The cloudburst is now being asked to keep its own schedule
the cloudburst is now being asked to keep its own schedule: microsoft’s mw nowcast runs six hour ensemble radar forecasts and doubles detection of the most intense rain across the us europe and china, a generative model called stormdit trained on 7,720 chinese storms holds skill through the two to six hour gray zone, and flowsnet keeps a nonzero hit rate on category four rain past the three hour limit.
the sky now files its own downpour log.
Context
MW-Nowcast (arXiv, September 2026, Microsoft authors) is a six-hour radar nowcaster. A deterministic decoder predicts storm structure shared by all ensemble members and a flow-matching generator samples only the residual. The abstract says it beats leading methods for heavy and extreme rain through the 6 h horizon across the United States, Europe and China, and for the most intense rain doubles the warning time, giving 6 h forecasts with skill previously limited to 3 h for the leading generative baseline.
StormDiT (arXiv, January 2026) is a unified generative diffusion transformer trained on 7,720 precipitation events from China. On a heavy-rain test set it keeps a Critical Success Index near 0.2 for echoes of 35 dBZ or more across a 6 hour forecast at 6 minute steps, with a spread-skill ratio of about 0.96. On the SEVIR benchmark it raised 1-hour heavy rain CSI from 0.054 to 0.130.
FlowsNet (EGU General Assembly 2026 abstract, Kumamoto University and others) fuses radar and gauge analyses, ground station data and satellite imagery to target rain of 50 mm per hour or more over Kyushu, Japan. It keeps a non-zero category-4 CSI through 6 h and from 4 to 6 h matches or exceeds the Japan Meteorological Agency's very-short-range forecast.
The note's 'doubles detection of the most intense rain' is looser than the paper: the abstract says it doubles the usable warning time, with 6 h skill matching what the leading baseline had at 3 h. A third-party summary adds that the baseline, NowcastNet, was retrained to 36 frames although it was designed for 3 h, and that recent diffusion nowcasters were not compared. StormDiT's 'skill through the gray zone' is a CSI near 0.2, which is modest in absolute terms. FlowsNet's result is a non-zero CSI, not a 'hit rate', and it covers one region of Japan. Each is a different test set, threshold and metric, so the three cannot be ranked against each other. The two arXiv items are preprints and FlowsNet is a conference abstract.
Related work
- MW-Nowcast: Six-hour ensemble nowcasting of extreme precipitation (arXiv) ↗Source for the six-hour ensemble, the US, Europe and China tests and the doubling claim.
- MW-Nowcast: Six-hour ensemble nowcasting of extreme precipitation (Paper Feed summary) ↗Third-party summary that describes the deterministic and generative split and the baseline caveat.
- StormDiT: A generative AI model bridges the 2-6 hour 'gray zone' in precipitation nowcasting (arXiv) ↗Source for the 7,720 events, CSI near 0.2 and the SEVIR result.
- An NWP-Free, Observation-Driven Deep Learning Approach to Heavy-Rainfall Nowcasting Beyond the Three-Hour Limit (EGU26 abstract) ↗Source for FlowsNet, its 50 mm per hour target and the 4 to 6 h comparison with JMA.
- Hybrid physics-AI outperforms numerical weather prediction for extreme precipitation nowcasting (npj Climate and Atmospheric Science) ↗An earlier peer reviewed hybrid approach to the same problem.
Watch next
- Check MW-Nowcast against a diffusion baseline such as DiffCast on the same data. Look for how StormDiT and FlowsNet hold up on events outside China and Kyushu.
Sources
- MW-Nowcast: Six-hour ensemble nowcasting of extreme precipitation (arXiv)arxiv.org
- StormDiT: A generative AI model bridges the 2-6 hour 'gray zone' in precipitation nowcasting (arXiv)arxiv.org
- An NWP-Free, Observation-Driven Deep Learning Approach to Heavy-Rainfall Nowcasting Beyond the Three-Hour Limit (EGU26 abstract)meetingorganizer.copernicus.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 01:03 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 →