Enhancing Model‐Based Systems Engineering with Large Language Models

Abstract Model‐Based Systems Engineering (MBSE) is a cornerstone of modern systems engineering, enabling the design and management of increasingly complex systems. However, creating models with modeling languages such as System Modeling Language (SysML) V2 remains a highly manual and time‐intensive process. As system complexity increases and interdisciplinary collaboration becomes critical, streamlining workflows and accelerating the modeling process are more important than ever. This paper investigates the use of two classes of Large Language Models (LLMs)— the general‐purpose reasoning model, GPT‐4, and the domain specific fine‐tuned model, CodeT5, to assist in generating SysML V2 models from natural language descriptions. Our findings reveal that while LLMs can generate initial “skeleton” models, relying solely on AI for entire model generation is not advisable due to critical errors and inconsistencies in the generated outputs. Instead, LLMs serve as effective tools for accelerating the modeling process by producing foundational prototypes that can be refined and improved through human intervention. With fine‐tuning techniques for CodeT5 and prompt engineering for GPT‐4, we evaluate their performance across key dimensions of syntax accuracy, semantic alignment, and structural correctness. Using a diverse set of metrics and statistical analyses, we identify the trade‐offs between task‐specific optimization (CodeT5) and broader contextual reasoning (GPT‐4). The potential of LLMs as accelerators in MBSE workflows cannot be ignored with the “early” modeling capabilities they present. This work contributes to the growing intersection of AI and MBSE, offering practical insights into the selection and application of LLMs in systems engineering and paving the way for future advancements in the field.