Abstract for: Exploring Black-Box Parameter Calibration Techniques in System Dynamics Models

System dynamics models often involve nonlinear feedback structures and complex interactions, making parameter calibration difficult to approach through equation-based methods. Black-box optimization techniques provide an alternative by treating the model as an input-output system, allowing calibration without requiring access to the underlying equations. This raises the question of how effectively these methods can preserve the dynamic behavior of the model. This study evaluates several black-box optimization techniques for parameter calibration, including Grid Search, Random Search, Tree-structured Parzen Estimator (TPE), Gaussian Process (GP), and CMA-ES. Two generic system dynamics models—a SEIR epidemiological model and a stock management model—are used with synthetic datasets generated under controlled noise conditions. The methods are compared in terms of numerical accuracy, parameter recovery, computational cost, and their ability to reproduce characteristic dynamic behaviors. The results show that black-box calibration techniques can be effectively applied to the selected system dynamics models, although their numerical performance varies across methods and experimental settings. Differences also emerge in their ability to reproduce the underlying system dynamics, particularly in peak magnitudes and oscillatory patterns. The findings indicate that calibration performance depends on model structure, noise conditions in observed data, and data availability. In some cases, numerical accuracy alone is not sufficient to ensure correct dynamic behavior. These findings highlight an important limitation of relying solely on statistical error metrics in system dynamics calibration. A parameter set that minimizes numerical error may still fail to reproduce the true dynamic behavior of the system. This study emphasizes the importance of evaluating calibration results from both numerical and behavioral perspectives. Incorporating dynamic pattern validation alongside traditional error measures can lead to more reliable and interpretable calibration outcomes in system dynamics models. Grammatical refining