AI CFD Scientist catches silent solver failures with vision-language physics gate
Open-source framework from RPI automates computational fluid dynamics research end-to-end, using a vision model to detect physical failures that solver logs miss.
AI CFD Scientist is an open-source framework from researchers at Rensselaer Polytechnic Institute that automates computational fluid dynamics research end-to-end — from reading papers through modifying C++ solver code to writing illustrated manuscripts. The system runs on OpenFOAM via Foam-Agent and centers on a vision-language model that inspects rendered flow fields before accepting any result, a design choice that addresses CFD's core challenge: solver completion does not guarantee physical validity, and many failure modes appear only in field-level imagery rather than error logs. Three execution pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator.
Under a shared GPT-5.5 backbone across five tasks, the framework autonomously discovered a Spalart-Allmaras runtime correction that cut lower-wall skin-friction RMSE against direct numerical simulation by 7.89 percent on the periodic hill at Reynolds number 5,600. Two general AI-scientist baselines — ARIS and DeepScientist — executed partial CFD workflows under matched token budget but lacked domain-specific validity gates and failed to produce defensible claims. A controlled planted-failure ablation showed the vision-language gate caught 14 of 16 silent failures that solver-level checks missed. Code, prompts, and run artifacts are available on GitHub.