Harnessing ChatGPT for Model Transformation in Software Architecture: From UML State Diagrams to Rebeca Models for Formal Verification

Software architecture relies heavily on modeling techniques to describe, analyze, and verify system designs. The Unified Modeling Language is widely recognized as the de-facto standard for modeling various types of systems. However, UML’s lack of formal semantics poses challenges for performing formal verification, a critical step in ensuring the correctness of architectural models. Rebeca, an actor-based modeling language, is designed to enable formal verification of concurrent reactive systems. Previous efforts to bridge UML and Rebeca through model transformations have required combining multiple UML diagrams and a deep understanding of Rebeca, limiting practical applicability.In this paper, we explore the potential of leveraging large language models, specifically GPT-4, to automate the transformation of UML state diagrams into Rebeca models. Using a few-shot learning approach, we investigated the feasibility of this translation process. Initial results revealed that UML state diagrams alone were insufficient for generating accurate Rebeca models. To address this limitation, we augmented the diagrams with metadata, enabling GPT-4 to produce models that required only minor corrections to be executable in Rebeca's model-checking tool, Afra. Our findings demonstrate that LLMs hold promise in facilitating model transformations for software architecture, particularly for translating UML state diagrams into Rebeca models for formal verification. While not yet fully automated, this approach significantly reduces the effort required for transformation, paving the way for further research into the integration of LLMs into model-driven engineering practices.