Integrated hydrologic models can simulate coupled surface and subsurface processes but are computationally expensive to run at high resolutions over large domains. Here we develop a novel deep learning model to emulate continental-scale subsurface flows simulated by the integrated ParFlow-CLM model. We compare convolutional neural networks like ResNet and UNet run autoregressively against our novel architecture called the Forced SpatioTemporal RNN (FSTR). The FSTR model incorporates separate encoding of initial conditions, static parameters, and meteorological forcings, which are fused in a recurrent loop to produce spatiotemporal predictions of groundwater. We evaluate the model architectures on their ability to reproduce 4D pressure heads, water table depths, and surface soil moisture over the contiguous US at 1km resolution and daily time steps over the course of a full water year. The FSTR model shows superior performance to the baseline models, producing stable simulations that capture both seasonal and event-scale dynamics across a wide array of hydroclimatic regimes. The emulators provide over 1000x speedup compared to the original physical model, which will enable new capabilities like uncertainty quantification and data assimilation for integrated hydrologic modeling that were not previously possible. Our results demonstrate the promise of using specialized deep learning architectures like FSTR for emulating complex process-based models without sacrificing fidelity.

Remotely sensed evapotranspiration (ETRS) is increasingly used for streamflow estimation. Earlier reports are conflicting as to whether ETRS is useful in improving streamflow estimation skills. We believe that it is because earlier works used calibrated models and explored only small subspaces of the complex relationship between model skills for streamflow (Q) and ET. To shed some light on this complex relationship, we design a novel randomized, large sample experiment to explore the full ET-Q skill space, using seven catchments in Vietnam and four global ETRS products. For each catchment and each ETRS product, we employ 10,000 SWAT (Soil and Water Assessment Tool) model runs whose parameters are randomly generated via Latin Hypercube sampling. We then assess the full joint distribution of streamflow and ET skills using all model simulations. Results show that the relationship between ET and streamflow skills varies with regions, ETRS products, and the selected performance indices. This relationship even changes with different ranges of ET skills. Parameter sensitivity analysis indicates that the most sensitive parameters could have opposite contributions to ET and streamflow skills. Conditional probability assessment reveals that with certain ETRS products, the probabilities of having good streamflow skills are high and increase with better ET skills, but for other ETRS products, good model skills for streamflow are only achievable with certain intermediate ranges of ET skills, not the best ones. Overall, our study provides a useful approach for evaluating the value of ETRS for streamflow estimation.