Abstract for: The Transformative Potential of Large Language Models for System Dynamics Modeling

The process of transforming a dynamic hypothesis into a causal loop diagram (CLD) that articulates the causal structure of a system is a key step in the development of System Dynamics (SD) models. Extracting key variables and causal relationships from text to build a CLD is an essential step in the development of feedback-rich models, and is often challenging and time-consuming for less-experienced modelers, limiting the application and use of SD tools. In this paper, we introduce and test a method for automating the translation of dynamic hypotheses (text) into CLDS using large language models (LLMs). We first describe how LLMs work, and how they can make the inferences neede to build CLDS using a standard digraph structure. We next develop a set of simple dynamic hypotheses and corresponding CLDS from leading SD textbooks and training courses for LLM prompting and testing, and define metrics for the comparison of automated CLDS to the test set. We then perform a series of preliminary tests using this method, demonstrating that, for simple model structures, LLMs are able to generate CLDS of a similar quality to CLDs built by leading SD experts. Finally, we discuss opportunities and challenges extending this approach to more complex dynamic structures, and, eventually, to simulation models.