The river is now being asked to predict its own surge
the river is now being asked to predict its own surge: ecmwf’s aifl learns global streamflow straight from data for operational flood forecasts, a gpu native solver maps ensemble inundation probabilities for the july 2025 central texas floods, and a convlstm with attention scores 0.84 nash sutcliffe across 14 flood events.
the river now files its own flood forecast.
Context
ECMWF's blog (August 27, 2026) introduces AIFL, a global streamflow model that learns from data. An LSTM is pre-trained on 40 years of ERA5-Land reanalysis (1980 to 2019) over more than 18,000 basins from the Caravan dataset, then fine-tuned on operational IFS control forecasts from 2016 to 2019 to cope with the shift from reanalysis to real forecasts. Its maps show many stations with Kling-Gupta efficiency scores well above 0.5 in Europe, North America, South Africa and Australia and New Zealand.
Inunda (arXiv, July 2026) solves the two-dimensional shallow water equations in PyTorch on a GPU with a mass-conservative scheme. For the July 2025 Central Texas flash flood it is driven by an 18-member 1 kilometer convection-allowing rainfall ensemble to make probabilistic flood forecasts, whose skill improves as lead time to the crest shortens. For Hurricane Harvey it reports a median gage Nash-Sutcliffe efficiency of 0.72 against 0.31 for the operational National Water Model v3.0.
The ConvLSTM paper builds a flood forecasting system with a ConvLSTM and attention network. It reports an average Nash-Sutcliffe efficiency of 0.84 and an RMSE of 0.18 meters over 14 flood events across five years, and says it outperforms conventional hydrodynamic and other machine learning models.
The note says AIFL learns for 'operational flood forecasts'. The ECMWF post describes a study that tackles the gap between training on reanalysis and running on forecasts, and says ECMWF is 'exploring how machine learning can complement' its flood systems. Whether AIFL is running operationally is not stated in the post, so that is unsupported here, not refuted.
The Texas result is a research hindcast and forecast experiment, and the 0.72 versus 0.31 comparison is for Hurricane Harvey in Harris County, not the July 2025 flood. A separate 2026 preprint on the Upper Guadalupe basin that uses 100-member ensembles with a different model found hindcasts matched the hydrograph shape better than the peak, so peak depth remains the hard part.
The 0.84 figure is an average over events from a single study area, an unnamed river's upper reaches in the PDF read. The paper's venue is not stated in what was read, and its claim to beat other models rests on the authors' own comparison. The three results use different metrics and different places, so they cannot be ranked against each other.
Related work
- AIFL: advancing global streamflow prediction with operational machine learning (ECMWF) ↗Source for AIFL's two-stage training and skill maps.
- Inunda: A GPU-Native, Differentiable Solver for High-Resolution Flood Inundation Modeling (arXiv) ↗Source for the GPU solver, the Texas ensemble case and the Harvey scores.
- Multi-Source Environmental Data Flood Forecasting System Based on ConvLSTM Networks (Semantic Scholar PDF) ↗Source for the ConvLSTM attention results.
- Probabilistic Flash-Flood Inundation under Ensemble Rainfall Uncertainty in the Texas Hill Country (SSRN) ↗Another study of 2025 Texas floods with 100-member rainfall ensembles and a different model.
- AIFL: A Global Daily Streamflow Forecasting Model Using a Deterministic LSTM Pre-trained on ERA5-Land and Fine-tuned on IFS (arXiv) ↗The AIFL paper itself.
Watch next
- Read the AIFL paper for its skill figures by region. Check whether the ConvLSTM paper names its river and venue.
Sources
- AIFL: advancing global streamflow prediction with operational machine learning (ECMWF)ecmwf.int
- Inunda: A GPU-Native, Differentiable Solver for High-Resolution Flood Inundation Modeling (arXiv)arxiv.org
- Multi-Source Environmental Data Flood Forecasting System Based on ConvLSTM Networks (Semantic Scholar PDF)pdfs.semanticscholar.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 3 October 2026 at 22:48 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →