Can AI Accelerate Design Cycles?

Generative design promises to turn days of drawing and checking into an hour of computing. On one well-defined NASA bracket, the engineer who ran it reports something close to that. But the speed comes from moving the engineer's work, not removing it: someone still has to write down every load, interface and keep-out zone, judge what comes back, and prove it works.

Updated October 20267 min readEvidence-graded8 sources

On this page

The BFCS verdictEvery article scored on the same four dimensions

Better

PartlyModerate confidence

A wider option space, not automatically a better choiceIn a 34-person lab study, a generative tool spread designs across far more of the objective space, but participants scored better on the shaded area-to-weight measure when adjusting the design parameters themselves. A NASA engineer reports stiffer generated parts than his human baseline.

Note and confidence

Moderate confidenceA wider option space, not automatically a better choiceIn a 34-person lab study, a generative tool spread designs across far more of the objective space, but participants scored better on the shaded area-to-weight measure when adjusting the design parameters themselves. A NASA engineer reports stiffer generated parts than his human baseline.

Faster

PartlyLow to moderate confidence

Large gains on well-posed parts, mostly reported by the teams using themIn his own comparison, a NASA engineer reports that a bracket which took two engineers two days by hand needed about 30 minutes of setup and an hour of computing with generative design. We found no independent study that timed whole design cycles, including setup, validation and manufacturing review.

Note and confidence

Low to moderate confidenceLarge gains on well-posed parts, mostly reported by the teams using themIn his own comparison, a NASA engineer reports that a bracket which took two engineers two days by hand needed about 30 minutes of setup and an hour of computing with generative design. We found no independent study that timed whole design cycles, including setup, validation and manufacturing review.

Cheaper

ConditionalLow confidence

Fewer engineering hours, if the part can be madeBy the NASA team's approximate figures, its evolved antenna took about three person-months against about five for the conventional design. Printed parts can add machining, heat treatment and qualification cost. The effort figures come from the NASA teams themselves; the Autodesk vendor case study gives no cost or time figures.

Note and confidence

Low confidenceFewer engineering hours, if the part can be madeBy the NASA team's approximate figures, its evolved antenna took about three person-months against about five for the conventional design. Printed parts can add machining, heat treatment and qualification cost. The effort figures come from the NASA teams themselves; the Autodesk vendor case study gives no cost or time figures.

Safer

ConditionalGovernance-gated

Only with a recorded constraint set and full validationA generated part is only as sound as the requirements it was given. NASA's own guide requires detailed finite element checks before anything is fabricated. In our reading, responsibility for the design stays with the engineer.

Note and confidence

Governance-gatedOnly with a recorded constraint set and full validationA generated part is only as sound as the requirements it was given. NASA's own guide requires detailed finite element checks before anything is fabricated. In our reading, responsibility for the design stays with the engineer.

The problem, in numbers

A conventional structural part is designed in loops: a designer draws it, an analyst checks it, and someone asks whether it can be made. At NASA's Goddard Space Flight Center, design, analysis and manufacturing are often done by separate groups. For one instrument bracket, an expert designer and analyst went through four iterations over two days; the fourth met the requirements but could not readily be machined or printed.^{Source 1}

In a survey for Engineers Australia's January 2025 report, 43% of working engineers reported using generative AI, against 74% of students, and 66% of engineers said they had received no generative AI training at work.^{Source 7} Those are self-reported figures for generative AI in general, and the 43% does not separate design software from chat tools. Generative design is a different idea: software that proposes shapes to meet a set of engineering requirements.

The evidence, in four numbers

2 days vs ~1.5 hours

of effort for one NASA instrument bracket, by hand versus with generative design

two engineers for two days, against about 30 minutes of one engineer's setup plus about an hour of computing (the 1.5 hours is our sum); the NASA author's own comparison, not an independent study (2022)

3 vs 5

person-months to design and fabricate an evolved versus a conventional spacecraft antenna

approximate figures reported by the NASA team (2006, abstract only); an antenna, not a structure

1.098 vs 0.453

spread of designs across the objective space, with and without a generative tool

hypervolume measure, 34 participants in a lab canopy task; the shaded area-to-weight score was better without the tool, p = 0.039 (Research in Engineering Design, 2024)

43%

of engineers in the workforce surveyed by Engineers Australia reported using generative AI

self-reported, Engineers Australia members (report published January 2025; the report does not state the exact base); generative AI in general, not generative design tools

What improves, and why

Generative design changes what the engineer spends time on. Instead of drawing a part and then checking it, the engineer encodes the problem: the desired mass, minimum factor of safety, interfaces with other parts, keep-out zones, loads, minimum natural frequency, materials and manufacturing method. The software then grows shapes that satisfy them, using topology optimisation, which strips out material that carries little stress.^{Source 1}

In our reading, that is why the speed can be large on well-posed parts. NASA's guide for its engineers says simple parts with known requirements can be designed and validated by finite element analysis in as little as a day by an experienced user.^{Source 2} It also says the method suits requirements that change over a mission: once the inputs are defined, a new design can be generated rather than redrawn.^{Source 2} NASA's earlier evolved antenna showed the same pattern. When the mission's orbit changed, the team evolved a new antenna in a few weeks.^{Source 3}

The second gain is breadth: a generative tool can return many candidate designs at once. An RMIT paper shows, in computational examples, that vastly different structural designs can be found by giving up a small amount of performance, for example 3%.^{Source 6} In a vendor case study, Autodesk reports that its software produced more than 150 valid options for a proof-of-concept General Motors seat bracket; the chosen design was 40% lighter and 20% stronger, and replaced eight parts with one.^{Source 8}

The software searches the space the engineer defines. Leave out a load, a bolt or a keep-out zone, and it will optimise confidently around the gap.

What the evidence actually says

The only time figures we found come from the people using the tools. In the NASA author's own bracket comparison, the human designs took two engineers two days. The generative design took about 30 minutes of one engineer's time to encode the requirements, plus about an hour of computing.^{Source 1} By the author's figures, the generated brackets had over three times the stiffness for their mass of the third human iteration, and in vibration tests of three of them the first natural frequency fell within 3% of prediction.^{Source 1} The organic shapes were readily machined, though the paper notes that 3D-printed parts often need extra machining, heat treatment and qualification testing, each adding cost and schedule.^{Source 1} The author reports similar gains across 20 other parts, all small to mid-sized metal structures.^{Source 1} This is one engineer's account of his own process, not an independent trial. For the antenna, the NASA team's abstract gives approximately three person-months to design and fabricate the evolved version against about five for the conventional one.^{Source 3}

Controlled studies are narrower and more mixed. In a lab experiment with 34 participants, mostly students, a multi-objective tool spread the designs produced across a much larger share of the objective space (a hypervolume of 1.098 against 0.453 without it). Yet on the shaded area-to-weight measure, participants did better adjusting the parameters themselves (p = 0.039).^{Source 4} Participants said they handed the numbers to the tool and concentrated on aesthetics; the task order was fixed, and many had not used the tool before.^{Source 4}

Interviews with six designers, each with over five years of design practice, show what setup involves. Setting up the tool is a constraint-driven process, with values taken from calculations, conservative estimates, experience and intuition. One designer iterated on the inputs 37 times. Requirements the tool could not take were worked around or added by hand; one designer added a maximum hole size at the end.^{Source 5} In our reading, the speed-up is real on bounded parts, but we found no independent study timing the full cycle: setup, iteration, validation and manufacturing review together.

Where it falls short

Self-reported speed. The headline time savings come from NASA engineers describing their own work; the vendor case study gives no times.^{Source 1,Source 3,Source 8} We found no independent industry study of end-to-end design time.
Wider is not better. More options did not produce better combined performance in the lab study.^{Source 4} Someone still has to choose well.
Not every requirement fits. RMIT's Xie writes that topology optimisation is rarely used in practice, especially on large architectural projects, because its forms often miss design requirements.^{Source 6} In one bridge example, keeping a gap for traffic cost 14% of stiffness but made the design far more likely to be accepted.^{Source 6}
Trust without understanding. Designers may not fully understand how the tool reached a result, and accept it on trust in the tool and their own setup.^{Source 5}

Where governance decides the outcome

The constraint set is the design decision, so it needs an owner. NASA's guide asks engineers to document every known requirement and its source, and warns that a missing preserve or obstacle means the model must be fixed and run again.^{Source 2} It also requires detailed finite element validation before anything is fabricated, because the optimisation mesh is too coarse to catch stress peaks at interfaces, with flight parts checked further by specialists.^{Source 2} Treat a generated shape as a proposal from a tool, reviewed and signed off by the engineer who answers for it. In Engineers Australia's survey, 82% of engineers said generative AI will always need oversight, though the report does not give that figure's base.^{Source 7} Record the inputs, the tool version and why one option was chosen, so a later reviewer can see what the software was told.

How to think about adoption

Start with bounded parts such as brackets, mounts and housings, where loads, interfaces and keep-out zones are known and a failure is caught in test. Write the requirement list before opening the tool, and treat a surprising result as a sign that something is missing from it. Time the whole cycle, not the computing: setup, iterations, validation, manufacturing review and any hand rework, against a recent conventional design of similar difficulty. Put experienced engineers on the setup, and ask manufacturers early whether the shapes can be made. Treat vendor speed claims as hypotheses to test on your own parts.

How we read this evidence

BFCS.ai does not fabricate figures. The design time figures here come from NASA engineers describing their own process; the software vendor case study gives none, and the controlled evidence is one lab study and one interview study, neither of which timed a full design cycle. We found no independent study measuring end-to-end design cycle time with generative design in industry, and we say so where it matters.

Sources

1
McClelland, R. (NASA Goddard Space Flight Center), "Generative Design and Digital Manufacturing: Using AI and robots to build lightweight instruments," conference paper, SPIE Optics and Photonics, August 2022 (NASA Technical Reports Server 20220012523; an author-reported case comparison, not an independent study; the designs were made with the Autodesk Fusion 360 Generative Design Extension). ntrs.nasa.gov
What we took from it
The author claims a greater than 10x reduction in development time and cost. EXCITE Tip/Tilt bracket: the human designs "took two days of effort by two engineers" (a design engineer and a structural analyst, four iterations, the fourth not readily manufacturable); the AI design took about 30 minutes of one engineer's effort to encode the requirements and about 1 hour of generative design computing (the ~1.5 hours in our stats is our sum of the two). Stiffness-to-mass ratio, excluding the unmanufacturable fourth human iteration: 240 Hz/kg (third human iteration) against 740 Hz/kg (AI, CNC) and 885 Hz/kg (AI, additive). Requirements encoded may include mass, minimum factor of safety, interfaces, keep-out zones, structural loads, minimum natural frequency, materials and manufacturing processes. In vibration tests of three brackets the first mode was within 3% of finite element predictions; failure loads were generally over 200% above predicted. Section 3: the organic shapes were readily CNC machined; for additive manufacturing, post-print machining (frequently needed because of poor tolerances), heat treatment and additional testing to qualify parts under NASA-STD-6030 each add to cost and schedule. Applied to 20 other applications with similar reported improvements; demonstrated so far on small to mid-sized metallic structures up to 0.6 m.
Back to text
2
McClelland, R. (NASA Goddard Space Flight Center, Code 550), "Evolved Structures Guide for GSFC Applications," guide presented to Penn State design for additive manufacturing students on 27 March 2024 and posted to the NASA Technical Reports Server in May 2024 (20240005675). ntrs.nasa.gov
What we took from it
The guide is written for Autodesk Fusion 360. For simple parts with known requirements, designs can be completed and validated by finite element analysis "in as little as 1 day by an experienced user"; document every known requirement and its source; all required preserve and obstacle geometry should be modelled, "otherwise, the model needs to be fixed and re-run"; the mass target "may need to be iterated" based on engineering judgement; before fabrication "the design must be validated by detailed FEA" because the optimisation mesh is coarse; flight parts are more extensively verified by analysis specialists; human manufacturing feedback is needed for additive and 5-axis machined parts.
Back to text
3
Hornby, G. S., Globus, A., Linden, D. S. and Lohn, J. D. (NASA Ames Research Center), "Automated Antenna Design with Evolutionary Algorithms," conference paper, AIAA Space 2006, San Jose, 19 September 2006 (NASA Technical Reports Server 20060024675; abstract only, and the effort figures are the team's own approximations). ntrs.nasa.gov
What we took from it
The engineer chooses the evolutionary technique, parameters and basic form of the antenna (a monopole wire for ST5, chosen from past experience); the evolved antenna "required approximately three person-months to design and fabricate whereas the conventional antenna required about five"; after the mission's orbit changed, a new antenna was evolved "in a few weeks".
Back to text
4
Saadi, J. I., Chong, L. and Yang, M. C. (Massachusetts Institute of Technology), "The effect of targeting both quantitative and qualitative objectives in generative design tools on the design outcomes," Research in Engineering Design 35, 2024, pp. 409-425 (peer-reviewed, open access). doi.org
What we took from it
Controlled lab experiment: 34 participants (30 students, 4 university staff and local designers) designed a 10-parameter canopy to maximise shaded area, minimise weight and meet an aesthetic brief, first by adjusting the 10 parameters themselves (the Free task) and then with a multi-objective optimisation tool returning the top 20 of 100 generated designs. The objectives and constraints were given to participants, not set by them. Hypervolume of designs 1.098 with the tool against 0.453 without, which the authors read as a broader spread across the objective space; on the shaded area-to-weight ratio participants did better without the tool (paired t-test, p = 0.039; the sentence reporting it says "qualitative", but the measure and figure are quantitative). Task order was fixed and many participants had not used the tool before.
Back to text
5
Saadi, J. I. and Yang, M. C., "Generative Design: Reframing the Role of the Designer in Early-Stage Design Process," Journal of Mechanical Design 145(4), 2023 (peer-reviewed; read as the accepted manuscript on the NSF Public Access Repository). doi.org
Accepted manuscript: par.nsf.gov
What we took from it
Six in-depth interviews with practising and student designers who use commercial generative design tools, all with over five years of general design practice. Defining the tool inputs is "a constraint-driven process"; initial values come from calculations, conservative estimates, experience and intuition; one designer iterated on the inputs 37 times; designers "may not fully understand" how the tool reached its designs and rely on trust in the tool and their own setup; requirements the tool could not take were handled by workarounds or by hand (one designer added a maximum hole size in the final design stage). The authors note the small sample.
Back to text
6
Xie, Y. M. (Centre for Innovative Structures and Materials, RMIT University, Melbourne), "Generalized topology optimization for architectural design," Architectural Intelligence 1, article 2, June 2022 (peer-reviewed journal, open access). doi.org
What we took from it
Use of existing topology optimisation techniques in practical applications, especially large-scale projects, "is rare because the generated forms often cannot satisfy all the design requirements of architects and engineers"; vastly different designs can be obtained by sacrificing a small amount (for example 3%) of structural performance; adjusting support locations by trial and error is "tedious and time-consuming"; in one bridge example, adding a prohibited region so traffic could pass underneath gave a design 14% less stiff but far more likely to be accepted. Both figures come from computational design examples, the bridge from his group's earlier work (Yang et al., 2019).
Back to text
7
Engineers Australia, "The impact of AI and generative technologies on the engineering profession," January 2025 (research by Ergo Strategy, commissioned in 2024). engineersaustralia.org.au
What we took from it
Online survey of 2,197 Engineers Australia members and educators (1,465 engineering professionals, 448 students, 53 educators and 231 others such as retirees), with the career sample weighted by age and company size; plus 34 interviews. 43% of engineers in the workforce reported using generative AI against 74% of students (the report does not state the exact base for either); 66% of engineers had received no generative AI training at work (career sample, N = 1,465); 82% of engineers believe generative AI "will always require oversight" (the report does not state the base for this figure). The survey covers generative AI generally, not generative design software.
Back to text
8
Vendor case study (self-reported, not peer-reviewed). adsknews.autodesk.com
What we took from it
Autodesk News, "How GM and Autodesk are using generative design for vehicles of the future," 3 May 2018. Seat bracket, described as an initial proof of concept by GM and Autodesk engineers working together: the software produced "more than 150 valid design options" from parameters the engineers set; the chosen design is 40% lighter and 20% stronger than the original and consolidates eight components into one 3D-printed part. No design time or cost figures are given; production status is not stated.
Back to text