W Studio Lab
← Work
03 / AI weather · independent research / 2026.03

One model run,
weather maps worth reading closely.

In shortAn open AI weather model produced a seven-day outlook for East China on a single cloud GPU in about half an hour. Research charts, not a forecast service.

The FuXi model (Fudan University) run through NVIDIA Earth-2 / earth2studio: GFS 0.25° initial state, initialised 2026-03-29 00UTC, 7 days / 168 hours, six variable groups over East China (105–130°E, 22–45°N). A single research case, not a forecast service.

FuXi · earth2studio · GFS 0.25° · INIT 2026-03-29 00UTC · 7 days / 168 h · 6 variable groups · ~24 min pure inference / ~34 min end to end · cloud NVIDIA T4

Seven days · six variables

Pick a day and all six charts follow; click any chart to enlarge.

500 hPa height + sea-level pressure

Z500 + SLP

500 hPa sits near 5.5 km and is the core level for the large-scale flow. Tighter contours mean stronger wind; the blue side is colder (troughs), the red-orange side warmer (ridges). Black contours are sea-level pressure, marking surface highs and lows. Across the seven days you can watch troughs and ridges travel west to east.

850 hPa wind speed + barbs

Wind 850 + barbs

850 hPa, about 1.5 km up, is the main lane for low-level jets and moisture transport. Deeper warm colours mean stronger wind; broad strong-wind bands usually mark a low-level jet. Barbs show direction; where southerly and northerly flows meet is often where fronts and rain sit.

700 hPa relative humidity + wind

RH 700 + wind

700 hPa, about 3 km up, is the most direct level for rain potential: green moist areas tend to match where it rains, brown dry areas the fair-weather zones. Follow the green across seven days and you are following the rain band.

Sea-level pressure + 2 m temperature

SLP + T2m

The chart closest to how the day feels: black isobars mark highs and lows, cool colours are colder, warm colours warmer. The boundary between air masses — the front — swings north and south over time, which is why East China’s spring weather changes so often.

10 m wind speed + barbs

Wind 10 m + barbs

Wind at 10 m, directly comparable with station observations and the base quantity for offshore work, shipping and wind-energy assessment. In the Northern Hemisphere air converges anticlockwise around lows and diverges clockwise around highs; compare with 850 hPa to read vertical wind shear.

24 h accumulated precipitation

Precip 24 h

Daily 24-hour accumulated precipitation (mm); lighter to darker means weaker to heavier. Read it with 700 hPa humidity and 850 hPa wind to see where the rain band comes from and where it goes.

Full configuration

Model
FuXi (Fudan University); natively 15-day, 7 days used here
Framework
NVIDIA Earth-2 / earth2studio
Initial state
GFS 0.25° global analysis (NOAA, public)
Resolution
0.25° (about 25 km)
Initialised
2026-03-29 00UTC
Lead time
7 days (168 h); charts at 20CST (= 12UTC) each day: +12h, +36h … +156h
Region
East China, 105–130°E, 22–45°N
Variables
500 hPa height + SLP; 850 hPa wind speed + barbs; 700 hPa RH + wind; SLP + 2 m temperature; 10 m wind speed + barbs; 24 h accumulated precipitation
Runtime
~24 min pure inference; ~34 min end to end (GFS download + inference + zarr output + charting)
Compute
Cloud NVIDIA T4 GPU, started on demand
Output
zarr

Pipeline

01GFS downloadNOAA public analysis, 0.25°
02FuXi inferenceearth2studio · ~24 min
03zarr outputsix variable groups
04Charting7 days × 6 groups, daily 20CST

~34 min end to end

Limits and notes

  • One retrospective case: the run initialised 2026-03-29 00UTC.
  • FuXi was trained on ERA5 reanalysis; this run was driven by the GFS analysis, and a different initial-state source introduces bias.
  • No verification of my own; no accuracy claim is made.
  • Research figures, not a forecast service.

Data: NOAA GFS (public) · Model: FuXi (Fudan University) · Inference: NVIDIA Earth-2 / earth2studio