Skip to content

Weather and climate models

seed pillar 7 · Domain codes and toolchains

Cheatsheet page

Gotchas, not tutorials. Write the entry the day you fight the tool.

WRF

Workflow: WPS (geogrid → ungrib → metgrid) → real.exe → wrf.exe.

./geogrid.exe          # static fields onto your domain
./link_grib.csh /path/to/GRIB/
./ungrib.exe           # decode driving data
./metgrid.exe          # horizontally interpolate onto the domain
./real.exe             # vertical interpolation, IC/BC
mpirun -np 256 ./wrf.exe

namelist.wps and namelist.input must agree on domain dimensions, nest ratios, projection and start/end times. Most WRF failures are a mismatch between the two, and the error messages point at the symptom rather than the inconsistency.

Settings that decide the answer:

Setting Note
dx, dy below ~4 km, turn the cumulus scheme off (cu_physics = 0)
time_step ~6× dx in km as a starting point; reduce if unstable
e_vert more levels near the surface for boundary-layer work
parent_grid_ratio 3 or 5; even ratios are discouraged
num_metgrid_levels must match what ungrib produced
bl_pbl_physics / sf_sfclay_physics must be a compatible pair

The PBL/surface-layer pairing is a real constraint — not all combinations are valid, and an invalid one may run and produce nonsense rather than stopping.

MPAS

Unstructured centroidal Voronoi mesh, global with a smoothly variable-resolution region — which avoids the abrupt nest boundaries that WRF has. Mesh files are downloaded or generated, not defined in a namelist. Output is on an unstructured grid, so standard lat/lon tooling needs a conversion step (convert_mpas).

ERF

Energy Research and Forecasting — an AMReX-based atmospheric model, GPU-first, with WRF-like physics. Relevant here as the modern rewrite of that lineage on block-structured AMR.

Radiation

RRTMG is the standard broadband scheme in WRF and many others: correlated-\(k\), separate longwave and shortwave. libRadtran is the reference line-by-line/DISORT toolkit — far more accurate, far too slow for inline use, and the right thing to validate against.

Post-processing

cdo sinfo wrfout_d01_*            # structure
ncl_filedump wrfout_d01_*         # WRF-aware dump

import wrf, netCDF4
nc = netCDF4.Dataset("wrfout_d01_2026-09-26_00:00:00")
slp = wrf.getvar(nc, "slp")               # diagnostics computed correctly
lats, lons = wrf.latlon_coords(slp)

Use wrf-python rather than reading raw variables: WRF stores perturbation fields, and several diagnostics require adding a base state that is easy to forget.

Failure modes

Cumulus scheme left on in the grey zone. At 3–4 km with cu_physics active, the scheme and the resolved dynamics both produce convection. Too much light rain, suppressed storms. See NWP.

Raw WRF fields read without the base state. Geopotential is PH + PHB, pressure is P + PB, potential temperature is T + 300. Reading T alone gives a field 300 K too cold, which is obvious; reading P alone gives a small perturbation that looks plausible.

Namelist mismatch between WPS and WRF. Domain sizes, num_metgrid_levels, or time range disagreeing. real.exe may succeed and wrf.exe fail much later with an unrelated message.

Timestep too large for the grid. CFL violation appears as cfl warnings then a segfault, often after hours. The 6×dx(km) rule is a start, not a guarantee — steep terrain needs less.

Spin-up included in verification. The first few hours reflect the initialisation, not the model.

Nest boundary too close to the feature of interest. Lateral boundary errors propagate inward; keep the region of interest well inside.

Terrain smoothing insufficient for steep topography. Produces the pressure-gradient error described in NWP — spurious circulations locked to the terrain.

Connections

Sources

  • WRF User's Guide and the WRF-ARW Technical Note (the Tech Note is where the physics options are actually described).
  • MPAS-Atmosphere User's Guide; ERF documentation.
  • Books — Kalnay, Warner.