Abstract for: From Data to Model Structure: A Generative Algorithm to develop System Dynamics Models

Developing simulation models, including System Dynamics (SD) models, remains largely a manual process that involves specifying equations to capture the dynamic behaviour of the system under study. While artificial intelligence (AI) applications are increasingly used for model calibration and analysis, their role in model development remains unexplored. We introduce a generative algorithm designed to derive system structures from observed behaviours. The algorithm constructs net flow equations based on time series data of stocks in the system. We describe the algorithm's operations and discuss its potential as a model development support tool within the SD modelling process. To evaluate its efficacy, we apply it to three test cases of increasing complexity: a simple goal-seeking system represented by a cooling teacup, an oscillatory predator-prey model, and a complex model involving interactions between workforce, production, and inventory management. The algorithm successfully identifies structures in simpler cases but achieves only partial success in the complex case. This exploration shows both the potential and limitations of using generative algorithms to support SD model development. We concludes with a discussion of the challenges encountered and reflects on the future of modeller-algorithm collaboration, setting the stage for further research on ‘human-in-the-loop’ AI support tools. The method itself uses AI and as non-native speakers we use AI to check our work for language errors.