Advancing continental-scale hydrology model calibration using large-sample emulators: from simple conceptual to complex process-based land models

Land/hydrologic model advances have significantly enhanced the capability to simulate complex hydrological processes. However, the accuracy of these simulations is often undermined by uncertainties in model parameters, many of which are poorly constrained by observations, as well as the high computation demand of sophisticated models, which restricts their optimization. To address these challenges, we developed a machine learning (ML)-based calibration approach using large-sample emulators (LSEs) to optimize and regionalize model parameters.  We have now evaluated an LSE for three models spanning a range of complexity: the conceptual HBV hydrologic model, the process-based SUMMA hydrologic model, and the Community Terrestrial Systems Model (CTSM) land model.

Our LSE approach leverages static catchment attributes and parameter values across basins to train ML emulators, which are then coupled with optimization algorithms (e.g., Genetic Algorithm) to iteratively refine parameter estimates. This iterative process enhances both the accuracy and number of parameter samples, progressively improving model performance. Results show that the LSE-based optimization achieved median modified Kling Gupta Efficiency (KGE') values of 0.65 for CTSM, 0.76 for SUMMA, and above 0.8 for HBV. Those values are competitive with or better than comparable model calibration results in past years, and outperform local calibrations based on single-site emulators (SSEs). This presentation will highlight the methodologies, key results, and challenges of implementing LSE-based calibration for hydrologic and land models at a continental scale, emphasizing the potential for regionalization and improved predictive capabilities in large-domain hydrologic modeling.