Why run your own WRF at all?
Global weather models stop being useful for soaring long before you reach the thermal layer. Here's why Convek runs WRF at 4 km instead of serving GFS like everyone else.
Every soaring weather app you've ever used is, underneath, one of two things. Either it's a thin wrapper around a global model - GFS, ECMWF, ICON - or it's a closed pipeline run by a company that won't let you have the data.
Global models are extraordinary pieces of engineering. They forecast the planet four times a day, from the surface to the stratosphere, in grids that cover every point on Earth. The problem is resolution. GFS runs at roughly 13 km across most of the globe, stepping down to 25 km at longer forecast ranges. ECMWF's HRES is a finer 9 km. Neither of those numbers sounds bad in isolation - until you think about what they mean for a thermalling pilot.
A 13 km grid cell is roughly 170 km² of averaged air. Inside that cell you have valleys, ridges, coastlines, different land cover, different albedo, and dozens of potential thermal sources. The model doesn't see any of it. It sees one average temperature, one average moisture, one average wind - and it advects that average forward in time. Everything that actually produces thermals happens below the resolution of the model.
This is why pilots who've looked at a global-model forecast and a high-resolution regional forecast on the same day will tell you they feel like different planets. The regional model resolves the valley wind, the sea breeze, the convergence line along the edge of a moisture boundary. The global model smears all of that out.
WRF - the Weather Research and Forecasting model - is an open-source mesoscale model designed to take global output as a boundary condition and refine it over a limited area. You hand it GFS as your forcing, you pick your domain, you pick your grid spacing, and it simulates the atmosphere at a resolution that actually resolves thermals. At 4 km, a grid cell is ~16 km² - an order of magnitude more faithful to the terrain.
Is 4 km enough? It's the threshold where convection starts being resolved explicitly rather than parameterised. Below 4 km, you're still relying on a cumulus scheme to fake the convection your grid can't see. At 4 km and finer, the model starts producing its own updrafts where the physics says they should happen. For soaring forecasting, that's the line. Higher resolution (1 km, 500 m) is useful over complex terrain, but the compute cost scales brutally - and the improvement over 4 km is marginal for most UK flying.
That's the foundation this whole series builds on. The next post walks through the UK domain we run - where the edges land, why, and how we chose the grid. After that, the physics stack: PBL scheme, microphysics, land surface, radiation. Then the optimisation work itself - tuning for thermals, getting cloudbase right, turning raw WRF output into the RASP fields the API actually serves.
If you're a pilot, the headline is: when a forecast feels right to you, it's usually because someone, somewhere, ran a regional model. Convek exists to give developers and integrators a clean API over that model, so the soaring ecosystem doesn't have to keep reinventing the pipeline.