WRF

WRF is the Weather Research and Forecasting model, the regional atmospheric model behind every downscaling result in this garden — the Galapagos refined analysis and CER v2 alike. It solves the compressible nonhydrostatic equations on a limited-area grid, driven at its lateral boundaries by a global model or reanalysis.

The ARW dynamical core is what most people mean by WRF.

Configurability is the defining property

WRF ships with many options for every physics component: several convection schemes, several boundary-layer schemes, several microphysics and radiation packages, multiple land-surface models. The combinatorial space is enormous.

That is genuinely useful — it is why WRF works across regimes from tropical islands to urban Europe — and it is the source of nearly every difficulty people have with it.

The options are not independent. The physics map covers the couplings; the short version is that convection, microphysics, and radiation exchange cloud fields, and the surface, boundary-layer and land-surface schemes exchange fluxes. Choosing each on its own merits produces configurations that are individually defensible and collectively inconsistent, and the resulting bias has no single cause to find.

Defaults are not neutral. They reflect the regimes the developers worked in, largely mid-latitude continental. A scheme performing well over the continental United States has no obligation to perform well over the equatorial Pacific — which is why the DARWIN experiments treated physics as variables to test rather than settings to inherit.

The workflow

WPS prepares the domains and interpolates the driving data; real builds initial and boundary conditions; wrf integrates. Most of the effort in a new setup goes into domain configuration and getting the static data right, not into running the model.

Two things I would tell anyone starting. Spin up the land surface — soil moisture inherited from a reanalysis is not in equilibrium with WRF’s own land-surface physics, and the resulting drift looks exactly like a model bias. And decide the output frequency deliberately, because I/O rather than computation is usually what limits a run on shared infrastructure.

Nesting bridges resolutions, and a nest whose parent ratio is too large produces artefacts at the boundary. Below about 4 km, switch the convection scheme off and run convection-permitting.

See also: ARW, the tropical setup, dynamical downscaling, and the physics map.