My submission

If you would like to skip ahead: the Submission page has the deliverables and a sample plot, and the End-to-end runbook has the commands to inspect a run.

This turns the WeatherMesh-3 research code into a scheduled forecast pipeline. Every six hours it fetches a live GFS analysis, runs the model on a GPU, validates the output, and writes netCDF, metadata, and a plot to S3.

Verified run

One cycle on live GFS data:

GFS init     2026-07-10T18Z
Forecast     +6 hours
Fetch        7.0 s
Preprocess   6.5 s
Inference    7.1 s
Cycle        45.6 s
Output       720 x 1440 x 157, no NaN or Inf

Two-metre temperature ran -69.9 to 44.7 C, MSLP 961.2 to 1050.3 hPa, peak 250 hPa wind 88.7 m/s. The netCDF is about 355 MB: one time step, 720 x 1440, 25 pressure levels.

Inspect the latest forecast:

aws s3 cp s3://wm3-forecasts-194290773983/latest.json latest.json
aws s3 cp s3://wm3-forecasts-194290773983/forecasts/init=2026-07-10T18Z/lead=006h/overview.png overview.png

Pages

• Initial thoughts: the approach and the constraints along the way.
• Architecture: what runs.
• Inference pipeline: live GFS to forecast.
• Inputs and outputs: shapes, fields, file layout.
• Operationalisation: schedule, monitoring, quarantine, alerts.
• Limits and tradeoffs: the known gaps.
• End-to-end runbook: commands to build and inspect.