Industries / Technology / Software and SaaS / Engineering design software
Simulation
Simulation software predicts how a design will behave before it is built: how a part bends or breaks, how air or water flows past it, how heat moves through it, and what happens to it in a crash.
6
Milestones achieved, from 1969 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1969
A shared stress program
In the 1960s Thomas Butler, an engineer at NASA's Goddard Space Flight Center, pressed for one general program to calculate stresses in complex structures. In 1966 NASA hired the MacNeal-Schwendler Corporation (MSC) of Los Angeles to write it, and the first working version of NASTRAN (NASA Structural Analysis) was delivered in 1969. NASA shared it through its software library, COSMIC; carmakers took it up after the 1973 oil embargo set off work on lighter cars, and in 1982 MSC won the right to sell its own versions, by NASA's account.
LessonA program paid for once and shared widely became a common tool across aircraft, spacecraft, ships and cars.
Credit
Thomas Butler, NASA Goddard; MacNeal-Schwendler Corporation, co-founded by Richard MacNeal
1978
A crash code given away
Lawrence Livermore National Laboratory needed to simulate a bomb casing striking the ground at about 40 m/s, and no 3D program could. John Hallquist, at first its only developer, wrote DYNA3D, whose manual was released in August 1976; in 1978 the lab put the code in the public domain without restrictions. By 1988 about 600 tapes of DYNA3D and related codes had gone out, the code was used for car crashes and metal forming, and in 1989 Hallquist left to found Livermore Software Technology Corporation, maker of LS-DYNA. These figures come from talk slides by David Benson of UC San Diego, who worked on the code.
LessonBenson writes that two companies selling DYNA3D-based software created "even more interest in the free public domain version", so giving the code away grew its paid market.
Credit
John Hallquist, Lawrence Livermore National Laboratory; David Benson
1983
Benchmarks to test solvers
Finite element analysis splits a structure into many small pieces whose equations a computer solves together. By the early 1980s industry used it on real problems, and engineers worried that the programs gave wrong answers. After lobbying by industry and universities, the UK Department of Trade and Industry funded a project at the National Engineering Laboratory in East Kilbride, Scotland, and NAFEMS (first the National Agency for Finite Element Methods and Standards) was founded in 1983. It wrote standard benchmark problems and published how a range of programs scored on them. NAFEMS says software makers "quickly responded by adopting these tests" and that most major vendors now run them as part of routine quality control.
LessonPublished scores on shared test problems pushed vendors to check and fix their own programs.
Credit
UK Department of Trade and Industry; National Engineering Laboratory; NAFEMS
1998
Rules for trusting a simulation
Verification checks that a program solves its equations correctly; validation checks how closely a model matches the real world for the use it is meant for. The American Institute of Aeronautics and Astronautics (AIAA) published a guide to both for fluid flow simulations in 1998 (AIAA G-077-1998), which the consultant Len Schwer calls the first modern standards document on the subject. The American Society of Mechanical Engineers (ASME) followed with a guide for solid mechanics in December 2006 (V&V 10), NASA with a standard for models and simulations in July 2008 (NASA-STD-7009), and ASME with a standard for fluid flow and heat transfer in 2009 (V&V 20).
Credit
AIAA Committee on Standards for CFD; ASME V&V committees; NASA
2004
Fluid code goes open source
Computational fluid dynamics (CFD) calculates how liquids and gases flow. OpenCFD, a British company, published OpenFOAM 1.0, a CFD toolbox anyone may use and change without a licence fee, in December 2004. Ownership moved several times: OpenCFD was bought by Silicon Graphics International in 2011, when the OpenFOAM Foundation was set up, and became part of the ESI Group in 2012, and releases have continued.
Credit
OpenCFD Ltd; OpenFOAM Foundation
2025
Chip software buys simulation
Synopsys, a maker of chip design software, completed its purchase of Ansys, a maker of simulation software, on 17 July 2025, after announcing it in January 2024. On 23 February 2026 Cadence, another chip design software maker, completed its purchase of Hexagon's design and engineering business for about €2.7 billion. The deal brought MSC Nastran, the commercial line that grew from NASA's program, and Adams, with Hexagon's strengths in structural analysis, acoustics and multibody dynamics (the motion of linked parts). Both buyers say they want to simulate whole products, from the chip to the finished system.
Credit
Synopsys; Cadence Design Systems; Hexagon
Watch for
GPUs take over fluid runs
A GPU, or graphics processor, is a chip built to do many calculations at once. Honda reported that unsteady, high-fidelity CFD in Ansys Fluent ran 34 times faster, at 38% lower cost, on four NVIDIA GB200 GPUs than on 1,920 cloud processor cores, and that GPUs made large runs feasible that had been impractical before. The figures were given at NVIDIA's own conference and repeated by an Ansys director.
Achieved whensomeone outside Honda, Ansys and NVIDIA runs one published case on both kinds of hardware and reports the times, the costs and whether the results match.
Claimed by
Honda, with Ansys and NVIDIA, at NVIDIA GTC 2026 (reported by Autocar Professional, September 2026)
Open
01
Agent fluid runs versus experiments
Give a coding agent only the written description of validation cases in the Turbulence Modeling Resource (TMR), a public set of test flows with measured data, first hosted by NASA Langley and guided by a working group of the AIAA, such as the 2D backward-facing step and the 2D NASA wall-mounted hump. The agent sets up, meshes and runs each case in OpenFOAM, refines the grid at least twice, and reports the quantities the TMR publishes. Score every run against the TMR's experimental data and against its results from established codes using the same turbulence model.
The briefCredit · unclaimed02
Stress answers without a licence
FEABench, a benchmark from Google researchers, poses physics and engineering problems in words, each with a published target number. Restate its 15 Gold problems for open-source solvers such as Elmer, FEniCS and CalculiX, have an agent solve each from the text alone, and score the number it reports against the target.
The briefCredit · unclaimed
Open 01 · the brief
Agent fluid runs versus experiments
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Give a coding agent only the written description of validation cases in the Turbulence Modeling Resource (TMR), a public set of test flows with measured data, first hosted by NASA Langley and guided by a working group of the AIAA, such as the 2D backward-facing step and the 2D NASA wall-mounted hump. The agent sets up, meshes and runs each case in OpenFOAM, refines the grid at least twice, and reports the quantities the TMR publishes. Score every run against the TMR's experimental data and against its results from established codes using the same turbulence model.
Why it is still open
Published tests of simulation agents mostly score whether a case runs or whether an AI judge approves it. Shi and Zhang (September 2026) found that of the suites they used "Only FoamBench and FEABench check the result numerically", that "a run that executes is not necessarily physically correct", and that scores fell by about 20 points under a strict field-level check. FoamBench compares an agent's result with a reference simulation and gives full marks within a 10% error. No agent test scored against the TMR's measured data was found.
Built with what exists
OpenFOAM; the TMR's case descriptions, grids and experimental data, which moved from NASA Langley to GitHub on 24 February 2026; a coding agent such as Claude Code; isolated sandboxes to run many cases in parallel (Note: "A thousand sandboxes"); Python for scoring.
First step
Run the 2D zero-pressure-gradient flat plate by hand in OpenFOAM against the TMR's published numbers, to settle the scoring script before handing any case to an agent.
Done looks like
at least ten TMR validation cases, each run five times by the agent from the written description alone with no human edits; for each run, the change in results between grids and the error against experiment and against the TMR's reference-code results; a public table with every log; and a second person who reruns two cases and gets results inside the reported spread.
Open 02 · the brief
Stress answers without a licence
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
FEABench, a benchmark from Google researchers, poses physics and engineering problems in words, each with a published target number. Restate its 15 Gold problems for open-source solvers such as Elmer, FEniCS and CalculiX, have an agent solve each from the text alone, and score the number it reports against the target.
Why it is still open
FEABench runs only in COMSOL Multiphysics, a commercial program that needs a licence, and in September 2026 the best agent harnesses tested by Shi and Zhang solved 5 of its 15 Gold problems (33.3%) within its 10% error bound. ALL-FEM, a suite of 31 FEniCS problems, is scored on whether the code runs and on an AI judge, not on the numbers. No numerically scored test of agents on open finite element solvers was found.
Built with what exists
the FEABench repository on GitHub for the problems and targets; Elmer, FEniCS and CalculiX; Gmsh for meshes; a coding agent such as Claude Code; repeated runs to measure consistency (Note: "All twenty times").
First step
Pick one FEABench Gold problem, rebuild it by hand in Elmer or FEniCS, and confirm that the open solver reaches the published target.
Done looks like
for each of the 15 problems, a human-made open-solver model that reaches the target within 10%, or a written reason why the problem cannot be posed in open tools; then ten agent runs per problem, with pass rates at 10% and at 2% error; all models, logs and scripts public.