Can AI Cut Prototyping Costs?

The problem, in numbers

Physical prototypes can be a slow and costly way to find out whether a design works. Each one has to be built, instrumented and tested, and destructive tests use it up. The simulation vendor Ansys, in a 2022 press release, put the cost of a single full-scale race car crash test at an estimated $500,000, and said simulation let NASCAR limit development of its Next Gen car to two full-scale physical crash tests, saving $1 million in physical testing material costs.^{Source 3} These are self-reported vendor figures, not audited.

One long public record of the trade-off comes from aerospace. Boeing reports that it tested 77 wings in wind tunnels for its 1980-era aircraft, the 767 and 757, then 18 for the 777 and 11 for the 787, while its computational fluid dynamics (CFD) runs grew to 60 times the 777-era level.^{Source 2} These are Boeing's self-reported figures from a 2008 conference presentation.

Researchers from the Defence Science and Technology Group (DST Group) wrote in a 2019 conference submission that a virtual fatigue test is the end goal for its Aerospace Division, giving Defence an agile and affordable way to certify aircraft structures for fatigue, and listed lower certification costs among the benefits of more accurate predictions.^{Source 1}

What improves, and why

Two kinds of model do the work. Physics-based simulation, such as finite element analysis and CFD, solves the governing equations directly. Machine-learning surrogates learn to approximate a simulation or an experiment, so many variants can be screened without running the full model each time. Physics-informed machine learning sits between the two, building physical laws into training so the model needs less data.^{Source 4}

The cost case is the same for each: replace some physical builds with computed ones, and use the builds that remain to check the model rather than to search the design space. The US Food and Drug Administration gives a plain example from medical devices: a validated finite element model picks the worst-case sizes in an implant family, so fewer articles need bench durability testing.^{Source 5} In our reading, that is the realistic pattern. The model narrows the field, and the physical test confirms the answer.

Physics-informed methods earn their place where test data is scarce. A 2021 peer-reviewed review reports a laboratory study in which combining many finite element simulations with sparse experimental data cut the error in inferring the yield stress of 3D-printed alloys from over 100% to below 5%.^{Source 4} That is a measured result quoted second-hand, in a materials-testing setting, not a saving on a product programme.

What the evidence actually says

The evidence is thinner than the promise. We found no independent, peer-reviewed study that measured the total cost of a development programme before and after machine-learning surrogates were introduced. What we found falls into three kinds.

Company-reported reductions. Boeing's self-reported wing record is the longest-running example we found, and it concerns physics-based CFD, not machine learning. Boeing reports testing 11 wings for both the 737NG and the 787, while its CFD runs rose from four times to 60 times the 777 baseline.^{Source 2} In our reading, extra computing stopped buying fewer wind-tunnel wings at that point. The same presentation still describes an "extensive wind-tunnel test program" for the 787, and treats certification by simulation as an open question.^{Source 2}

Vendor case studies. The Ansys figures are self-reported by the software vendor. Its release also says physical crash tests were later used to verify the simulation models.^{Source 3} In our reading, the physical test did not disappear, it moved from search to confirmation.

Research on the methods. The peer-reviewed literature is candid about limits. The 2021 review warns that purely data-driven models can fit observations well yet give physically inconsistent predictions through extrapolation or biased data, and that for well-posed forward problems needing no data assimilation, conventional grid-based solvers still outperformed physics-informed neural networks when it was written.^{Source 4} It also says meaningful benchmarks are still needed.^{Source 4} DST Group's Australian programme compares blind predictions with fatigue test data, aiming to close the gaps between predictions and test results.^{Source 1}

Where it falls short

• Outside the envelope, the model guesses. A surrogate trained on one range of loads, materials or shapes can extrapolate badly beyond it.^{Source 4}

• Validation is a cost, not a formality. The FDA treats calibration against the data used to build a model as weak evidence compared with validation against independent data.^{Source 5} It notes that most ASME V&V 40 validation activities assume well-controlled bench testing.^{Source 5}

• The standards are written for physics-based models. FDA medical device guidance on simulation credibility does not apply to standalone machine-learning models, and NAFEMS's new validation guidelines, described in a conference presentation, focus on physics-based models.^{Source 5,Source 6}

• Adoption has lagged the promise. In July 2024, US auditors found four of six hypersonic weapon efforts had not adopted leading practices for digital engineering tools. Three have since met the related recommendations; one program office concluded a complete digital twin was not practical or cost effective.^{Source 7}

Where governance decides the outcome

The method for deciding how far to trust a model already exists. ASME V&V 40, a standard the FDA recognises, scales the evidence a model needs to its risk: how much the decision leans on the model, and how serious the outcome would be if that decision were wrong.^{Source 5} In our reading of that logic, a surrogate that only ranks concepts for a later physical test has low influence. One that replaces a qualification test has high influence and needs far more evidence. A review presented at the NAFEMS World Congress (a conference presentation, not peer-reviewed) points to the weak spot in practice: simulation quality assurance in industry depends largely on individual analysts, and designers and managers rarely question the results.^{Source 6} In our reading, every surrogate needs a written context of use, a named owner, validation data it was not trained on, and a rule for when a physical test is still required. The US auditor found Defense policy did not describe how to apply leading testing practices under digital engineering, and GAO listed all 13 of its recommendations as open as of March 2026.^{Source 7}

How to think about adoption

Start where the physics is well understood and you already hold test data: a component family, a known material, a familiar load case. Use simulation and surrogates to cut the search, not the sign-off. Record which designs the model ranked and which physical tests confirmed it, so the saving can be measured rather than assumed. Budget for validation tests from the start, because they are what turns a cheap prediction into a trustworthy one. Treat vendor reductions in prototype counts as claims to test against your own programme. And keep a physical test wherever a wrong answer could put people at risk, until the model's credibility evidence matches the consequence.