Skip to main content

Version 2.0

This version is an overhaul to the model that enables a larger range of input and outputs of the model.

Key updates​

  • Datasets are expanded with the added capability of weather stations, radiosondes, and topographical information.
  • Microwave sounder data is expanded to the AMSU-A sensor on 5 more satellites.
  • An initial effort to support winds was added, including from ERA5 (13 levels) and radiosondes (2 levels).
  • The underlying data projection was upgraded to HealPix too ensure each modeled pixel is of near equal size. Data is reprojected to the latitude-longitude grid for dissemenation, and resulted in a slightly lower spatial resolution of 0.23.
  • Multi-resolution data capability to better handle cross sensor differences in spatial/temporal resolution.
  • In all, this model ingests 17 modalities with resolutions varying by a factor of 2.
  • Lead_time coordinate is added to enable more efficient data analysis of forecasts across different runs.

Variables​

  • qv = Specific humidity (37 levels, pressure_37)
  • temp = Temperature (37 levels, pressure_37)

Coordinates​

  • lat: Latitude, 0.23 degrees
  • lon: Longitude, 0.23 degrees
  • time: Initial time of forecast
  • lead_time: Forecast lead time as a datetime.timedelta
  • pressure_37: 37 levels of atmospheric pressure from 1 to 1000 hpa
  • pressure_13: 13 levels of atmospheric pressure from 50 to 100 hpa
  • pressure_2: 2 levels of atmospheric pressure from 1000 to 1 hpa

Location​

AWS S3 - s3://zeusai-data/prod/earthnet/v2/forecast/{year}/{month}/{day}/earthnet.v2.forecast.6h.{year}{month}{day}{hour}00.zarr

Format​

<xarray.Dataset> Size: 4GB
Dimensions:      (lat: 782, lead_time: 7, lon: 1565, pressure_13: 13,
                  pressure_37: 37, pressure_2: 2, time: 1)
Coordinates:
  * lat          (lat) float64 6kB -89.88 -89.65 -89.42 ... 89.42 89.65 89.88
  * lead_time    (lead_time) timedelta64[ns] 56B 00:00:00 01:00:00 ... 06:00:00
  * lon          (lon) float64 13kB -179.9 -179.7 -179.4 ... 179.4 179.7 179.9
  * pressure_13  (pressure_13) int64 104B 50 100 150 200 ... 700 850 925 1000
  * pressure_37  (pressure_37) int64 296B 1 2 3 5 7 10 ... 900 925 950 975 1000
  * pressure_2   (pressure_2) int64 16B 300 500
  * time         (time) datetime64[ns] 8B 2025-07-13T16:00:00
Data variables:
    qv           (time, pressure_37, lat, lon, lead_time) float32 1GB dask.array<chunksize=(1, 10, 500, 500, 7), meta=np.ndarray>
    u_1          (time, pressure_13, lat, lon, lead_time) float32 445MB dask.array<chunksize=(1, 13, 500, 500, 7), meta=np.ndarray>
    u_2          (time, pressure_2, lat, lon, lead_time) float32 69MB dask.array<chunksize=(1, 2, 500, 500, 7), meta=np.ndarray>
    temp         (time, pressure_37, lat, lon, lead_time) float32 1GB dask.array<chunksize=(1, 10, 500, 500, 7), meta=np.ndarray>
    v_2          (time, pressure_2, lat, lon, lead_time) float32 69MB dask.array<chunksize=(1, 2, 500, 500, 7), meta=np.ndarray>
    v_1          (time, pressure_13, lat, lon, lead_time) float32 445MB dask.array<chunksize=(1, 13, 500, 500, 7), meta=np.ndarray>
    w_1          (time, pressure_13, lat, lon, lead_time) float32 445MB dask.array<chunksize=(1, 13, 500, 500, 7), meta=np.ndarray>

Dependencies​

  • pandas >= 2.3.1 (we found earlier versions are not reading times)
  • zarr >= 3.0 (preferred, migration to zarr v3 upcoming)

What's next?​

  • In version 2.1 we are aiming to expand the vertical resolution of radiosonde inputs and incorporate surface level ocean winds from the ASCAT sensor.
  • More work will be done on handling multiresolution data more effectively.