General Motors has cut its finite element analysis runtime from 15 hours to one minute by replacing brute-force simulation with probabilistic AI methods, according to chief product officer Sterling Anderson. Full vehicle dynamics simulations that previously took 15 to 18 hours now complete in under one minute. The shift collapses GM's traditional engineering relay — design to aerodynamics to structures to software — into a single concurrent optimization loop.
Anderson, who joined GM just over a year ago after co-founding Aurora in 2016, calls this the third epoch of engineering. The first was empirical iteration. The second introduced virtual tools like computational fluid dynamics and finite element analysis, but kept engineering siloed. The third, he argues, is a probabilistic collapse of those silos into one model.
“Our FEA runs that historically were 15 hours per run? They're now one minute.”— Sterling Anderson, GM Chief Product Officer
The speed gain is the news, but the throughput gain matters more. When a simulation runs overnight, an engineer gets one answer per day and prays the setup was correct. When it runs in 60 seconds, the engineer iterates thousands of times in an afternoon and explores design space that was previously inaccessible. Anderson said the team is now "pumping through iterations at a much faster clip" and running broader test sets than the calendar previously allowed.
Key facts
- 01GM has compressed finite element analysis runs from 15 hours to one minute using probabilistic AI methods.
- 02Full vehicle dynamics simulations that took 15 to 18 hours now finish in under one minute.
- 03Engineers can now run thousands of designs of experiments against a single maneuver, like a 40 mph collision avoidance test.
- 04Sterling Anderson joined GM as chief product officer just over a year ago after co-founding Aurora in 2016.
- 05GM is sharing tooling monthly with its NASCAR and Formula One motorsports programs.
Jason Fischer, GM's executive director of virtual integration engineering, walked through a concrete example: Consumer Reports' avoidance test, in which a car must swerve at speed to dodge an obstacle. GM now models every sensor, electronic control unit, domain controller, and software stack in one virtual environment, then runs thousands of designs of experiments against the maneuver. The output is a vehicle tuned not for the specific test but hardened against real-world variance.
Crash engineering benefits similarly. Engineers can identify structural weak points and reinforce them before a physical prototype ever hits a barrier at 40 mph. Fischer said the time savings are not about engineers getting home earlier — they are about how many follow-up questions can be asked in the same shift.
GM holds IP on the integration architecture that lets vehicle physics, ECU behavior, and software run together in the same simulation. Fischer said the company can change physical parameters and run thousands of designs of experiments to see how the control logic responds, which is how a setup gets hardened against conditions the test track never reproduces.
The tooling extends beyond crash and dynamics. HVAC system design — historically a sequence of independent component optimizations followed by integration calibration — is now done as a single concurrent balance of airflow, refrigerant behavior, and cabin comfort. Work that took weeks or months now takes hours or days. Digital twins of assembly lines are also built before physical hardware is installed, letting GM debug factories virtually.
Motorsports is the cross-pollination layer. Fischer said GM co-develops these tools with its NASCAR and Formula One programs, with monthly technology transfers between the racing side and the production side. Racing teams have always operated under tighter iteration windows than production engineering, and the tooling overlap means production cars now benefit from techniques refined under race-weekend pressure.
The skeptical read is that AI-accelerated simulation only shifts the bottleneck. Faster iteration is only useful if the underlying physics models are accurate, and probabilistic surrogates trained on simulation data can drift from physical reality in ways that show up only at homologation or, worse, in the field. The IBM and Dallara research GM cites as analogous shows correlation between AI-accelerated CFD and physical results, but correlation at one design point does not guarantee it across the full operating envelope. Validation against physical prototypes still has to happen — the question is how much of it.
For the auto industry, the implication is a widening gap between automakers that have rebuilt their engineering stack around concurrent virtual development and those still running sequential analysis. Development cycles measured in years compress when every engineering loop runs three orders of magnitude faster. GM is also signaling that the competitive moat in carmaking is shifting from physical tooling and supplier relationships toward the integration software that ties physics, electronics, and code into one optimization problem — a moat that looks much more like what software companies have than what Detroit has historically defended.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
