Bootstrapping SysML with LLM-based inference and integrated discrete-event simulation

Abstract

The demand for systems engineering methodologies with integrative artificial intelligence (AI) has been increasing. Model-based systems engineering offers a disciplined, structured methodology. However, it encounters difficulties with semantic interpretation and domain adaptation, especially across different contexts. In this work, we examine the potential of generative AI to address this challenge. We implement a dual approach to enrich the modeling experience by incorporating domain adaptations via large language models and executable semantics via discrete-event simulation. The result is a bootstrapped, end-to-end automated system model construction from minimal entry points, featuring built-in, generic executable capability that adheres to the simulation-to-production system principle. We will demonstrate how a user of such an approach can produce a sound, semantically rich model with advanced simulations from a minimal textual entry. We also discuss mechanisms for incorporating knowledge and expertise through a convenient yet effective human-in-the-loop integration. We demonstrate the approach through a detailed semiconductor wafer fabrication experiment and further illustrate its generality across diverse domains through extensive generative and simulation-based evaluations.