Source-linked AI summary
OpenFOAMGPT 2.0: end-to-end, trustworthy automation for computational fluid dynamics
Jingsen Feng, Ran Xu, Xu Chu
TL;DR
CFD workflows are computationally complex, expertise-intensive, and labor-intensive, limiting access to simulations. OpenFOAMGPT 2.0 coordinates specialized agents to turn natural-language requests into end-to-end CFD simulations and post-processing; the authors report 100% reproducibility across test cases, including demanding parametric studies.
Problem
CFD simulations involve computational complexity, domain expertise, and labor-intensive workflows that limit accessibility.
Source: §1 · p. 1Method
Four specialized agents coordinate pre-processing, prompt generation, simulation execution, and post-processing in an end-to-end CFD workflow.
Source: §2 · p. 2Results
100% reproducibility across all test cases, including demanding multi-case parametric studies.
Source: §3.6 · p. 9Takeaways & Limitations
Abstract
from arXiv · showhide
We propose the first multi agent framework for computational fluid dynamics that enables fully automated, end to end simulations directly from natural language queries. The approach integrates four specialized agents Pre processing, Prompt Generation, OpenFOAMGPT (simulator), and Post processing decomposing complex computational fluid dynamics workflows into collaborative components powered by large language models. Extensive validation through diverse case studies, including Poiseuille flows, single and multi phase porous media flows, and aerodynamic analyses, demonstrates 100% success and reproducibility rates across over 450 simulations. Rigorous trustworthiness verification confirms that properly designed multi agent systems can achieve the reliability standards necessary for zero tolerance scientific computing applications while significantly lowering entry barriers. The framework establishes a foundation for conversation-driven simulation workflows in computational science, potentially accelerating discovery and innovation through more accessible tools for complex numerical simulations. Results reveal that multi-agent architectures, when properly specialized and orchestrated, can effectively handle the stringent requirements of computational physics while maintaining the intuitive interface of natural language interaction.
1 Introduction
CFD simulations are important across scientific and engineering fields, but their complexity, expertise demands, and labor-intensive workflows limit accessibility. The paper presents an end-to-end multi-agent framework that turns natural-language queries into executable simulations and analyzed results.
1 Introduction
- CFD workflows require specialized expertise and substantial labor, limiting their accessibility despite applications across many fields.The cited applications include aerospace engineering, weather prediction, biomedical flows, and Geo-Engineering.Source: §1 · p. 1
- LLM integration with CFD must address precise numerical specifications, complex geometry, sophisticated physics modeling, and result interpretation.Source: §1 · p. 1
- The framework builds on OpenFOAMGPT to convert natural-language queries into executable CFD simulations with visualization and analysis.The authors describe its design as end-to-end, with specialized agents and orchestration intended to maintain trustworthiness and reproducibility.Source: §1 · p. 2
- The authors report that conversation-to-simulation workflows reduce CFD technical barriers while increasing productivity for domain experts.They also report experimental validation across diverse flow scenarios.Source: §1 · p. 2
2 Multi-Agent LLM framework for CFD
The framework divides end-to-end CFD automation among specialized agents, built-in modules, and workflow information. Its agents formalize simulation requirements, generate case-specific prompts, execute and refine simulations, then produce visualizations and analyses.
2 Multi-Agent LLM framework for CFD
- Four specialized agents collaboratively orchestrate CFD workflows from pre-processing through simulation and post-processing without human intervention.Figure 1 distinguishes interfaces, LLM-based agents, built-in code modules, workflow information, and text logs by color.Source: §2 · p. 2, Figure 1 · p. 3
- Pre-processing classifies single- versus multi-case scope and selects blockMesh for simple geometries or snappyHexMesh for complex geometries.It identifies parametric variables when relevant and defaults to single-case, blockMesh setups for computational efficiency in simple geometries.Source: §2 · p. 2
- Prompt Generation uses the query directly for a single case and decomposes parametric variables into case-specific prompts for multi-case studies.The variables can include mesh resolution, physical properties, and boundary conditions.Source: §2 · p. 2
- Fresh prompts for each parametric case avoid incremental edits to interconnected OpenFOAM configuration files, which the authors describe as error-prone and inefficient.The prompts are organized in a Prompt Pool for sequential parametric studies.Source: §2 · p. 3
- OpenFOAMGPT generates case configurations, runs simulations in a standardized OpenFOAM environment, and uses error logs to refine failed configurations iteratively.Configuration generation uses temperature 0 to produce deterministic outputs; execution uses an OpenFOAM v2406 Docker container.Source: §2 · pp. 3–4
- The Post-processing Agent matches simulation outputs to requested analyses and generates Python visualizations and ParaView VTK files.Outputs can include field contours, streamlines, vectors, forces, moments, and comparative plots for parametric studies.Source: §2 · p. 4
3 Experiments
Five representative case studies tested the workflow across single- and multi-phase channel and porous-media flows and motorbike aerodynamics. Under well-formulated queries and carefully designed system prompts, the framework completed simulations and parametric studies reproducibly.
3 Experiments
- 100% success across single-case and parametric simulations, with 100% reproducibility across all test cases.The validation covered five representative CFD applications, from Poiseuille flows to porous media and motorbike aerodynamics.Source: §3 · p. 4, §3.6 · p. 9
3.1 Single-phase Poiseuille flow
- The single-phase Poiseuille simulation’s numerical velocity profile showed excellent agreement with the analytical solution.The post-processing agent made the comparison without additional guidance.Source: §3.1 · p. 5, Figure 2 · p. 6
3.2 Multi-phase Poiseuille flow
- Three parametric series completed successfully, producing visualizations of grid convergence and saturation and viscosity-ratio effects in multiphase Poiseuille flow.The studies varied mesh resolution, wetting-phase saturation, and viscosity ratio.Source: §3.2 · p. 5, Figure 2 · p. 6
3.3 Single-phase flow in porous media
- For porous flow, 600 × 300 cells differed by less than 2% from 700 × 350; the permeability REV was 3 × 1.5 mm, and higher gradients departed from Darcy linearity.The workflow also handled interconnected geometry, meshing, and flow-regime studies.Source: §3.3 · p. 6, Figure 3 · p. 7
3.4 Multi-phase flow in porous media
- The workflow completed all three porous-media drainage studies, handling geometry, interface physics, and variations in contact angle, inlet velocity, and viscosity ratio.It automatically generated visualizations and identified displacement efficiency.Source: §3.4 · p. 7, Figure 4 · p. 8
3.5 Aerodynamics of a motorbike
- For motorbike aerodynamics, the workflow extracted drag coefficients across velocities, visualized flow at 100 m/s, and accepted a provided mesh without geometry preprocessing.The visualization shows helmet wake vortices, fairing boundary-layer development, and rear-wheel turbulence.Source: §3.5 · pp. 7–8, Figure 5 · p. 8
3.6 Trustworthiness verification
- Trustworthiness testing included ten repetitions of an eight-case porous-flow sequence and two 100-simulation experiments, one with pressure differences varied from 1 Pa to 10 Pa.The varied-pressure runs are shown in Figure 6.Source: §3.6 · pp. 8–9, Figure 6 · p. 9
4 Conclusion
OpenFOAMGPT 2.0 turns natural-language queries into end-to-end CFD results with visualization and analysis through four specialized agents. The paper reports 100% reproducibility and describes extensions to other precision-demanding computational domains and future work on more complex industry challenges.
4 Conclusion
- Four specialized agents handle pre-processing, prompt generation, simulation, and post-processing to turn natural-language queries into end-to-end CFD results.The system includes visualization and analysis of results.Source: §4 · p. 9
- The framework handles diverse fluid dynamics problems with 100% reproducibility and is presented as meeting reliability standards for scientific computing.The paper also characterizes its approach as maintaining numerical rigor and physical accuracy while reducing expertise barriers.Source: §4 · p. 9
- The architectural principle is described as readily extendable to other computational domains requiring comparable precision, with future work targeting more complex industry-level challenges.Source: §4 · p. 9