Google researchers have open-sourced their hydrology modeling framework on GitHub, giving National Meteorological and Hydrological Services access to the AI architecture powering Google Flood Hub predictions. Floods rank among the most destructive natural hazards globally, often striking with minimal warning and causing lasting damage. The move follows years of developing state-of-the-art AI models for more accurate flood forecasting.

The Python package uses PyTorch and Long Short Term Memory networks to process climate, soil, topography, and land cover data alongside meteorological forecasts. It trains on the open-source Caravan dataset while supporting custom data integration for local watershed modeling. The release includes both the original 2024 benchmarked model and an upgraded version extending predictive horizons by six days in gauged basins.

Partnership with the Czech Hydrometeorological Institute validated the approach, with CHMI developing an adapter for the Delft-FEWS operational forecasting platform used worldwide. This integration demonstrates how national agencies can incorporate machine learning into existing workflows while maintaining data sovereignty. The framework addresses World Meteorological Organization findings that Indigenous and Local Knowledge integration remains rare in hazard warnings.

By democratizing access to advanced forecasting tools, Google enables resource-constrained regions to bypass costly traditional infrastructure while empowering local experts to refine predictions with specialized datasets. The Apache 2.0 licensed repository includes comprehensive documentation and tutorial notebooks for immediate implementation.