AI Agents Ran 140 Physics Simulations on Their Own — With a Human Check

Engineers who design jet engines, car bodies, or medical pumps rely on something called CFD — computational fluid dynamics, which is just a fancy way of saying 'simulate how air and liquids move on a computer.' Setting up each simulation is slow, fiddly work: you have to describe the shape, build a grid of tiny cells over it, write scripts to run the job, and then process the results. It's the kind of work that eats days.

A team of researchers wanted to see if AI coding agents (AI that can write and run code for you) could take over that grunt work. They gave GPT-5.5, running inside OpenAI's Codex tool, a job: build a 'case bundle' for an OpenFOAM fuel-injector study. A case bundle is basically a recipe kit — the simulation settings, the shape-processing steps, the scripts to run it on a remote supercomputer, the results-processing code, and a record of who reviewed what. The AI wrote all of it, under human review.

Then came the real test. The team replayed that reviewed recipe kit to run and analyze 140 different injector shapes on a remote high-performance computing system. No starting from scratch each time — just re-running a trusted recipe. They also ran a second experiment using a different agent, Pi, wired up to four different AI models. All four runs succeeded and produced verified results, though they used different amounts of compute and tool calls along the way.

The takeaway isn't that AI replaced the engineers. It's that the work split cleanly in two: AI handles the repetitive execution, and humans keep the engineering judgment — deciding whether a result is physically sensible, not just whether the code ran. That division is the whole point. It means a small team could screen hundreds of design ideas instead of a handful.