CFD Mesh-Gen AI: Natural Language to OpenFOAM Simulation

This model, developed by Arun Govind Neelan, is a specialized 32.8 billion parameter Qwen2.5-Coder-32B LLM fine-tuned with QLoRA to automate the entire Computational Fluid Dynamics (CFD) workflow from natural language input. It processes geometry descriptions to generate meshes and subsequently runs OpenFOAM simulations, providing visualized results.

Key Capabilities

• End-to-End Pipeline: Converts natural language geometry descriptions into Gmsh meshes, then executes OpenFOAM CFD simulations, and finally visualizes the results as PNGs.
• Two-Phase Architecture: Utilizes a "Mesh Agent" (fine-tuned Qwen2.5-Coder-32B) for geometry description to Gmsh script conversion and a "Simulation Agent" (OpenFOAM 11) for CFD execution.
• Hybrid Mesh Generation: Employs the LLM for complex geometries and a deterministic fast-path for common shapes (e.g., backward-facing step, NACA airfoils, cylinders) to ensure accuracy and efficiency.
• Self-Correction: Includes a loop that feeds Gmsh error messages back to the LLM for automatic syntax correction, improving mesh generation reliability.
• Automated Simulation Setup: Automatically assigns boundary conditions, detects 2D/3D cases, and extracts flow parameters from prompts for OpenFOAM execution.
• API Endpoints: Provides a FastAPI server with endpoints for mesh generation, simulation, and result download, facilitating integration into other systems.

Good for

• Automated CFD Pre-processing: Ideal for researchers and engineers looking to quickly generate meshes and set up CFD simulations from simple text prompts.
• Rapid Prototyping: Enables fast iteration on design concepts by streamlining the geometry-to-simulation process.
• Educational Purposes: Can serve as a tool for demonstrating the CFD workflow and the integration of LLMs in scientific computing.
• Specific Geometry Generation: Excels at generating meshes for common CFD geometries like pipes, cavities, airfoils, and steps, with support for flow conditions.