02 - Systems Engineering and Integration
AUTOMATIC GENERATION OF PHYSICAL SYSTEM MODELS FOR AIRCRAFT SYSTEMS BASED ON LLMS
Y.L. Ren¹, L.F. You¹, X.H. Gao¹, Z.Q. Wang¹, C.Y. Chang¹, Z.Q. Guo¹; ¹Beihang University, China
Modelica is an object-oriented multi-domain unified modeling language widely applied in the simulation of Cyber-Physical Systems (CPS) and Model-Based Systems Engineering (MBSE).Although Large Language Models (LLMs) have enabled code generation,the automated modeling of Modelica remains a major challenge due to the scarcity of domain-specific training corpora and the high dependency on professional expertise.This paper presents a domain-specific LLM framework tailored for Unmanned Aerial Vehicle (UAV) Modelica modeling.Retrieval-Augmented Generation (RAG) technology is leveraged to construct a specialized knowledge base by performing structured extraction, intelligent chunking, and vectorization of UAV domain materials and Modelica specifications. This knowledge base is validated through recall testing.The LLM is integrated into an AI-Agent architecture, which utilizes API interfaces to access the knowledge base and perform adaptation.Then,specialized prompt templates are designed, and systematic transformation rules and constraints are established to convert modeling requirements described in natural language into JSON format expression and ultimately into high-fidelity Modelica code, thereby realizing the conversion process from unstructured text to high-fidelity models.The application in UAV system modeling scenarios demonstrates that the proposed system effectively comprehends modeling intent and automatically generates syntactically correct and executable domThe main contributions of this work are threefold. First, the proposed structured knowledge framework ensures unified domain organization, significantly enhancing knowledge retrievability and modeling consistency. Second, the RAG-enhanced generation mechanism fuses semantic matching with precise querying, thereby improving code accuracy and professional quality. Third, the collaborative AI-Agent architecture enables automated multi-model specialization, effectively optimizing system intelligence and modelin
DOI: 10.71945/icas2026_0479