Methodology and published results

We publish what we measured, misses included, and say plainly what has not been measured yet.

The models

Wind, storm and hail forecasts run on AceCAST, the WRF weather model built to run natively on NVIDIA GPUs. The worldwide wind forecast (experimental) runs on MPAS.

Speed on the product today: a 12-hour hail forecast, 500 km across at 1 km, took 62 minutes from order to finished forecast on one GPU (Sep 28, 2026). Times vary with the weather and the queue.

Hub-height wind: twelve nights in Oklahoma

At an 88-turbine Oklahoma wind farm in 2023, forecasts were scored every ten minutes against public DOE AWAKEN lidar at 80 m. Nightly error is the root-mean-square error of wind speed. The 1 km run blended with HRRR had 11 to 13 percent less error than HRRR alone.

On its own, the 1 km run did not reliably beat HRRR at the tower. The blend is research and not on sale; what you buy is the 1 km run.

Mean nightly error over 12 nights, by lidar tower
Lidar tower (80 m)US national model (HRRR) alone1 km run blended with HRRRLess errorNights better
Tower A12.58 m/s2.25 m/s12.9%11 of 12
Tower H2.46 m/s2.18 m/s11.3%11 of 12

Not yet scored

  • Storm forecasts are not yet scored against observations.
  • Hail size is experimental guidance, not yet systematically scored.
  • Worldwide wind forecast (experimental): Hub-height winds (80–200 m), 10 m gust, turbulence (TKE), rotor-layer shear and veer, and 1 h wind-speed change are produced by the model but not yet scored against observations.

How gaps are shown

Missing, withheld and not-yet-evaluated are shown as exactly that, never as clear. Absence of data is never a clear-weather signal.

The Logbook records what the forecast said and what was not assessed; no entry is marked right or wrong yet.

Aerava is not a safety system. Its checks are asset protection and planning information, never a safety verdict; your lift plan, site procedures and the manufacturer’s limits decide whether work goes ahead.

Runs and files

Credits held for a run are returned to your balance automatically when the run fails, cannot be submitted or is canceled before it starts running. Returned credits are credits, not money.

Maps, charts, point series, tables and Logbook history are kept until you ask us to delete them or your account is closed. Raw model output and data packages are available for at least 30 days after each run.