Industries / Industrial equipment and services / 3D printing
Designing for printing
Designing for printing is the work of turning an idea into a digital model a printer can build: shaping the part in software, checking it against design rules, saving it in a file format machines can read, and slicing it into the layer-by-layer instructions the printer follows.
7
Milestones achieved, from 1987 to 2018.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1987
Triangles become the shared format
STL describes only the outside surface of a part, as a mesh of flat triangles, and it became widely used for 3D printing. Wikipedia says the Albert Consulting Group developed it in 1987 for the first commercial printers of 3D Systems; the Library of Congress credits Charles Hull of 3D Systems and dates the first documentation, the "StereoLithography Interface Specification," to 1988. The format has no standard way to record color or material, and no units, so the receiving software has to ask whether the numbers are millimeters or inches. An outside count shows what this costs: of 10,000 models featured on Thingiverse between 2009 and 2015, Qingnan Zhou and Alec Jacobson found about 50% were not solid and about 22% were non-manifold (meshes with edges that do not enclose a clean, closed volume).
LessonA format simple enough for every machine to read spread everywhere, and because it checks nothing, broken models pass down to every user. ).
Credit
3D Systems, with the Albert Consulting Group (per Wikipedia) or Charles Hull (per the Library of Congress)
1988
Software decides where material goes
In November 1988 Martin Bendsøe and Noboru Kikuchi published "Generating optimal topologies in structural design using a homogenization method" in the journal Computer Methods in Applied Mechanics and Engineering. The method treats the design of a part as a question of where, inside a block, material should be. This is topology optimization: software that removes material wherever the loads on a part do not need it. Altair built a commercial program on the idea, OptiStruct, which Industry Week magazine named its "Technology of the Year" in 1994; by the recollection of its former simulation manager, General Motors Powertrain bought the first copy. The free-form shapes this software produces are often hard to machine or cast, which is why it became a key part of design for printing.
Credit
Martin Bendsøe and Noboru Kikuchi; Altair
2008
A public library of things
Bre Pettis writes that he and Zach started Thingiverse "on a Saturday in August 2008" to give people a place to share digital designs for things; Wikipedia gives November 2008 and names his partner as Zach "Hoeken" Smith. The two formed MakerBot Industries in January 2009, and Thingiverse ran as its companion site. By MakerBot's count, Thingiverse reached 1 million uploads and 200 million downloads in October 2015. Later libraries followed the same pattern, including Printables from Prusa Research in Prague (renamed from PrusaPrinters on 30 March 2022) and MakerWorld from Bambu Lab in Shenzhen (announced in August 2023).
opendesignnow.orglaopinion.comen.wikipedia.orgblog.prusa3d.comblog.bambulab.com
Credit
Zach Smith and Bre Pettis, MakerBot
2011
Open slicers take root
A slicer cuts a 3D model into thin layers and writes the path the print head follows in each one. Slic3r, which its own site says was "born in 2011 within the RepRap community" of open-source printer builders, is maintained by Alessandro Ranellucci and Joseph Lenox and released under an open license, the GNU AGPL version 3. Printer makers built on it instead of starting over: Prusa Research's PrusaSlicer is "originally based on Slic3r," Bambu Lab's Bambu Studio is "based on PrusaSlicer," and the community-run OrcaSlicer draws on both. Separately, Cura, written by David Braam and also first released in 2011, became the slicer of the Dutch printer maker Ultimaker and, by Ultimaker's account, one of the most widely used open-source slicers; it is released under the LGPL-3.0 license.
LessonBecause the code was open, printer makers in Prague and Shenzhen could ship slicers for their own machines by building on one shared base.
slic3r.orggithub.comgithub.comgithub.comultimaker.comgithub.com
Credit
Alessandro Ranellucci and the RepRap community; Prusa Research; Bambu Lab; David Braam and Ultimaker
2011
A fuller format few used
ASTM approved the Additive Manufacturing File Format (AMF), standard F2915-11, in 2011, and it later became the joint ISO/ASTM 52915 standard. Written in XML (a plain-text tagging language), it adds what STL lacks: units, color, several and blended materials, lattices, and curved triangles. Most of the industry kept using STL. In December 2024, Bob Zollo of Avante Technology said "awareness of the AMF standard and benefits is not as well known," that ISO and ASTM are restricted by their policies from promoting it, and that STL stays in use because of its "ease of adoption, minimal cost."
LessonA better standard does not replace an entrenched one just by being better; someone has to build it into the software people already use.
Credit
ASTM Committee F42 on additive manufacturing
2015
3MF carries the whole job
Microsoft, Autodesk, Dassault Systèmes, FIT AG/netfabb, HP, Shapeways and SLM Solutions launched the 3MF Consortium on May 1, 2015, starting from a file format Microsoft had been developing and donated. A 3MF file packs the model, its units, colors and materials, and extensions such as beam lattices (open frameworks of thin struts) into one zipped package. This time the companies writing the format also built it into their products: by the consortium's count in November 2021, 3MF was "implemented in nearly 50 products across 38 companies," and the Library of Congress keeps a format description of it. When Bambu Lab announced MakerWorld in August 2023, it let designers upload their slicer settings as 3MF files beside a model, so a download carries the print settings as well as the shape.
Credit
the 3MF Consortium, starting from Microsoft's donated draft
2018
Design rules become a standard
ISO and ASTM International published ISO/ASTM 52910:2018, "Additive manufacturing — Design — Requirements, guidelines and recommendations," and European standards bodies adopted it as EN ISO/ASTM 52910:2019. It applies to parts made by any printing process, and it is written for designers and their managers, students of mechanical and computer-aided design, and people who write design guides. By its own scope, it identifies issues to consider but does not give specific design solutions or process- or material-specific data.
LessonA shared list of issues helps teaching and review, but the numbers a designer needs, such as the thinnest wall one machine can print in one material, still have to come from somewhere else.
Credit
ISO and ASTM International
Watch for
Text prompts to editable CAD
Autodesk announced "neural CAD," a family of AI foundation models (large models trained on broad data) that it says can create designs from a text prompt, with "upcoming commercial availability" in its Fusion and Forma software and no release date given.
Achieved whentext-to-geometry ships to Fusion users and someone outside Autodesk publishes parts made from prompts that print and meet their stated dimensions.
Claimed by
Autodesk, at Autodesk University, September 2025
Photo to color print
Meshy says its Meshy-6 image-to-3D engine now runs inside MakerLab, the AI tool section of Bambu Lab's MakerWorld, and turns a single image into a print-ready model. An export costs 2 MakerLab credits and can be a single-color STL or a 3MF file with colors mapped for Bambu Lab's multi-material feeder (AMS).
Achieved whenan independent tester prints a stated sample of generated models and publishes how many printed without manual repair.
Claimed by
Meshy, March 2026
Open
01
Printed benchmark for AI parts
A public set of about 30 written specifications for small working parts a desktop printer can make, such as a clip that must grip a 12 mm rod, a hinge that must open 90 degrees, or a hook that must hold 2 kg. Each AI tool designs each part from the text alone. Every design is then sliced with locked settings, printed on the same printer in the same filament, and tested with gauges and weights. The score records whether it sliced, whether it printed, whether it fit, and whether it carried the load.
The briefCredit · unclaimed02
Scoring printability checkers
A public test set of models, each reviewed separately by at least three experienced print engineers for three common design faults: overhangs steeper than a set angle, which need supports (temporary structures printed underneath); walls thinner than the process minimum; and closed hollows that would trap unfused powder or resin. Then run the open printability checkers that already exist against it and publish how often each one agrees with the experts.
The briefCredit · unclaimed
Open 01 · the brief
Printed benchmark for AI parts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of about 30 written specifications for small working parts a desktop printer can make, such as a clip that must grip a 12 mm rod, a hinge that must open 90 degrees, or a hook that must hold 2 kg. Each AI tool designs each part from the text alone. Every design is then sliced with locked settings, printed on the same printer in the same filament, and tested with gauges and weights. The score records whether it sliced, whether it printed, whether it fit, and whether it carried the load.
Why it is still open
Searches of arXiv, GitHub and Hugging Face in November 2026 found benchmarks that score AI-made CAD in software, and a few papers that printed a handful of results without scoring them. MUSE (The Hong Kong Polytechnic University, Lingnan University and Curvature Flow, 2026) scores AI-generated assemblies for function, manufacturability and fit with written rubrics and an AI judge, and its authors write that "the reference designs have not been physically manufactured or assembled." The open cadclamp benchmark on GitHub has 47 prompts, among them a snap clip, a hinge and a print-in-place bearing, but it judges printability from the mesh and simulated slicer runs, and its fit and load checks are "advisory" geometry. FunCADBench (2026) says its authors "3D-printed and assembled prototypes to complement the simulation-based evaluation", as a few qualitative examples with no scores. CADPrompt, CAD-Recode and Text2CAD-Bench score geometry in software only.
Built with what exists
Claude Fable 5.2 (see the Note "Three working devices") and other AI models writing code for open CAD tools (CadQuery, build123d or OpenSCAD); cadclamp's clip and hinge prompts as a starting set, so its software scores can be set beside the physical results; PrusaSlicer or OrcaSlicer, run with one fixed settings file; one desktop filament printer; calipers, printed go/no-go gauges, and hanging weights.
First step
Write 10 specifications with one pass/fail test each, then design and print a human-made reference part for each to confirm the test can be passed, before running the first AI tool.
Done looks like
30 specifications, at least four AI tools, three copies of each part, and a published table giving each tool's slice rate, print-success rate, fit pass rate and load pass rate, with every prompt, model file, settings file and test photo released.
Open 02 · the brief
Scoring printability checkers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of models, each reviewed separately by at least three experienced print engineers for three common design faults: overhangs steeper than a set angle, which need supports (temporary structures printed underneath); walls thinner than the process minimum; and closed hollows that would trap unfused powder or resin. Then run the open printability checkers that already exist against it and publish how often each one agrees with the experts.
Why it is still open
The checkers exist; a test of whether they are right does not. The open augura package (October 2026) checks overhangs and wall thickness, @casys/mcp-dfm checks overhangs and wall thickness and says its result is "not a general printability or supplier-acceptance verdict", and the printing rules in the open text-to-cad project cover overhangs, wall minimums and trapped volumes but are, in its own words, "general, no printer's or shop's". AgentsCAD (Carnegie Mellon University, 2026) finds and fixes overhangs only and reports a single birdhouse case study. The closest labeled set, about 150 Printables models in a September 2025 paper by Janik Henn and coauthors, was labeled by the authors themselves, who say the labels "may reflect subjective judgments based on our own experience", and covers overhangs, bed adhesion, warping and fragility, not thin walls or cavities. Shapeways wrote in 2014 that it checked its automatic printability tool against its own engineers, without publishing numbers.
pypi.orgjsr.ioraw.githubusercontent.comarxiv.orgarxiv.orgshapeways.com
Built with what exists
Thingi10K, 10,000 real models with mesh-defect counts, as a source of test inputs; the checkers named above, plus the support previews in PrusaSlicer or OrcaSlicer as a baseline for overhangs; a labeling tool such as Label Studio; Python scripts, which an AI model such as Claude Fable 5.2 (see the Note "Three working devices") can help write.
First step
Pick 50 Thingi10K models, write a one-page labeling guide that defines each fault with a picture, and send the first batch to three reviewers.
Done looks like
A test set of at least 200 models, each labeled by at least three reviewers; a published measure of how often the reviewers agree; and, for each checker and fault type, its precision (the share of its flags the experts agree with) and recall (the share of expert-found faults it catches), with code and labels under open licenses.