AutoFOAM: Self-Refining Autonomous Agent for OpenFOAM Simulations

A recent paper published on arXiv (2608.00003) presents AutoFOAM, an innovative large language model (LLM) agent that automates Computational Fluid Dynamics (CFD) simulations through the open-source solver OpenFOAM. This agent interprets natural-language commands to autonomously create, assess, execute, and refine simulations without the need for manual setup. Initially trained on Qwen-coder 2.5-14B, it undergoes fine-tuning with 252 text prompts related to 7 OpenFOAM solvers, 13 parameterized mesh templates, and a y plus-aware numerical policy. Central to its design is a complex evolution loop with 7 stages. To mitigate model deterioration during self-training, the agent utilizes three anti-collapse strategies: RAG-augmented retry context, surgical dictionary-level patching, and prompt-diversity paraphrasing. This advancement seeks to alleviate the knowledge and time demands of OpenFOAM, enhancing accessibility for engineers. The authors of the paper contribute to the fields of computational engineering and AI-driven automation.