Trustworthy generation of SysML v2 models for physical systems: a multi-agent approach

Abstract

Model-Based Systems Engineering (MBSE) increasingly depends on architectural models of physical and engineered systems that are both syntactically valid and semantically consistent with their originating requirements. Authoring such models by hand in SysML v2 is slow and demands scarce expertise, a recognised “modelling bottleneck”. Large Language Models (LLMs) can draft SysML v2 from heterogeneous inputs, natural-language requirements, voice, and even simulation models, but their probabilistic nature yields frequent violations of the language’s formal ANTLR grammar and silent semantic inconsistencies, making outputs untrustworthy for simulation or verification. This thesis investigates a multi-agent, LLM-based methodology that couples constrained generation with a layered validation authority (grammar-grounded syntactic repair and standards-aligned semantic judgement) and connects simulation-oriented physical models (Modelica/Simscape) to SysML v2 architecture. This paper states the problem and research questions, frames the expected contributions as Design Science artefacts, summarises the planned evaluation across the electronic and aerospace domains, and reports an already substantial set of results, including published and in-press journal contributions and several peer-reviewed papers, together with the timeline to completion in 2029.