Researchers tested whether large language models can handle one of systems engineering's most tedious jobs: mapping how every part of a machine depends on every other part.
The team compared a plain LLM, retrieval-augmented generation (RAG), and graph-based RAG (GraphRAG) for building Design Structure Matrices, spreadsheets that show which components in a complex system connect to, depend on, or interfere with each other. They ran the methods on two test cases with known reference architectures: a power screwdriver and a CubeSat. Each method was scored on two tasks - filling in relationships once a list of components was already given, and the harder job of first identifying the components themselves, then mapping how they connect. Performance was checked both component-by-component and against the full architecture.
Design Structure Matrices are normally built by hand by engineers who know a system inside out, which makes them slow to produce and hard to keep current as a design changes. A model that can draft a usable first pass could shift that work from blank-page assembly to error-checking, which matters most for CubeSats and other cyber-physical systems where design cycles are already tight.
The catch: the paper reports real computational and design hurdles, and the harder identification task is exactly where hallucination-prone LLMs are least trustworthy. The authors published their code, which is the right call if they want other engineers to stress-test the claims rather than take them on faith.