Sap flow observations provide a basis for estimating transpiration and understanding forest water use dynamics and plant-climate interactions. This study developed a continental modeling approach using Long Short-Term Memory networks (LSTMs) to predict hourly tree-level sap flow across Europe based on the SAPFLUXNET database. We developed models with varying levels of training sets to evaluate performance in unseen conditions. The average Kling-Gupta Efficiency was 0.77 for gauged models trained on 50 % of time series across all forest stands and was 0.52 for ungauged models trained on 50 % of the forest stands. Continental models matched or exceeded performance of specialized and baseline models for all genera and forest stands, demonstrating the potential of LSTMs to generalize hourly sap flow across tree, climate, and forest types. This work highlights hence the potential of deep learning models to generalize sap flow for enhancing tree to continental ecohydrological investigations.

It has been proposed that conservation laws might not be beneficial for accurate hydrological modeling due to errors in input (precipitation) and target (streamflow) data (particularly at the event time scale), and this might explain why deep learning models (which are not based on enforcing closure) can out-perform catchment-scale conceptual and process-based models at predicting streamflow. We test this hypothesis at the event and multi-year time scale using physics-informed (mass conserving) machine learning and find that: (1) enforcing closure in the rainfall-runoff mass balance does appear to harm the overall skill of hydrological models, (2) deep learning models learn to account for spatiotemporally variable biases in data (3) however this “closure” effect accounts for only a small fraction of the difference in predictive skill between deep learning and conceptual models.