Industries / Technology / Software and SaaS / Engineering design software
AI-assisted design
AI-assisted design is software that proposes, rebuilds or checks a design, of a part, a building or a circuit, by searching many candidates or by learning from past designs, instead of a person drawing every feature.
6
Milestones achieved, from 2006 to 2026.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2006
An evolved antenna flies
A team at NASA Ames Research Center let an evolutionary algorithm, a program that breeds candidate designs, scores each in a simulator and keeps the best, design the antenna for the three Space Technology 5 (ST5) satellites. When a change of orbit changed the requirements, the team adjusted the scoring and evolved a new set of antennas "in less than a month". One design was built, tested and flown on the three spacecraft, launched on 22 March 2006; the authors call it "the first computer-evolved antenna to be deployed for any application" and "the first computer-evolved hardware in space."
LessonSearch by software works where a simulator can score designs quickly and well, and the engineer still chooses the basic form, here a single wire.
Credit
Gregory Hornby, Jason Lohn, Derek Linden and Al Globus, working at NASA Ames Research Center
2018
A million CAD models shared
Sebastian Koch, Daniele Panozzo and seven co-authors released the ABC dataset, one million CAD (computer-aided design) models in which each model is a set of exactly described curves and surfaces, so that learning methods could train on real engineering geometry and be compared on the same data. Other teams use it: the organizers of the SHARP 2023 challenge on rebuilding CAD from scans name it among the datasets the field works from.
Credit
Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin and Daniele Panozzo
2021
Models learn the CAD steps
DeepCAD, presented at the ICCV 2021 conference, describes a shape as the sequence of operations a designer would use, such as sketching a profile and extruding it, rather than as a mesh or a cloud of points; its authors call it "the first 3D generative model" of this kind. They released 178,238 models with their construction sequences, so a generated model can be opened and edited like one a person built. Later work builds on that set: GenCAD-3D (2025) extends the DeepCAD dataset to train a method that rebuilds CAD from scans.
Credit
Rundi Wu, Chang Xiao and Changxi Zheng
2023
A chip layout claim retested
In June 2021 a Google team reported in Nature that reinforcement learning, in which a program improves by trial and reward, produced chip floorplans (the placement of large blocks such as memories) "superior or comparable to those produced by humans" in under six hours. A UC San Diego team rebuilt the test in public with open test cases; in the March 2026 version of its study, a stronger form of an older method, simulated annealing, gave shorter wiring than Google's released AlphaChip code in 7 of 9 cases, while AlphaChip gave better timing in 6 of 9, and the team writes that checking the claims has been hampered because neither the data nor the code is fully available. Nature put an editor's note on the paper in September 2023 saying its claims "have been called into question", then removed it on 26 September 2024 after a review by the editors and a panel, adding a clarifying addendum. The Google authors replied in November 2024 that the critics did not pre-train the method and used "20x fewer RL experience collectors and half as many GPUs".
LessonA claim about design software is settled only when its code, data and test cases are released with it.
Credit
Anna Goldie, Azalia Mirhoseini, Jeff Dean and colleagues, Google; Chung-Kuan Cheng, Andrew Kahng and colleagues, UC San Diego
2024
AI misreads the rulebook
Anna Doris and Faez Ahmed of MIT, Daniele Grandi of Autodesk Research and colleagues built DesignQA, a public test drawn from the rules of the Formula SAE student racing competition, CAD images and engineering drawings. The models they tried, GPT-4o, GPT-4, Claude Opus, Gemini 1.0 and LLaVA 1.5, struggled "to reliably retrieve relevant rules", to recognize components in CAD images and to analyse engineering drawings. The test has a public leaderboard and was set as a hackathon problem at the ASME IDETC conference in 2025.
Credit
Anna C. Doris, Daniele Grandi, Ryan Tomich, Md Ferdous Alam, Mohammadmehdi Ataei, Hyunmin Cheong and Faez Ahmed
2026
Runs, but fails the brief
Harmanjot Singh, Abhra Dubey and Jorge Alejandro Amador Herrera of the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi tested eight AI models on CADEngBench: 600 tasks of writing or editing parametric CAD programs, each checked for valid geometry, design and manufacturing rules, behaviour when its parameters change, and response to load in a finite element analysis (a calculation that splits a part into small pieces to find how it bends). Of 2,400 generated programs, 1,030 ran, and only 432 of those also met their stated engineering requirements, so 58.1% of the programs that ran broke at least one. Editing a supplied model proved easier than making a new one.
LessonA model that looks right and runs is not yet a model that does its job.
Credit
Harmanjot Singh, Abhra Dubey and Jorge Alejandro Amador Herrera
Watch for
Flow results in minutes
Siemens says Simcenter PhysicsAI, an add-on to its STAR-CCM+ fluid simulation software, trains AI reduced-order models (fast stand-ins for a full simulation) on a customer's past CFD (computational fluid dynamics) results and runs what-if studies "in the order of 1,000x faster" than the usual workflow, with predictions "up to 100x faster on GPU versus CPU". It says built-in error measures compare predictions with full simulations.
Achieved whena user outside Siemens publishes, for designs not used in training, the predicted and fully simulated results side by side, with the error and the time saved.
Claimed by
Siemens Digital Industries Software, May 2026
Scans into editable CAD
Backflip AI says its updated model turns 3D scans and mesh files into editable CAD models with a feature tree, the list of steps that built the part, in one to five minutes and for about $10 a part, against the "$1,500 or more" of a skilled person's reverse engineering, and that it does best on "moderate-complexity 3-axis CNC milled and turned parts."
Achieved whenan independent tester rebuilds a stated set of real scanned parts and publishes each model's deviation from the original part and how many needed repair by hand.
Claimed by
Backflip AI (chief executive Greg Mark), August 2026, reported by DEVELOP3D
IDS files drafted by AI
Ryo Kanazawa and eleven co-authors released Ishigaki-IDS, an open-weight AI model (its trained weights can be downloaded) that drafts IDS files, the checkable lists of what a building model must contain, from written requirements. On their own benchmark of 166 expert-written cases in Japanese and English, its 8-billion-parameter version scored 0.651 on a validator-pass measure based on buildingSMART's IDS audit tool, against 0.331 for Claude Opus 4.5, the strongest general model they tried; in a check with six BIM practitioners, drafting with it cut total work time by 54.7%.
Achieved whena team outside the authors uses it on a real project's requirements and publishes the pass rate and the time saved.
Claimed by
Ryo Kanazawa and co-authors, June 2026
Open
01
Scans to CAD, scored
A public set of real parts with their original CAD models, each part scanned, and a scoring script that measures how closely each tool's rebuilt, editable model matches the original, and whether its feature tree still rebuilds after simple edits. Turning a physical part back into CAD is still mostly skilled hand work.
The briefCredit · unclaimed02
Fit bugs from real hardware
A test set built from real fit problems reported on open-hardware projects. Each task gives the part files from before the fix, the issue text, and a check, such as an overlap or clearance test against the mating part, that the project's own fix passes. AI design reviewers are scored on whether they find the problem from the files alone, and AI design agents on whether their fix passes the check. Software has such a test, built from real GitHub issues (SWE-bench); mechanical design does not.
The briefCredit · unclaimed
Open 01 · the brief
Scans to CAD, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of real parts with their original CAD models, each part scanned, and a scoring script that measures how closely each tool's rebuilt, editable model matches the original, and whether its feature tree still rebuilds after simple edits. Turning a physical part back into CAD is still mostly skilled hand work.
Why it is still open
Backflip's cost and time figures, under Watch for, are its own. The closest public set, from Nomi Yu, Md Ferdous Alam, A. John Hart and Faez Ahmed (GenCAD-3D, accepted in the ASME Journal of Mechanical Design), holds 51 laser-scanned parts, all 3D printed, and scores the team's own method. The SHARP 2023 challenge used scans "obtained by virtually scanning" CAD models, not scans of real parts.
Built with what exists
NIST's free test models, which "can be used without any restrictions", including a four-part box assembly remodelled in 2018 to be easier to manufacture, made by an online machining service and on a desktop printer; a desktop structured-light scanner and phone photogrammetry; the tools under test, such as Backflip AI's Fusion add-in, GenCAD-3D's code once released, and an AI model writing CadQuery code from a mesh, such as Claude Fable 5.2 (see the Note "Three working devices"); Open3D or trimesh to measure distances between surfaces; CadQuery or FreeCAD to apply the test edits.
First step
Print five NIST test parts, scan them with a phone photogrammetry app, and measure one tool's rebuilt models against the original STEP files.
Done looks like
At least 30 parts, at least 10 of them machined in metal, each scanned with two scanners; for each tool, the mean and 95th-percentile distance from the original surface, the share of key dimensions within 0.1 mm, the share of models that still rebuild after three set dimension changes, and minutes per part, all published with the scans.
Open 02 · the brief
Fit bugs from real hardware
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test set built from real fit problems reported on open-hardware projects. Each task gives the part files from before the fix, the issue text, and a check, such as an overlap or clearance test against the mating part, that the project's own fix passes. AI design reviewers are scored on whether they find the problem from the files alone, and AI design agents on whether their fix passes the check. Software has such a test, built from real GitHub issues (SWE-bench); mechanical design does not.
Why it is still open
Public CAD benchmarks set their own tasks rather than use reported defects. CADGenBench asks for parts from engineering drawings and for requested changes to STEP files, and CADEngBench (2026) scores generation and editing of 300 parametric parts and 150 pairs of parts. Searches in November 2026 found no benchmark built from defect reports. The raw material exists: closed issues such as "Bowden retainer does not fit over a T-nut" (Voron 2), "Bottom panel clip does not fit side fan support" (Voron Trident, two printed parts that collide when screwed together) and "3D printed battery holder doesn't fit battery" (NASA JPL's Open Source Rover).
Built with what exists
GitHub's issue search and commit history to find fit issues closed by a change to a part file; trimesh for overlap tests between mesh files and FreeCAD or CadQuery for STEP files; AI models that read images and write CAD code, such as Claude Fable 5.2 (see the Note "Three working devices"), each run several times per task because one pass says little about reliability (see the Note "All twenty times").
First step
Pull ten closed "does not fit" issues from the Voron repositories, find the commit that changed each part, and confirm that the files from before and after can be loaded and tested for overlap.
Done looks like
At least 40 tasks from at least five projects, each with a check that fails on the old files and passes on the project's fix; published catch rates (problem found and placed on the right part) and fix rates for at least three AI agents and one human designer; every task used under its project's own license, such as GPL-3.0 for Voron 2 and Apache-2.0 for the Open Source Rover.