabdose

FAQ

Where is the thick section on my injection molded part?

Under something you added on purpose: a coil pillar, a magnet pocket, a boss pad, a raised logo. It stacks on the nominal wall and stays invisible from outside. On the charging base here, a 3.4mm plinth on a 2.0mm floor makes 5.40mm, reported from both sides as face 0, the 4900 square mm outer bottom, and face 28, the 576 square mm plinth top at (0.00, 0.00, 5.40). The ABS maximum is 3.56mm.

Is average wall thickness a useful number for DFM?

Not on a part with local features. Here the whole-part figures read 12.157mm average, 12.157mm maximum, 4.717mm minimum, and none of the three is any wall on the part. The real walls are a 2.00mm dominant section, a 5.40mm thick spot and a 0.77mm thin spot.

What causes sink marks in injection molding?

A section thick enough that its core cools and shrinks after the skin has frozen, pulling the surface inward. That is why the dimple lands on the visible face opposite the thick feature. In this run the tool ties them together: sink mark risk high, root cause wall_too_thick, affected areas faces 0 and 28.

Does coring out a thick section actually fix the report?

It fixed this one. Coring the plinth from the open side removed both wall_too_thick findings, removed both sink mark risks, dropped warpage from high to low, and moved production readiness from difficult to concerns. The tooling estimate stayed at $14,000, unchanged.

Does changing material fix a thick section?

Not this one. The same geometry in PP kept both thick findings, since 5.40mm still exceeds PP's 3.81mm maximum. What it removed was a different finding, the 0.77mm thin spot, which clears PP's 0.64mm minimum. The report got shorter without the thick section changing.

Why does the report say it could not measure some faces?

Because it would rather say so than guess. Of 33 faces, 15 were measured, 16 skipped as slender_face, the gate that keeps rib and rim edges from being read as walls, and 2 skipped as no_opposing_face. Those carry no verdict rather than a fabricated one.

Fabdose is a desktop tool for checking STEP and STP files against injection molding design rules on Windows and Mac. It measures wall thickness, draft and undercuts face by face, before the file goes out for a quote. Your CAD geometry is processed on your own computer; defect findings and descriptions are processed by AI. The three runs above are real, unmodified engine runs against a synthetic representative charging base built for this post; the dollar figures are the tool's built-in static estimates, not factory quotes.