A catalogue of design patterns for using LLMs in systems engineering
What happened
On 16 October 2026 Systems Engineering, the research journal of the International Council on Systems Engineering (INCOSE), published by Wiley, put out an open-access Early View paper: "A Design Pattern Catalog for Applying Large Language Models in the Systems Engineering Lifecycle". It comes from the Information Technology Group at Wageningen University & Research in the Netherlands.
The paper sets out a catalogue of design patterns for using large language models (LLMs) in systems engineering work. Each pattern records a recurring problem, its context, an LLM-based solution, an implementation structure and the consequences seen in use. The paper says it was inspired by the software design patterns of Gamma and colleagues and by architectural patterns in model-based systems engineering (MBSE).
The patterns cover conceptual design, requirements engineering, system design and architecture, verification and validation, operations and maintenance, and support that runs across the whole life cycle. The paper also deals with concerns that cut across every stage: traceability, keeping a human in the loop, trustworthiness, prompt engineering and fitting the patterns into the life cycle. It concentrates on techniques used while a model is in operation, such as prompt engineering and retrieval-augmented generation (RAG), rather than retraining or adapting the model itself.
The paper says the patterns came from a literature analysis, exploratory experiments with LLM-supported tasks, and case studies in three domains: physical protection systems, space systems, and command and control systems. The tasks covered were generating requirements, summarising, recording design rationale, developing test cases and maintaining traceability.
Why it matters
The paper's own starting point is that LLM use in systems engineering is fragmented, ad hoc and hard to reproduce, and that there has been no systematic way to capture and reuse what works. It states that it introduces no new machine learning algorithms. Its contribution is to give emerging practice a structure tied to life cycle activities.
The paper is open about the limits of its evidence. It lists five threats to validity:
- the case studies were prospective, on systems still in development, which may limit confidence in long-term effects in operation;
- the sample of domains was relatively small;
- the people reviewing the patterns knew LLMs well, which may have introduced confirmation bias;
- effectiveness was judged partly on users' qualitative views;
- LLM technology is changing fast, so later models may behave differently.
It also says the set of patterns is not complete, and it lists explainability, risk management and certification in mission-critical and safety-critical work as future research.