Extending AI's scientific discovery capabilities beyond software and dry lab sciences into high-fidelity physical simulators, particularly computational fluid dynamics (CFD), has been a significant hurdle. Traditional AI agents falter because solver completion doesn't guarantee physical validity, and many critical failure modes manifest visually in field-level imagery, eluding solver logs.
Closing the Physical Validity Gap with Vision
The breakthrough lies in the introduction of AI CFD Scientist, an open-source AI scientist designed to navigate the full scientific discovery loop within CFD. This framework uniquely integrates literature-grounded ideation, validated execution, and crucially, vision-based physics verification. A central component is a vision-language gate that scrutinizes rendered flow fields before any result is deemed acceptable, rerouted for further analysis, or incorporated into a manuscript. This addresses the core limitation of previous AI approaches in physical sciences: ensuring not just computational success, but physical realism.
Autonomous Discovery and Validation in OpenFOAM
AI CFD Scientist operates through three coupled pathways within the OpenFOAM environment via Foam-Agent. These pathways enable parameter sweeps, case-local C++ library compilation for novel physical models, and open-ended hypothesis searches against reference comparators. Demonstrating its efficacy, the system autonomously discovered a Spalart-Allmaras runtime correction that reduced lower-wall Cf RMSE against DNS by 7.89% on a periodic hill at Reh=5600. Importantly, when compared against matched LLM costs, established AI-scientist baselines like ARIS and DeepScientist executed only partial workflows, lacking the domain-specific validity gates to produce defensible scientific claims. A planted-failure ablation study highlighted the vision-language gate's strength, detecting 14 out of 16 silent failures that solver-level checks missed.
