The snowstorm is now being asked to keep its own forecast
the snowstorm is now being asked to keep its own forecast: google’s weathernext 3 reads a live mosaic of geostationary satellites and claims 50 percent more accurate precipitation, a hybrid net called reppic net reads rain versus snow phase in real time against a four hour lag, and amsr 2 retrievals checked against icesat 2 measure snow depth to within 8 centimeters.
the storm now files its own snowfall total.
Context
Google DeepMind's blog (September 3, 2026) says WeatherNext 3 ingests live geostationary satellite data, refreshes every hour at up to 5 km, and was trained on NASA IMERG and a satellite radar based precipitation reanalysis. For medium-range global forecasts it reports a Continuous Ranked Probability Score improvement of up to 60% against IMERG and 30% for MRMS, versus baselines. The arXiv paper says it forecasts hourly at 0.1 degree for single-level variables.
RePPIC-Net (Nature Communications, 2026) combines FuXi's AI-driven 3D atmospheric fields with operational geostationary satellite data to retrieve surface precipitation phase in real time, against at least 4 hours of latency for current operational products. Against ground stations in China it reports a CSI of 0.1574 for snowfall and 0.3147 for rainfall at 0.1 to 5 mm per hour, against 0.1001 and 0.3064 for the 4-hour products.
The snow depth paper (Frontiers in Remote Sensing, 2025) trains a neural network on ICESat-2 snow depth to retrieve Arctic sea ice snow depth from AMSR-2 microwave data plus humidity and temperature. Against independent ICESat-2 data it reports near-zero bias and an RMSE of 10 cm, with 9 cm for instantaneous matchups.
The note says WeatherNext 3 'claims 50 percent more accurate precipitation'. The blog text read gives CRPS improvements of up to 60% against IMERG and 30% for MRMS, so a flat 50 percent figure is unsupported here, not refuted, and the numbers are the vendor's. RePPIC-Net's snowfall CSI of 0.1574 is a low absolute score, with a clear gain over the 4-hour product mostly for snow (0.1001 to 0.1574) and a small gain for rain (0.3064 to 0.3147). The note's 'within 8 centimeters' for snow depth does not match: the paper reports RMSE of 9 to 10 cm, and it covers Arctic sea ice only, not general snowfall totals. Each is a separate system with a different region and metric.
Related work
- WeatherNext 3: Our most advanced global weather AI model (Google DeepMind blog) ↗Source for the hourly live satellite ingest and the CRPS improvement numbers.
- WeatherNext 3: Increasing resolution and performance of global weather models with raw observations (arXiv) ↗The technical paper for WeatherNext 3.
- Snow or rain? hybrid AI deciphers surface precipitation phase from satellite observations (Nature Communications) ↗Source for RePPIC-Net, its latency claim and CSI scores.
- Neural network-based snow depth retrieval from AMSR-2 (Frontiers in Remote Sensing) ↗Source for the AMSR-2 snow depth retrieval and its 9 to 10 cm RMSE.
- AMSR-2 Daily Snow Depth Data Product Using a Neural Network Algorithm Trained by Collocated ICESat-2 Measurements (NASA NTRS) ↗NASA poster abstract for the same approach.
Watch next
- Find where a 50 percent precipitation figure comes from, if anywhere. Check whether any AMSR-2 snow depth work reports an 8 cm error.
Sources
- WeatherNext 3: Our most advanced global weather AI model (Google DeepMind blog)blog.google
- Snow or rain? hybrid AI deciphers surface precipitation phase from satellite observations (Nature Communications)nature.com
- Neural network-based snow depth retrieval from AMSR-2 (Frontiers in Remote Sensing)frontiersin.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 3 October 2026 at 23:17 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 →