AI Agent Framework Translates Legacy Fortran Code to Devito

A new approach for AI agents merges Retrieval-Augmented Generation (RAG) with open-source Large Language Models to break down older finite-difference code and adapt it for the Devito framework. This method employs a LangGraph architecture with multiple iterative stages. It creates a knowledge graph for Devito using techniques like document parsing, extracting relationships, structure-aware segmentation, and community detection based on Leiden. The GraphRAG optimization enhances query efficiency across different semantic areas, such as seismic wave simulation and computational fluid dynamics. Moreover, a reverse engineering component develops three-tier query strategies for RAG retrieval by analyzing Fortran source code. You can find more about this research in arXiv:2601.18381.