Surrogate-Assisted Global Calibration of a Three-Dimensional Hydrodynamic Lake Model Using Machine Learning

Calibration of hydrodynamic models is a cumbersome process, whereby model coefficients are manually and/or systematically adjusted to minimize model error metrics relative to observations. Manual adjustment may leave models under calibrated or introduce subjectivity in parameter selection, while systematic calibration across a large parameter space may be computationally infeasible. To address these issues, we developed a Python-based automatic calibration framework that combines quasi-random Sobol sampling to explore the parameter space, eXtreme Gradient Boosting (XGBoost) as a machine learning surrogate to approximate calibration error, and a Genetic Algorithm to identify candidate optimal parameter sets, which are then evaluated using the full hydrodynamic model. The framework was applied to a three-dimensional Aquatic Ecosystem Model (AEM3D) hindcast simulation of Lake Erie from 2017 to 2019 to minimize the root-mean-square (RMSE) in water temperature, current velocity, and water level relative to observations. Through optimized adjustment of seven model parameters, RMSE improvements of 11%, 13%, and 73% were achieved, respectively, using a limited number of calibration runs. The framework also provides information on the sensitivity of the simulated output variables to variations in calibration parameters. This computationally efficient calibration approach reduces the number of expensive model simulations required for calibration, enables practical optimization of three-dimensional hydrodynamic models and is suitable for incorporation into hindcast, real-time and forecasting workflows.