On June 3, 2026, Google Research released an open-source hydrology modeling framework designed to help National Meteorological and Hydrological Services integrate advanced AI-based flood forecasting into their operational workflows. The framework allows researchers and forecasters to train models using architectures and data pipelines consistent with those powering Google’s Flood Hub, which provides riverine flood alerts. By making the architecture available on GitHub under an Apache 2.0 license, Google aims to enable local experts to retain control over sensitive data while refining models with specialized, region-specific datasets.

The hydrology model operates as a Python package leveraging PyTorch, incorporating inputs such as climate patterns, soil composition, topography, and land cover. The system architecture includes Long Short Term Memory (LSTM) networks and provides a training pipeline compatible with the open-source Caravan dataset, which includes historical river data. The release contains two model versions: the original 2024 benchmark model and an upgraded version currently utilized on Flood Hub. This updated model utilizes an ME-LSTM architecture, which integrates disparate weather data—including inputs from Graphcast, the European Centre for Medium-Range Weather Forecasts (IFS), and NASA satellite estimates—into a unified prediction system. According to the internal benchmarking study, this updated version extends the reliable predictive horizon by 6 days in gauged basins and by 1 day in ungauged basins compared to the previous iteration.

To facilitate adoption, the framework allows for integration into existing industry platforms such as Delft-FEWS, a tool maintained by the Deltares research institute. A collaboration with the Czech Hydrometeorological Institute (CHMI) served as a primary validation step, demonstrating that the AI-driven model produces forecasts comparable in quality to traditional, locally calibrated conceptual models. By enabling the direct integration of Indigenous and Local Knowledge (ILK) into risk knowledge production, the open-source workflow aligns with the World Meteorological Organization’s 2025 report on multi-hazard early warning systems. This release aims to provide an accessible, scalable alternative for resource-constrained regions, reducing the reliance on costly traditional infrastructure while empowering global teams to improve local flood preparedness through high-caliber, AI-enabled insights.

On June 3, 2026, Google Research released an open-source hydrology modeling framework designed to help National Meteorological and Hydrological Services integrate advanced AI-based flood forecasting into their operational workflows. The framework allows researchers and forecasters to train models using architectures and data pipelines consistent with those powering Google’s Flood Hub, which provides riverine flood alerts. By making the architecture available on GitHub under an Apache 2.0 license, Google aims to enable local experts to retain control over sensitive data while refining models with specialized, region-specific datasets.

The hydrology model operates as a Python package leveraging PyTorch, incorporating inputs such as climate patterns, soil composition, topography, and land cover. The system architecture includes Long Short Term Memory (LSTM) networks and provides a training pipeline compatible with the open-source Caravan dataset, which includes historical river data. The release contains two model versions: the original 2024 benchmark model and an upgraded version currently utilized on Flood Hub. This updated model utilizes an ME-LSTM architecture, which integrates disparate weather data—including inputs from Graphcast, the European Centre for Medium-Range Weather Forecasts (IFS), and NASA satellite estimates—into a unified prediction system. According to the internal benchmarking study, this updated version extends the reliable predictive horizon by 6 days in gauged basins and by 1 day in ungauged basins compared to the previous iteration.

To facilitate adoption, the framework allows for integration into existing industry platforms such as Delft-FEWS, a tool maintained by the Deltares research institute. A collaboration with the Czech Hydrometeorological Institute (CHMI) served as a primary validation step, demonstrating that the AI-driven model produces forecasts comparable in quality to traditional, locally calibrated conceptual models. By enabling the direct integration of Indigenous and Local Knowledge (ILK) into risk knowledge production, the open-source workflow aligns with the World Meteorological Organization’s 2025 report on multi-hazard early warning systems. This release aims to provide an accessible, scalable alternative for resource-constrained regions, reducing the reliance on costly traditional infrastructure while empowering global teams to improve local flood preparedness through high-caliber, AI-enabled insights.