LLMs Calibrate Grey-Box Simulation Models More Efficiently

Calibrating grey-box simulation models, which are models where the underlying code is known but the combined effect of parameters is hard to predict, is a complex optimization challenge. It often involves expensive model evaluations, high-dimensional parameter spaces, and the need to respect specific plausibility constraints. Traditional optimization methods like Nelder-Mead can be inefficient, while advanced Bayesian Optimization (BO) methods, though more efficient, require significant setup for constraint handling. This research introduces an innovative "agentic calibration" method that leverages large language models (LLMs) to act as the optimizer. A key advantage is how constraints are managed: they are simply specified in plain language within the LLM's system prompt, eliminating the need for complex modeling machinery. Evaluations on an anal cancer simulation model, under both unconstrained and clinically constrained scenarios, showed compelling results. The agentic method achieved lower errors with fewer model evaluations compared to both Nelder-Mead and Bayesian Optimization in unconstrained settings. Under constrained conditions, it matched BO's performance while significantly outperforming Nelder-Mead. Although the per-iteration inference time for the LLM is higher, the substantial reduction in expensive simulation evaluations makes this approach highly suitable when simulation runtime is the dominant cost. Furthermore, the LLM's iterative rationale provides an auditable and explainable search process, enhancing transparency.