New arXiv survey maps 6 stages where AI and simulation converge

Feldkamp and colleagues' overview details how AI can support, augment, or even replace components of simulation studies. For example, LLMs can automate model specification or assist in input modeling, while simulation executors benefit from AI-driven optimizations. The authors structure the survey around the M&S lifecycle—from model specification and input modeling through execution, experimentation, verification and validation (V&V), and output analysis. At each stage, they examine selected studies that illustrate how techniques such as large language models have reshaped simulation practices. They also point out that simulations are not just consumers of AI: they act as data generators, training environments, and evaluation platforms for AI systems, creating a feedback loop that accelerates both fields.

The report highlights the growing availability of data and computational resources, coupled with the rise of generative AI, as key drivers of this convergence. Despite the promise, the authors note significant limitations: LLMs can introduce biases, make errors in complex simulation logic, and require careful validation against traditional simulation frameworks. The survey also identifies open challenges, such as maintaining traceability and reproducibility when AI is embedded in M&S workflows. By providing a structured overview and a conceptual roadmap, Feldkamp and colleagues aim to help researchers and practitioners navigate a rapidly changing ecosystem, making this an essential reference for anyone working at the intersection of AI and simulation.