Abstract. Differentiable, learnable process-based hydrologic models (abbreviated as δ or delta models) with regionalized parameterization pipelines were recently shown to provide daily streamflow prediction performance that closely approach state-of-the-art long short-term memory (LSTM) deep networks. Meanwhile, δ models provide a full suite of diagnostic physical variables and guaranteed mass conservation. Due to their physical constraints, we hypothesize that they are suitable for making extrapolated predictions. Here, we ran experiments to test (1) their ability to extrapolate to regions far from streamflow gauges; and (2) their ability to make credible projections of long-term (decadal-scale) change trends. We evaluate the models based on daily hydrograph metrics (Nash-Sutcliffe model efficiency coefficient, etc.), as well as projected decadal streamflow trends. The results show that, for spatial interpolation (test in randomly sampled ungauged basins, or PUB), δ models have mixed comparisons with LSTM, presenting better trends for annual mean flow and high flow but slightly worse for low flow. For spatial extrapolation (test in regionally held out basins, or PUR, representing a highly data-scarce scenario), δ models’ advantages in mean and high flows are more prominent. In addition, an untrained variable, evapotranspiration, retained good seasonality even for extrapolated cases. δ models’ parameterization pipeline produces parameter fields that maintain remarkably stable spatial patterns even in highly data-scarce scenarios, which explains their robustness. Combined with their interpretability and ability to assimilate multi-source observations, δ models are strong candidates for regional and global scale hydrologic simulations for climate change impact assessment.
Introduction
Hydrologists face a persistent challenge: predicting water flow in "ungauged" regions where historical data is absent. While pure deep learning models, such as Long Short-Term Memory (LSTM) networks, have achieved high accuracy in data-rich environments, they often function as "black boxes." These models frequently fail to guarantee mass conservation and struggle with spatial extrapolation—making predictions for regions significantly different from their training data.
This research explores a "middle path" using differentiable, learnable process-based hydrologic models (referred to as δ models). By combining the flexible learning capabilities of neural networks with the structural rigor of physical equations, these hybrid models aim to provide both high predictive accuracy and robust generalizability.
Methodology
The authors evaluate δ models against state-of-the-art LSTMs using the CAMELS dataset, which includes 671 basins across the contiguous United States (CONUS). The δ model framework utilizes the HBV (Hydrologiska Byråns Vattenbalansavdelning) conceptual model as its physical backbone.
A neural network (LSTM unit) is employed to learn the mapping from static basin attributes and meteorological forcings to the physical parameters of the HBV model. Because the entire pipeline is differentiable, it can be trained end-to-end. The study specifically tests two scenarios:
- Prediction in Ungauged Basins (PUB): Randomly sampled basins are held out for testing (spatial interpolation).
- Prediction in Ungauged Regions (PUR): Entire geographic regions are held out (spatial extrapolation), representing a highly data-scarce scenario.
Results and Key Insights
The results demonstrate that while LSTMs perform slightly better in interpolation (PUB) tasks, δ models excel in extrapolation (PUR) and long-term trend projection.
Robustness in Data-Scarce Scenarios
In the PUR test, δ models achieved a median Nash-Sutcliffe Efficiency (NSE) of 0.59, outperforming the pure LSTM's 0.56. The physical constraints of the δ model act as a "guardrail," preventing unrealistic predictions in regions far from the training data.
Capturing Climate Trends
δ models proved significantly more reliable for climate change impact assessments. For annual mean flow trends in the extrapolation scenario, δ models achieved an $R^2$ of 0.88, compared to 0.72 for the LSTM. This suggests that the hybrid approach better captures the underlying physical drivers of long-term change.
Internal Physical Consistency
A major advantage of δ models is their ability to produce physically consistent internal variables. Simulated evapotranspiration (ET) showed a median correlation of 0.84 with independent MODIS satellite data, despite the model only being trained on streamflow. This indicates that the physical backbone forces the neural network to find solutions that respect the broader water cycle.
Conclusion
Differentiable δ models represent a powerful tool for regional and global scale hydrologic simulations. By enforcing mass conservation and leveraging stable parameterization pipelines, they provide a more robust alternative to pure machine learning for predicting water resources in ungauged regions and under changing climate conditions. Future work will focus on improving low-flow predictions and incorporating multi-source observations, such as satellite soil moisture, to further refine model physics.
Summary updated