Preprint reports fewer simulator runs with AI-assisted calibration

Researchers report in a preprint that a workflow combining a language model with a conventional optimizer reached target calibration accuracy in 9 simulator runs, compared with 12 for the language model alone, 17 for Nelder-Mead and more than 60 for Bayesian optimization. Calibration adjusts simulator parameters to match reference data. The study tested a cylinder-wake simulation and a damped-oscillator problem. The language model used descriptions of physical errors to suggest initial parameter changes, then passed accumulated evaluations to a conventional optimizer. Its early advantage disappeared by fifteen evaluations, and variability in language-model outputs produced a 9.8-fold performance spread.