Open-Source ME-LSTM Framework Extends Flood Forecasts by 6 Days
Google Research has open-sourced the hydrology architecture behind Flood Hub, enabling local agencies to run ME-LSTM forecasting models on private data.
On June 3, 2026, Google Research released the AI-based hydrology modeling framework behind its Flood Hub platform to the public via GitHub. The official open-source release provides National Meteorological and Hydrological Services (NMHSs) with the complete architecture and training pipeline for Google’s Global Flood Forecasting Model Version 2 (v2).
By releasing the underlying machine learning components, Google is decoupling its core predictive engine from its centralized platform. This allows local meteorological agencies to deploy state-of-the-art AI forecasting within their own infrastructure while maintaining strict governance over sensitive regional data.
Architecture and Lead Times
The v2 forecasting system shifts to a Multi-Explanation Long Short-Term Memory (ME-LSTM) architecture. To generate accurate predictive horizons, the model natively integrates meteorological forcings from GraphCast, Google’s machine learning weather simulator.
This integration of deep learning weather models with hydrological simulation yields distinct performance gains over the previous generation. Research published in EGUsphere earlier this year demonstrated the impact of the v2 system on operational timelines.
| Basin Type | V2 Extension Over V1 |
|---|---|
| Gauged Basins | +6 days reliable lead time |
| Ungauged Basins | +1 day reliable lead time |
The framework is compatible out-of-the-box with the open-source Caravan dataset for global streamflow analysis. However, its primary operational value lies in local extensibility. NMHS forecasters can now feed specialized, non-public gauge telemetry into the ME-LSTM model locally to refine warnings for specific terrain constraints.
The Groundsource Foundation
The hydrology framework release caps a series of rapid 2026 updates to Google’s crisis response stack. In March, the company launched Groundsource, a methodology utilizing Gemini models to automate data extraction from decades of public reporting. That system successfully identified over 2.6 million historical flood events across more than 150 countries.
The Groundsource dataset directly enabled Google to expand Flood Hub into urban flash flood forecasting, achieving predictive capabilities up to 24 hours in advance. Currently, the centralized Flood Hub covers over 1,800 sites in 80 countries, reaching an estimated population of 460 million people.
Operational Deployment Strategy
Prior to the GitHub release, Google validated the operational viability of the ME-LSTM system through internal trials with the Czech Hydrometeorological Institute (CHMI). This testing phase ensured the framework’s inference patterns and ingestion pipelines matched the operational realities of a national weather service.
For developers building climate resilience tools, the open-source pipeline changes how predictive models handle spatial generalization. Instead of fine-tuning models exclusively on public datasets, regional agencies can fork the v2 architecture and combine the GraphCast forcings with proprietary ground-truth measurements to close the accuracy gap in ungauged river systems.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Run TPU Workloads on Google Cloud with Ray 2.55
Learn how to provision Google Cloud TPUs, handle slice topologies, and deploy machine learning models using Ray 2.55 and the KubeRay Operator.
Score Smoothing Drives Generative Novelty in Diffusion Models
Google Research demonstrates that generative novelty in diffusion models is a predictable mathematical consequence of score smoothing during training.
Google SensorFM Trains on 1 Trillion Minutes of Wearable Data
Google Research launched SensorFM, a foundation model pre-trained on one trillion minutes of wearable data to power generalized health prediction agents.
Zero-Shot TabFM Skips XGBoost Tuning for Sub-Second Predictions
Google Research released TabFM, a zero-shot foundation model that frames tabular data prediction as an in-context learning problem to bypass the fit step.
Google Finds Reasoning Tokens Expand LLM Parametric Recall
Google Research proves that generating reasoning tokens allows language models to retrieve unreachable parametric facts via a computational buffer effect.