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

Problem
This preprint addresses the gap in the integration of machine learning (ML) with high-fidelity physical simulations in computational fluid dynamics (CFD). While recent advancements in large language model (LLM)-based agents have made strides in scientific discovery across various domains, extending these capabilities to CFD is challenging due to the complexities of physical validity and the presence of failure modes that are not captured in solver logs. The authors propose the AI CFD Scientist, which aims to create a comprehensive workflow that encompasses ideation, execution, verification, modification, and documentation within a single framework.

Method
The core technical contribution is the AI CFD Scientist framework, which operates on OpenFOAM and utilizes a vision-language physics-verification gate. This gate inspects rendered flow fields to ensure physical validity before results are accepted or documented. The framework consists of three pathways: (1) parameter sweeps using a fixed solver, (2) case-local C++ library compilation for new physical models, and (3) open-ended hypothesis generation against a reference comparator. The system is built on a shared GPT-5.5 backbone, which facilitates the autonomous discovery of new models and corrections. The authors specifically highlight the discovery of a Spalart-Allmaras runtime correction that improves the lower-wall skin friction coefficient (Cf) RMSE by 7.89% against direct numerical simulation (DNS) on a periodic hill at a Reynolds number of 5600.

Results
AI CFD Scientist demonstrates significant improvements over existing baselines, such as ARIS and DeepScientist, which execute partial CFD workflows but lack the necessary domain-specific validity checks. Under matched LLM costs, the AI CFD Scientist’s performance is quantified by a 7.89% reduction in RMSE for the discovered runtime correction. Additionally, a controlled ablation study reveals that the vision-language gate successfully detects 14 out of 16 silent failures that traditional solver-level checks miss, underscoring its effectiveness in ensuring scientific rigor.

Limitations
The authors acknowledge that the framework’s reliance on a specific LLM backbone may limit its generalizability across different domains or types of simulations. Furthermore, while the vision-language gate shows promise, its performance may vary with different flow scenarios or physical models not covered in the study. The paper does not address the computational overhead introduced by the verification process, which could impact scalability in larger simulations.

Why it matters
The AI CFD Scientist represents a significant advancement in the integration of AI with physical simulation, providing a robust framework for autonomous scientific discovery in CFD. By incorporating a vision-language verification mechanism, it enhances the reliability of results, making them defensible for scientific claims. This work has implications for future research in automated scientific workflows, potentially influencing how AI can be applied to other domains requiring high-fidelity simulations and rigorous validation processes.

Authors: Nithin Somasekharan, Rabi Pathak, Manushri Dhanakoti, Tingwen Zhang, Ling Yue, Andy Zhu, Shaowu Pan
Source: arXiv:2605.06607
URL: https://arxiv.org/abs/2605.06607v1