Abstract
To track rapid changes within our water sector, Global Water Models (GWMs) need to realistically represent hydrologic systems' response patterns - such as baseflow fraction - but are hindered by their limited ability to learn from data. Here we introduce a high-resolution physics-embedded big-data-trained model as a breakthrough in reliably capturing characteristic hydrologic response patterns ('signatures') and their shifts. By realistically representing the long-term water balance, the model revealed widespread shifts - up to ~20% over 20 years - in fundamental green-blue-water partitioning and baseflow ratios worldwide. Shifts in these response patterns, previously considered static, contributed to increasing flood risks in northern mid-latitudes, heightening water supply stresses in southern subtropical regions, and declining freshwater inputs to many European estuaries, all with ecological implications. With more accurate simulations at monthly and daily scales than current operational systems, this next-generation model resolves large, nonlinear seasonal runoff responses to rainfall ('elasticity') and streamflow flashiness in semi-arid and arid regions. These metrics highlight regions with management challenges due to large water supply variability and high climate sensitivity, but also provide tools to forecast seasonal water availability. This capability newly enables global-scale models to deliver reliable and locally relevant insights for water management.
Introduction
Effective water resource management requires Global Water Models (GWMs) that can accurately project flood hazards and water availability. However, traditional GWMs often rely on rigid parameterizations and coarse resolutions, limiting their ability to ingest and learn from massive modern observational datasets. Conversely, while purely data-driven deep learning models like LSTMs achieve high accuracy, they often function as "black boxes" that do not conserve mass or provide intermediate physical variables such as evapotranspiration (ET) or groundwater recharge. This study introduces δHBV2, a high-resolution, physics-embedded machine learning model designed to bridge this gap by capturing shifting hydrologic "signatures"—the characteristic ways landscapes respond to rainfall.
Methodology
The δHBV2 model integrates neural networks with the conceptual HBV rainfall-runoff physical model. It employs a Long Short-Term Memory (LSTM) network to generate dynamic parameters (accounting for processes like vegetation dynamics) and a Multilayer Perceptron (MLP) for static geographic attributes. Operating at the high-resolution MERIT unit basin scale (~37 km²), the model is fully differentiable, allowing for end-to-end training via gradient descent. Runoff is routed through the global river network using a differentiable Muskingum-Cunge (δMC) model, which utilizes Kolmogorov-Arnold Networks (KANs) to predict channel geometry and roughness. The model was trained on 4,746 global basins using MSWEP V2.8 precipitation and MSWX V1 temperature data.
Global Performance and Shifting Signatures
δHBV2 demonstrates superior accuracy compared to established GWMs, achieving a mean absolute bias of 32.4 mm/yr for natural rivers—a 20–55% improvement. It explains the vast majority of spatial variability in large-river mean annual runoff, with R² values ranging from 0.92 to 0.97.
A key finding of the study is that the fundamental partitioning of water into "green water" (evapotranspiration) and "blue water" (runoff) is not static. The model revealed regional shifts in these ratios of up to 20% over the last two decades. As shown in the global maps, mid-latitudes are seeing an increased fraction of surface runoff, while regions like Central Europe and South America are experiencing increased green water partitioning.
Regional Trends and Validation
The model's ability to capture temporal trends was validated against observations in North America and Europe. δHBV2 significantly outperformed traditional models in capturing the slope of hydrologic changes (R² = 0.68). One critical application of this model was identifying a severe decline in freshwater inputs to European estuaries (e.g., along the North Sea), with some sites dropping by more than 1.5% per year. These shifts increase salinity and turbidity, threatening local macrofaunal communities. Furthermore, the model highlights that in arid regions, summer streamflow is highly sensitive to precipitation from the preceding six months, suggesting that seasonal outlooks can reliably predict water availability in these vulnerable areas.
Discussion and Limitations
Despite its advancements, δHBV2 faces challenges in simulating flow-attenuating effects in large inland lakes, wetlands, and major floodplains. Additionally, it does not yet explicitly account for human interventions such as large-scale water withdrawals or complex dam operations. Performance also remains lower in regions with sparse streamflow gauge data, such as the Congo and Niger basins. Future iterations will focus on incorporating human water usage and additional physical constraints to improve realism in these complex environments.
Summary updated