AI CFD Scientist
Open-source AI scientist for computational fluid dynamics that spans literature-grounded ideation, validated solver execution, vision-based physics verification, source-code modification, and figure-grounded writing in one inspectable OpenFOAM workflow.
| Affiliation | Rensselaer Polytechnic Institute, Pan group (csml-rpi) |
| First introduced | 2026-05 (arXiv:2605.06607) |
| Lifecycle stages | Multi-stage (ideation + execution + verification + source-code modification) + Writing (LaTeX manuscript generation) |
| Autonomy level | Semi-autonomous — user supplies a research topic and optional base case; the system runs end-to-end across three coupled pathways |
| Domain focus | Computational fluid dynamics (OpenFOAM-based) |
| Availability | Open source — code, prompts, and run artifacts at csml-rpi/cfd-scientist |
Approach
A central Orchestrator dispatches work across three first-class pathways running on OpenFOAM via Foam-Agent: (i) regular experimentation via literature-aware ideation, requirement validation, mesh-independence gating, and parametric sweeps; (ii) code modification that patches and compiles case-local C++ model libraries (headers, custom solvers, function objects); (iii) open-ended discovery that wraps both modules in an outer hypothesis loop driven by an LLM hypothesis generator and cross-case analysis. A vision-language physics-verification gate inspects PyVista-rendered flow fields before any result is accepted, rerun, or written into a manuscript — addressing failure modes (wrong geometry, missing flow features, sign errors in derived quantities) invisible to solver logs. Outputs include a LaTeX manuscript synthesized against figure-grounded evidence.
Validation
Five tasks under a shared GPT-5.5 backbone. In the open-ended task, the system autonomously discovered a Spalart–Allmaras runtime correction that reduced lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Re_h = 5600. A controlled planted-failure ablation showed the vision-language gate detected 14 of 16 silent failures missed by solver-level checks. Under matched LLM cost, two general AI-scientist baselines (ARIS, DeepScientist) executed partial CFD workflows but lacked the domain-specific validity gates.
Notable results
- First open-source CFD-specific AI scientist spanning ideation through manuscript writing in a single inspectable workflow.
- Autonomous discovery of a Spalart–Allmaras turbulence-model runtime correction validated against DNS reference data.
- Vision-language physics gate catches silent solver failures invisible to solver logs.
Primary paper
Other references
None yet.
Code
csml-rpi/cfd-scientist — code, prompts, and run artifacts.