AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents

Researchers present AI CFD Scientist, an open-source AI agent framework that autonomously conducts computational fluid dynamics research by combining literature review, physics simulation, vision-based verification, and manuscript generation. The system demonstrates measurable improvements in turbulence modeling and detects failure modes that traditional solver checks miss, representing a significant step toward AI-driven scientific discovery in high-fidelity physical simulation.

The emergence of AI CFD Scientist reflects a maturing capability in autonomous scientific research tools. While LLM-based agents have achieved notable success in software-only domains like machine learning and chemistry, extending them to physics simulations introduces unique challenges: solvers can complete without producing physically valid results, and failures often manifest only in visual field data rather than error logs. This research addresses those gaps through a vision-language verification gate that inspects rendered flow fields before accepting results.

The work builds on growing momentum in AI-assisted scientific discovery, where agents iteratively propose hypotheses, execute experiments, analyze results, and refine approaches. What distinguishes this contribution is its domain-specific architecture: three coupled pathways handle parameter optimization, C++ code generation for new physical models, and open-ended hypothesis testing against reference data. The demonstrated 7.89% improvement in turbulence modeling on a benchmark case shows tangible scientific value, while the ablation study revealing detection of 14 of 16 silent failures highlights a critical advantage over baseline approaches.

For the broader AI and scientific computing communities, this signals that autonomous research workflows are becoming viable beyond theoretical domains. The open-source release amplifies impact by enabling reproducibility and community extension. The vision-language verification approach—treating physics validity as a learnable classification problem—may transfer to other simulation-heavy fields including materials science, climate modeling, and engineering design.

Key challenges remain around scaling to more complex physics, reducing computational overhead of verification loops, and validating discoveries against real-world experiments rather than DNS benchmarks.

• →Vision-language verification gates detect physics invalidity better than solver-level checks alone, addressing a critical limitation in autonomous scientific workflows.
• →AI CFD Scientist autonomously improved turbulence modeling by 7.89% on a standard benchmark, demonstrating measurable scientific contribution beyond proof-of-concept.
• →The framework's open-ended hypothesis search capability suggests AI agents can operate beyond pre-defined parameter spaces in physical simulation domains.
• →Open-source release enables reproducibility and community validation, reducing barriers to adoption in academic and industrial CFD research.
• →Success here suggests autonomous discovery workflows may scale to other simulation-intensive fields like materials science and climate modeling.