AI finds simulation models via natural language queries in new study
Open-source embeddings match proprietary models for simulation model discovery with recall@5 metrics
An experimental study investigated how AI retrieval methods can discover simulation models using natural language queries. Published on arXiv:2606.30846 and accepted for WSC 2026, the work evaluated data representation, transformer-based embeddings (including open-source models), and reranking strategies. Key results: data representation significantly impacts retrieval, open-source embeddings achieve high performance, and reranking boosts accuracy for complex queries—providing a baseline for AI-driven model discovery and advancing toward composability and interoperability in M&S.
- Open-source transformer embeddings (e.g., Sentence-BERT) achieved comparable recall@5 to proprietary models on simulation model retrieval tasks
- Reranking strategies improved nDCG@5 by up to 20% for complex multi-condition queries
- Data representation significantly impacts performance—raw text descriptions outperformed structured metadata for model discovery
Why It Matters
Enables researchers to find and reuse simulation models via plain English, accelerating M&S workflows and promoting composability.