Automating CFD Simulations with Local Large Language Models

Sammanfattning

This thesis develops a locally deployed LLM-assisted system for automating CFD simulations from natural-language physical modelling tasks, including OpenFOAM case setup generation, solver execution, error diagnosis, and iterative correction. Unlike many recent LLM-driven CFD agents that mainly rely on frontier cloud models, this work investigates whether open-weight local LLMs with stronger data privacy, internet independence, and economic attractiveness can support practical CFD automation. The proposed framework combines a knowledge retrieval layer constructed from OpenFOAM tutorials and documentation with an agentic workflow consisting of five functional roles. To improve reliability under the relatively weaker reasoning, limited knowledge capacity, and less stable instruction-following ability of local LLMs, a series of robustness-oriented strategies such as hybrid retrieval, constrained prompts, cross-file dependency support, and self-correction loops are incorporated, shifting part of the reasoning burden from the model to the system design. The system is evaluated on representative and industrially motivated physical modelling tasks covering laminar incompressible flow, buoyancy-driven heat transfer, wall-bounded internal turbulence, and external aerodynamics over a Volvo EX90 geometry with multiple STL surfaces. The results show that locally deployed LLMs can generate executable and physically meaningful OpenFOAM cases with system-level support and iterative refinement. The study demonstrates the practical feasibility of local LLM-assisted CFD automation, while also indicating future directions in quantitative validation and CFD accuracy control for production-grade applications.