Scalability and Computational Performance of an Ecohydrological Model Using Machine Learning-Based Prediction

Despite the widespread use of distributed hydrological models for operational forecasting, climate change impact assessment, and large ensemble experiments, their computational performance and scalability are rarely systematically and reproducibly quantified. This lack of explicit information limits researchers’ and practitioners’ ability to anticipate runtimes, allocate computational resources efficiently, and design feasible modeling experiments. To address this gap, a general methodological framework was developed to assess computational scalability by varying spatial and temporal resolutions, input/output gauge densities, and hardware configurations. This framework was evaluated using the TETIS v9.1 ecohydrological model as a case study. Runtimes were systematically recorded, and a Random Forest regression model was trained to predict computational performance based exclusively on user-defined configuration variables. Model robustness was further assessed through a Monte Carlo uncertainty analysis. The results reveal clear scaling patterns: spatial resolution and output-gauge density exert the strongest influence on runtime, while temporal resolution shows nonlinear effects that depend on catchment size. The predictive tool achieved high accuracy for large hydrological simulations, with increased uncertainty limited to extremely short runtimes on high-speed processors. This study introduces a transferable framework to support efficient experimental design and operational hydrological modeling and provides the first reproducible characterization of TETIS computational scalability.