Skip to content
windborne

Inputs and outputs

The input is a live GFS analysis in GRIB2. The output is a CF-1.8 netCDF forecast, with JSON metadata and a PNG summary alongside.

flowchart LR
  GFS[GFS f000 GRIB2<br/>138 selected records] --> PRE[normalize and assemble]
  T0[initialization time] --> WM3
  PRE --> T1[GFS input<br/>720 x 1440 x 157]
  PRE --> T2[HRES-shaped input<br/>720 x 1440 x 157]
  T1 --> WM3[WeatherMesh-3]
  T2 --> WM3
  WM3 --> OUT[ERA5-space output<br/>720 x 1440 x 157]
  OUT --> NC[CF-1.8 netCDF]
  OUT --> META[validation metadata]
  OUT --> PNG[overview plot]

Input fields

Section titled “Input fields”
Group Variables Treatment
Pressure geopotential, temperature, u wind, v wind, specific humidity Read on the union of required GFS and HRES levels
Core surface 10 m u and v wind, 2 m temperature, mean sea-level pressure Required; a missing field raises an error
Extra surface total cloud cover, 2 m dewpoint, 100 m u and v wind Required; cloud cover converted from percent to fraction
  • Source grid is 721 x 1440 at 0.25 degree, North Pole first; the South Pole row is dropped to make the 720 x 1440 mesh.
  • Values are normalized with constants/normalization.json, arranged as pressure, core surface, extra surface, then zero padding, giving 157 channels.
  • The model also gets the initialization time as a Unix timestamp and derives solar angle internally.

Model call

Section titled “Model call”
x = [gfs_tensor[None], hres_tensor[None], t0]
out = model(x, [6])
pred = out[6][0]

The prediction is de-normalized with the output mesh means and standard deviations. The raw array stays available to validation, so clamping cannot hide NaN, Inf, or consistency violations.

netCDF contents

Section titled “netCDF contents”
Group Variables Dimensions
Surface temperature_2m, dewpoint_2m, pressure_msl, wind_u_10m, wind_v_10m, wind_speed_10m, wind_u_100m, wind_v_100m, wind_speed_100m, total_cloud_cover, total_precipitation_6h (time, lat, lon)
Upper air geopotential, temperature, wind_u, wind_v, specific_humidity (time, level, lat, lon)
  • One valid time and 25 pressure levels from 10 to 1,000 hPa: time=1, level=25, lat=720, lon=1440.
  • Product constraints: negative specific humidity clipped to zero, dewpoint capped at air temperature, six-hour precipitation limited to 0 to 1,000 mm.
  • Validation records the raw humidity and dewpoint violation fractions and precipitation diagnostics before those constraints.

S3 layout

Section titled “S3 layout”
forecasts/init=<YYYY-MM-DDTHHZ>/lead=<NNNh>/weathermesh3.fNNN.nc
forecasts/init=<YYYY-MM-DDTHHZ>/lead=<NNNh>/overview.png
forecasts/init=<YYYY-MM-DDTHHZ>/lead=<NNNh>/metadata.json
forecasts/init=<YYYY-MM-DDTHHZ>/metadata.json
latest.json

Invalid forecasts use the same partition layout under quarantine/ and never update latest.json.