Vertical levels: why the bottom 3 km gets most of the detail
WRF gives you a vertical stack of eta levels to spend however you like. For soaring, almost all of them should live in the lowest few kilometres.
A WRF run has two resolutions. The horizontal one everyone talks about (4 km, 1 km, 500 m) and the vertical one nobody talks about, which is just as important for soaring. Vertical resolution is how finely the model slices the atmosphere from the ground up to the model top.
WRF uses terrain-following eta coordinates. The level list is a set of dimensionless values between 1.0 at the surface and 0.0 at the model top, and WRF stretches the actual physical heights to fit the terrain below. You pick how many levels you want and where they sit. The default configuration spaces them fairly evenly, which is fine for weather research and not quite right for soaring.
Soaring happens in a narrow slice of atmosphere. On a good UK day the boundary layer tops out at 1800 m AGL, on an excellent day maybe 2500 m. Thermals, cloudbase, convergence, the inversion that caps the lot, all of it is in the lowest 3 km. Nothing above that matters for the forecast you are serving to a pilot, other than as context for what is diffusing downward.
The Convek UK run currently uses 35 eta levels with WRF's default vertical stretching, model top at 50 hPa. We have not yet hand-tuned the level distribution for soaring. With 35 levels and the default stretching, the bottom 3 km gets roughly a dozen levels, which is enough to resolve a daytime PBL but is not as sharp as it could be near the surface or at typical inversion heights.
Why this matters in practice. A coarse vertical grid smears out inversions. If your lowest four levels are at 0, 200, 500, and 1000 m, a 300 m-deep morning inversion gets represented by one point. The model cannot decide whether it is an inversion or a gentle gradient, and the breaking of that inversion (which is what triggers the first thermal of the day) is delayed or fuzzed. Pilots feel this as a forecast that either misses the start time by two hours or refuses to show convection at all until midday.
Packing the levels low also helps cloudbase. Cloudbase is a function of how the temperature and dewpoint profiles cross in the lower troposphere. If the crossing point sits between two widely-spaced levels, WRF has to interpolate, and small errors in the interpolation become hundred-foot errors in `cloudbase_agl_ft`. Dense levels around the typical 600 to 2000 m AGL band keep the crossing well resolved.
The cost of vertical resolution is modest compared to horizontal resolution. Adding 10 more levels (35 to 45) is roughly 20 to 30 percent extra compute - level count drives 3D advection and most physics close to linearly, and packing levels tightly low down can also force a smaller vertical timestep. Halving the horizontal grid (4 km to 2 km) is roughly 8x (four times the columns, half the time step). Vertical is still the cheaper place to buy forecast quality, which is why hand-tuned eta packing is on the optimisation list before any 2 km work.
There is a compute cost we do pay: the model top. Ours sits at 50 hPa, around 20 km altitude. That is higher than a pure soaring forecast needs, but WRF is unstable if the top is too low, particularly with the radiation schemes we use. 20 km is the compromise, with the sparse upper levels doing the job of keeping the stratosphere physically sensible without eating compute.
Next post in the series is the first field explainer: what `wstar_ms` actually represents, and why it is not a thermal strength readout. After that we come back to physics and start arguing about microphysics.