Towards hybrid modeling of the global hydrological cycle

Abstract. State-of-the-art global hydrological models (GHMs) exhibit large uncertainties in simulations due to the complexity, diversity, and heterogeneity of land surface subsurface processes, as well scale dependency these processes associated parameters. Recent progress machine learning, fueled by relevant Earth observation data streams, may help overcome challenges. But learning methods are not bound physical laws, their interpretability is limited design. In this study, we exemplify a hybrid approach modeling that exploits adaptivity neural networks for representing uncertain within model structure based on principles (e.g., mass conservation) form basis GHMs. This combination knowledge can potentially lead data-driven, yet physically consistent partially interpretable models. The (H2M), extended from Kraft et al. (2020), simulates dynamics snow, soil moisture, groundwater storage globally at 1? spatial resolution daily time step. Water fluxes simulated an embedded recurrent network. We trained simultaneously against observational products terrestrial water variations (TWS), grid cell runoff (Q), evapotranspiration (ET), snow equivalent (SWE) with multi-task approach. find H2M capable reproducing key patterns cycle components, performances being least par four state-of-the-art GHMs which provide necessary benchmark H2M. neural-network-learned responses antecedent moisture states qualitatively our understanding theory. contributions groundwater, snowpack variability TWS plausible ranges traditional identifies somewhat stronger role transitional tropical regions compared With findings analysis, conclude provides new data-driven perspective machine-learned parameters complementary existing frameworks. approaches have potential better leverage ever-increasing streams advance understandings system capabilities monitor it.