AI design in engineering means software that proposes a part's shape from stated loads, mounting points, and keep-out zones, and NASA has used two versions of it: an evolutionary algorithm that designed the X-band antenna on the three Space Technology 5 spacecraft launched on March 22, 2006, and the commercial generative design software behind the "evolved structures" built at Goddard Space Flight Center. Both produce shapes a human would be unlikely to draw, and both still depend on engineers to set the requirements and to check the result.
Key takeaways
- The Space Technology 5 antenna is described by its authors as the first computer-evolved hardware in space.
- Goddard's evolved structures use commercial generative design software, with a CAD specialist defining the fixed features first.
- NASA reports weight savings of up to two-thirds and stress factors almost ten times lower than a part by an expert human, as the center's own claims rather than independent tests.
- The lead engineer says the algorithms need a human eye because they can make structures too thin.
- Results come from specific parts, so they do not describe AI design in general.
The antenna that evolved
NASA's Space Technology 5 mission flew three small spacecraft that needed a compact antenna with a wide beamwidth and a wide bandwidth. Researchers at NASA Ames Research Center, with a collaborator from Linden Innovation Research, tried a different route from hand design. In a 2004 paper listed on the NASA Technical Reports Server, Jason Lohn, Gregory Hornby, and Derek Linden describe two evolutionary algorithms that generate and score many candidate antenna shapes. One used a genetic-algorithm representation that did not allow branching arms; the other used a tree representation that did.
A later journal article in Evolutionary Computation (2011) reports that the best antennas from both methods beat a hand-designed antenna from the contractor in testing. After the spacecraft's orbit changed, the requirements changed too, and the team evolved a new set of antennas in under a month by adjusting the fitness function, the scoring rule that tells the algorithm what a good antenna is. One design was approved for the three spacecraft, which launched on March 22, 2006. The authors call it the first computer-evolved hardware in space.
Evolutionary search is an older branch of AI than the neural networks behind current chatbots. It is included here because it shows the same pattern: people define the goal, and software explores more shapes than a person could draw.
Evolved structures at Goddard
Goddard Space Flight Center, in Greenbelt, Maryland, lists generative design as a capability on its engineering technology pages. Research engineer Ryan McClelland pioneered the approach for specialized, one-off parts, using commercially available AI software together with digital manufacturing. The workflow starts with a CAD specialist, who defines mounting surfaces, bolts and fittings, and keep-out zones for laser paths and optical sensors, and who leaves room for technicians' hands during assembly. The software then generates the structure.
NASA's page names an aluminum scaffold for the back of the EXCITE balloon-borne telescope, planned for a test flight in fall 2023, and an instrument optical bench. Trade coverage in February 2023 also reports use in balloon observatories, space weather monitors, space telescopes, and the Mars Sample Return mission, but that list comes from a news article rather than the NASA page, and the current status of each project was not checked.
What NASA claims, and how to read it
| Claim on the Goddard page | Who is the source | What is missing |
|---|---|---|
| Up to two-thirds less weight than traditional components | NASA Goddard | Which parts, and compared with which baseline |
| Stress factors almost ten times lower than a part by an expert human | McClelland, based on NASA-standard stress analysis | Number of parts compared, and test conditions |
| Complex structures in an hour or two; design, analysis, and fabrication in as little as one week | NASA Goddard | Whether the time includes review and certification |
| Performance improved three times and development time and cost cut ten times | Summary on the same page | Method behind the ratios |
These are center-reported figures for a small number of custom parts. They suggest the method can be useful when parts are one-offs and weight matters, as on a spacecraft, but they do not show the same gains for mass-produced goods, buildings, or consumer products. Independent comparisons were not found.
Where the human stays in the loop
McClelland says the algorithms need a human eye, and that left alone the software can make structures too thin. Generated parts are also run through NASA-standard validation software to find potential failure points. The design brief matters as much as the algorithm: if a keep-out zone or load case is missing from the requirements, the software will optimize a part that fails in a way no one specified.
That pattern appears in the antenna story as well. The fitness function encoded what the team wanted, and when the mission changed, a person rewrote the function. Optimization software rewards whatever it is told to reward, so design judgment moves into specifying and checking, and does not disappear.
Questions to ask about any AI-designed part
- What requirements, loads, and constraints were given to the software?
- Who reviewed the output, and with which analysis tools?
- What was the baseline for the claimed improvement?
- Was the part built and tested, or only simulated?
- Can the shape be manufactured and inspected with the processes available?
This article gives general information about published engineering examples. It is not engineering, safety, or procurement advice.
FAQ
Is an evolutionary algorithm really AI?
It belongs to the broad field of AI, in the family of search and optimization methods, though it differs from neural networks trained on large datasets.
Does AI design parts without human input?
Not in the cases described here. Engineers set the requirements, review the output, and run validation before a part is used.
Were the evolved structures flown?
The Goddard page described a scaffold planned for a test flight in fall 2023. The outcome was not confirmed in the sources reviewed, so it is not reported here.
Do these results apply to other industries?
The principle, defining constraints and letting software search, applies widely. The reported savings come from specific spacecraft parts and should not be assumed elsewhere.
Related reading
- AI in ship design
- What counts as artificial intelligence
- AI's role in modern society
- AI compared with data science
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Generative Design, Engineering and Technology Directorate, NASA Goddard Space Flight Center, accessed October 2026
- Lohn JD, Hornby GS, Linden DS, An Evolved Antenna for Deployment on NASA's Space Technology 5 Mission, NASA Technical Reports Server, 2004
- Hornby GS, Lohn JD, Linden DS, Computer-Automated Evolution of an X-Band Antenna for NASA's Space Technology 5 Mission, Evolutionary Computation 19(1), 2011
- NASA is using AI to help design lighter parts, Popular Science, February 16, 2023