Study proposes AI-driven method to discover reusable simulation models via natural language queries

A new arXiv preprint presents an experimental study on using AI to discover reusable simulation models from natural language queries. The work, titled 'How Can AI Find My Model? A Model-Finding Experimental Study Considering Data Formats, Embeddings, and Retrieval Strategies,' evaluates how data representation, transformer-based embeddings, and retrieval strategies affect the accuracy of model retrieval.

The authors report that data representation choices and reranking methods significantly influence performance, particularly as query complexity increases. They find that open-source embedding models can achieve high performance on standard information retrieval metrics, including recall@5 and nDCG@5.

The study positions its findings as a baseline for AI-driven model discovery, emphasizing the role of such systems in advancing toward AI-driven composability and interoperability in modeling and simulation. The paper has been accepted for publication in the Proceedings of the 2026 Winter Simulation Conference (WSC 2026).

The research team includes authors from Old Dominion University and other institutions, and the work is supported by major funders as acknowledged in the preprint. The authors propose that their approach could help practitioners more efficiently locate and reuse models, reducing duplication and improving workflow integration.