关键词: OpenFOAM, 多模态大模型, 知识图谱, 多智能体系统, 代码生成

Abstract: Addressing the dual challenges of OpenFOAM?s high learning curve due to its reliance on command-line operations and the generally long development cycles and high customization costs of traditional simulation interfaces, this paper proposes the AutoCode4OF framework, which aims to achieve end-to-end automatic generation of a complete OpenFOAM executable interface code from multimodal inputs. The main innovations of the framework include: (1) At the input level, it integrates multimodal information such as images, natural language text, and existing code snippets to jointly represent user intent; (2) In terms of knowledge processing, by constructing a professional knowledge graph in the field of computational fluid dynamics (CFD) and introducing retrieval-augmented generation (RAG) along with dual front-end and back-end validation mechanisms, it significantly enhances the physical rationality and reliability of the generated code; (3) In system architecture design, it adopts a multi-agent collaborative working mechanism, decomposing the overall task into specialized modules such as knowledge retrieval, task planning, code generation, material setup, and testing verification, with each module collaborating to ensure the quality and completeness of the output. Experimental results show that AutoCode4OF achieves scores of 0.956, 0.997, and 100% in code quality, functional completeness, and compilation success rate, respectively. It demonstrates strong applicability and stability across various stages, including mesh generation, solution computation, and post-processing result validation, highlighting its high practical engineering application value and providing new insights for the intelligent development of scientific computing software.

Key words: OpenFOAM, multimodal large language model, knowledge graph, multi-agent systems, code generation