AI CFD Scientist: Open-Ended Discovery with Physics-Aware AI Agents
Researchers have introduced AI CFD Scientist, an open-source artificial intelligence framework designed to automate scientific discovery in computational fluid dynamics (CFD). Unlike previous AI agents limited to software or biological research, this system integrates literature-grounded ideation, validated execution, and vision-based physics verification within a single workflow using OpenFOAM. A key innovation is the vision-language physics-verification gate, which inspects rendered flow fields to detect silent failures that standard solver logs miss. In tests utilizing a GPT-5.5 backbone, the agent autonomously discovered a Spalart-Allmaras runtime correction, reducing lower-wall Cf RMSE by 7.89% against DNS benchmarks. Comparative analysis showed that general AI-scientist baselines lacked the necessary domain-specific validity gates for defensible claims. The study demonstrates that integrating visual verification significantly enhances reliability, detecting 14 out of 16 planted failures. The code, prompts, and artifacts are publicly available, marking a significant step toward open-ended, physics-aware AI discovery in high-fidelity physical simulations.
Wire timeline
AI CFD Scientist: Open-Ended Discovery with Physics-Aware AI Agents
Researchers have introduced AI CFD Scientist, an open-source artificial intelligence framework designed to automate scientific discovery in computational fluid dynamics (CFD). Unlike previous AI agents limited to software or biological research, this system integrates literature-grounded ideation, validated execution, and vision-based physics verification within a single workflow using OpenFOAM. A key innovation is the vision-language physics-verification gate, which inspects rendered flow fields to detect silent failures that standard solver logs miss. In tests utilizing a GPT-5.5 backbone, the agent autonomously discovered a Spalart-Allmaras runtime correction, reducing lower-wall Cf RMSE by 7.89% against DNS benchmarks. Comparative analysis showed that general AI-scientist baselines lacked the necessary domain-specific validity gates for defensible claims. The study demonstrates that integrating visual verification significantly enhances reliability, detecting 14 out of 16 planted failures. The code, prompts, and artifacts are publicly available, marking a significant step toward open-ended, physics-aware AI discovery in high-fidelity physical simulations.
cs.AI updates on arXiv.org