LLM-Assisted Semantic Alignment and Integration in Collaborative Model-Based Systems Engineering Using SysML v2

Cross-organizational collaboration in Model-Based Systems Engineering (MBSE) faces many challenges in achieving semantic alignment across independently developed system models. SysML v2 introduces enhanced structural modularity and formal semantics, offering a stronger foundation for interoperable modeling. Meanwhile, GPT-based Large Language Models (LLMs) provide new capabilities for assisting model understanding and integration. This paper proposes a structured, prompt-driven approach for LLM-assisted semantic alignment of SysML v2 models. The core contribution lies in the iterative development of an alignment approach and interaction prompts, incorporating model extraction, semantic matching, and verification. The approach leverages SysML v2 constructs such as alias, import, and metadata extensions to support traceable, soft alignment integration. It is demonstrated with a GPT-based LLM through an example of a measurement system. Benefits and limitations are discussed.

💡 Research Summary

The paper addresses a fundamental obstacle in modern Model‑Based Systems Engineering (MBSE): when multiple organizations (e.g., an OEM and its suppliers) develop subsystem models independently, structural heterogeneity and semantic misalignment hinder holistic analysis and integration. While SysML v2 introduces a richer textual syntax, explicit alias/import mechanisms, and extensible metadata that provide a formally defined semantic core, these features alone do not solve the coordination problem. At the same time, large language models (LLMs) such as GPT‑4o have demonstrated impressive natural‑language understanding and generation capabilities, suggesting they could mediate between informal documentation and formal model artifacts. However, LLMs suffer from non‑deterministic outputs, hallucinations, and a lack of traceability—issues that are unacceptable in rigorous engineering processes.

To bridge this gap, the authors formulate two research questions: (1) how can LLM technology be combined with SysML v2 constructs to enable semantic alignment and integration of cross‑organizational MBSE models? (2) how can a structured alignment approach be designed to support company‑specific semantic extensions, thereby allowing the LLM to perform context‑aware model understanding and integration? The paper builds on prior work that identified three integration concepts—unified modeling, transformation‑based integration, and soft alignment—and argues that soft alignment is the most compatible with LLM assistance because it preserves each partner’s original model while mapping relevant elements through lightweight extension packages.

The core contribution is a systematic, prompt‑driven workflow that couples LLM assistance with human verification at every stage. The workflow consists of four iterative phases:

1. Model Extraction – The SysML v2 model is exported to a machine‑readable JSON representation that retains alias and import information, enabling downstream processing without altering the source model.
2. Semantic Matching – The LLM receives the extracted elements together with domain documents (requirements, design descriptions) and a user‑defined extension library that encodes domain‑specific alignment concepts (e.g., allocation types). Using natural‑language similarity and the extension semantics, the LLM proposes candidate mappings.
3. Verification & Annotation – Each mapping is automatically annotated with a confidence score and a textual rationale. Human engineers review, correct, or approve the suggestions, thereby injecting expert judgment and ensuring traceability.
4. Additive Integration – Approved mappings are materialized as a new SysML v2 package that references original elements via the import mechanism. This “soft alignment” preserves the original models, supports incremental updates, and produces both JSON and human‑readable outputs.

A key methodological innovation is the use of an agile test‑process and Design Research Methodology to iteratively refine the prompts. Early experiments with a monolithic prompt produced inconsistent outputs and no traceability. By decomposing the interaction into staged prompts, inserting verification checkpoints, and embedding confidence scores and rationale into the model metadata, the authors achieved markedly higher stability and reproducibility across repeated runs. The approach was demonstrated on a measurement‑system case study involving an OEM and a supplier model, showing that the LLM could correctly align sensor definitions, data‑flow relationships, and calibration requirements when guided by the structured workflow.

The paper also discusses limitations. The evaluation is confined to a small set of models, and the creation of the domain‑specific extension library requires substantial upfront expertise. The approach’s scalability to large, highly complex models remains untested, and changes in LLM versions may necessitate prompt re‑engineering. Future work is proposed to extend the framework to broader domains, integrate formal verification tools (e.g., Gamma, Imandra) for automated consistency checks, and explore automated confidence calibration techniques.

In conclusion, the study demonstrates that the combination of SysML v2’s formal semantics with LLM‑driven natural‑language processing, when organized into a disciplined, traceable, and human‑in‑the‑loop workflow, can substantially reduce the manual effort required for semantic alignment in collaborative MBSE. The soft‑alignment strategy preserves each organization’s autonomy while enabling incremental, auditable integration, offering a promising pathway toward AI‑augmented systems engineering practice.