TLDR: This paper introduces a prompt-driven approach using Large Language Models (LLMs) and SysML v2 to semantically align and integrate system models in collaborative Model-Based Systems Engineering (MBSE). It proposes a staged process with human verification, leveraging SysML v2’s ‘alias’ and ‘import’ constructs for “soft alignment” that preserves original model structures. The method aims to improve efficiency, accuracy, and traceability in cross-organizational system development, demonstrating its feasibility through examples while acknowledging limitations in deep semantic understanding and prompt reusability.
In the complex world of modern system development, multiple organizations often collaborate to build intricate systems. This collaboration, especially in Model-Based Systems Engineering (MBSE), frequently encounters a significant hurdle: ensuring that independently developed system models speak the same language, semantically speaking. This challenge, known as semantic misalignment, can lead to inconsistencies and difficulties in integrating different parts of a system.
A recent research paper explores an innovative solution to this problem by combining the power of SysML v2, a new generation of system modeling language, with Large Language Models (LLMs) like GPT. The paper, titled “LLM-Assisted Semantic Alignment and Integration in Collaborative Model-Based Systems Engineering Using SysML v2,” proposes a structured, prompt-driven approach to help align these diverse system models.
The core idea is to leverage SysML v2’s enhanced structural modularity and formal semantics, which provide a stronger foundation for interoperable modeling. Simultaneously, GPT-based LLMs offer new capabilities for understanding and integrating model information. The researchers, Zirui Li, Stephan Husung, and Haoze Wang from Technische Universität Ilmenau, have developed an iterative approach that involves extracting model information, matching semantic elements, and verifying the alignment.
Instead of forcing different organizations to adopt a single, unified modeling method, which is often impractical, the paper advocates for a “soft alignment” strategy. This approach creates new alignment packages that map elements between different models while preserving their original structures. SysML v2 constructs such as ‘alias’ (for lightweight name binding), ‘import’ (for element visibility and reuse), and metadata extensions are crucial here. These tools allow for traceable and flexible integration without altering the core models.
The proposed LLM-assisted integration approach is a systematic, staged process with alternating AI support and human verification. It aims to improve efficiency, semantic accuracy, structural stability, and traceability compared to traditional manual methods.
Also Read:
- LLMSymGuard: Enhancing Language Model Safety with Interpretable Internal Concepts
- Enhancing Urban Mobility Simulations with AI: The Preference Chain Approach
Key Mechanisms of the Approach
-
Additive Modeling: LLMs generate new content in a separate package, referencing original model elements via SysML v2’s ‘import’ mechanism, thus preventing changes to the original model structure.
-
Staged Process: The alignment is broken down into structured sequences, with LLM guidance and human checkpoints at each stage to ensure accuracy.
-
Confidence-Scored Mapping Suggestions: Each suggested model mapping comes with a confidence score from the LLM, helping engineers interpret the results.
-
Mapping Verification: All suggested mappings are checked against SysML syntax and semantic documentation.
-
Coverage Check: The LLM verifies that all model elements and previous stage outputs have been processed correctly, ensuring traceability.
-
Standardized Output Format: Using formats like JSON ensures structural consistency in LLM outputs, simplifying subsequent processing.
-
Structured Comment Support: The generated SysML v2 model includes comments explaining the rationale behind alignment decisions, improving user understanding.
The process is divided into seven stages, starting from preparation and syntax confirmation, moving through model element summarization, match candidate suggestion, mapping verification, aligned package generation, consistency checks, and finally, export and documentation. Each stage requires explicit user confirmation, allowing for iterative optimization if needed.
An interesting aspect highlighted in the paper is the prompt-driven realization and verification. During initial experiments, the LLM sometimes misinterpreted user-provided semantic extension libraries. However, with minimal human intervention and clear prompt clarification, the LLM quickly learned and consistently applied the correct usage patterns, demonstrating its ability to generalize from targeted corrections.
While this approach shows significant promise in enhancing structural and semantic transparency, the authors also discuss its limitations. Current methods primarily focus on naming-level consistency and need further development to address deeper semantic discrepancies, such as differences in interface granularity or design intent. Future work will also explore creating stable, reusable prompt modules, integrating ontologies for more grounded outputs, and improving feedback mechanisms for better user trust and traceability.
In conclusion, this research provides a foundational process for LLM-assisted SysML v2 multi-source model collaboration and semantic alignment. By combining structured prompt engineering, SysML v2 constructs, and human-in-the-loop interaction, it offers a pathway to more efficient and reliable integration in collaborative MBSE environments. You can find the full research paper here.
