WRF at 4 km - the engine behind Convek.
Convek runs the Weather Research and Forecasting model four times a day, post-processes the output into RASP-style soaring fields, and serves the result as JSON. This page explains what that means, why it matters, and where the model's limits are.
What WRF and RASP are
WRF is the Weather Research and Forecasting model, developed at NCAR. It is the most widely used open-source mesoscale atmospheric model in the world, used by operational weather agencies and research institutions to produce high-resolution forecasts over limited areas.
RASP (Regional Atmospheric Soaring Prediction) is a convention - a set of post-processed output fields derived from mesoscale model output and tailored for soaring. RASP fields include thermal strength, glider ceiling, cloudbase, trigger temperature, surface and upper-level winds, and soundings. Convek produces RASP-style fields on top of WRF output.
Why regional beats global for soaring
Global models (GFS, ECMWF HRES, ICON) run at roughly 9-25 km resolution. Inside a 13 km grid cell there are valleys, coastlines, different surface types - all averaged into a single value. The physics that produces thermals (terrain-forced uplift, sea breezes, convergence lines) happens below that resolution.
A regional model like WRF takes a global forecast as its boundary condition and refines it over a limited area at much higher resolution. At 4 km, convection starts being resolved explicitly - the model produces its own updrafts where the physics says they should happen, instead of relying on a cumulus scheme to fake them.
This is the threshold that matters for soaring: below 4 km, forecasts lose touch with the terrain; at 4 km and finer, they start tracking the thermal structure pilots actually fly.
Pipeline specifications
Physics stack
The schemes we run - chosen for convective-scale UK forecasting.
Boundary layer (PBL)
YSU - non-local scheme tuned for convective growth
Microphysics
Thompson - aerosol-aware, cloud/ice-resolving
Land surface
Noah-MP - multi-layer soil, vegetation, snow
Radiation
RRTMG - longwave + shortwave, aerosol-aware
Surface layer
Revised MM5 Monin-Obukhov
Dynamics
ARW core, non-hydrostatic, 3D Runge-Kutta time integration
These choices are tracked and revised - see the Optimisation blog series for the thinking behind each scheme.
From WRF to RASP - derived fields
Raw WRF gives you temperature, humidity, and winds on model levels. Convek's post-processor derives the soaring-specific fields pilots actually plan around:
wstar_msThermal updraft velocity - derived from surface heat flux and boundary-layer depth
hglider_agl_mGlider-capped thermalling ceiling - PBL height with moisture + glider performance assumptions applied
cloudbase_agl_ftConvective condensation level - parcel-based LCL with microphysics correction
thermal_trigger_temp_cSurface temperature required to generate positively buoyant thermals
day_ratingComposite soaring label: poor, marginal, fair, good, or excellent
Update schedule
Convek runs four forecast cycles a day - at 00z, 06z, 12z, and 18z UTC - to match the GFS global cycle that provides initial conditions.
Each cycle produces a +48h forecast at hourly resolution. The API automatically serves from the freshest completed cycle; typical data lag from cycle time to availability is ~90 minutes.
Honest about limitations
WRF at 4 km is the best open-source starting point for soaring, but no forecast is perfect. Specific weaknesses we track and work on:
- Marginal days. Forecast skill drops sharply on days near the thermal trigger threshold - small errors in surface temp can flip a rateable day to unflyable.
- Convergence lines. Narrow convergence features often sit at or below our grid scale - we can hint at them but not resolve them sharply.
- Sea breeze timing. Coastal sea-breeze fronts are sensitive to sea-surface temperature; we improve this with regular SST updates.
- Cloudbase wobble. Sensitive to microphysics assumptions - we track cloudbase accuracy against pilot reports continuously.
We publish our thinking on these in the Optimisation blog series.
Query the model yourself.
Free tier - 25 queries/day. Any UK coordinate, any time of day.