Physics options

wrf_gpu v0.4.0 runs most WRF 4.7.1 physics schemes on the GPU. This page lists every option code of the seven physics selectors with its status in this release. The codes are the same numbers you put in the &physics section of a WRF namelist.input.

What the status words mean¶

Notes used in the tables:

1. Passed the comparison, but its compiled GPU code exceeds our code-quality limits for speed and robustness (for example large per-thread memory use). It runs correctly; its speed has not been measured and may be lower than with the default physics. The resource list names each case.
2. Needs the Triton 3.6 compiler (pip install "triton==3.6.*"). The code-quality scan of its compiled GPU code could not be completed.
3. Runs, but cannot be validated: the original WRF crashes on the test case at its first output. Use it with cu_physics=0.
4. Passed a 24-hour comparison outside the release test set, and has the code-quality issue of note 1.

How the options were checked¶

The release check ran 66 configurations, each a scheme with a fixed set of companion schemes, compared hour by hour with the original CPU-WRF: 65 over 24 hours and the Goddard radiation pair over six hours. They are mostly single-domain forecasts; two-way nesting was checked on a two-domain case. Every option marked Validated 24 h or Default passed in one of them, and Goddard radiation is marked Validated 6 h. Options marked Experimental, CPU only, Python API only or Not supported are outside this check; their notes and status say why. Most configurations (51 of 66) used the bundled Swiss Alps case; options that need special surfaces, such as lakes, cities, open sea or wind farms, used cases that contain them. The validation page explains the method and the limits.

Schemes that must be paired¶

As in WRF, some schemes only work together. gpuwrf run and gpuwrf doctor check this before the forecast and name the setting to change. The most common pairings:

• QNSE: bl_pbl_physics=4 with sf_sfclay_physics=4.
• TEMF: bl_pbl_physics=10 with sf_sfclay_physics=10, without a land model or with Noah-MP.
• EEPS and KEPS (bl_pbl_physics=16/17): sf_sfclay_physics=1, without a land model or with Noah-MP.
• Goddard radiation: ra_lw_physics=5 and ra_sw_physics=5 together, with Thompson microphysics and use_mp_re=1.
• Ferrier microphysics (mp_physics=5/15): cu_physics=0.
• Grell-family and Tiedtke cumulus: cudt=0, so the scheme runs every step.

Microphysics mp_physics¶

Cumulus cu_physics¶

Planetary boundary layer bl_pbl_physics¶

Surface layer sf_sfclay_physics¶

Land surface sf_surface_physics¶

Longwave radiation ra_lw_physics¶

Shortwave radiation ra_sw_physics¶

Noah-MP options¶

With sf_surface_physics=4, the Noah-MP selectors of the &noah_mp namelist are read and checked. The default values are validated as part of the default suite. Of the other values, dveg=7/8/9, opt_run=8 and opt_infdv=1/2/3 passed 24-hour comparisons; the rest are Experimental.

Crop, irrigation, tile drainage and groundwater options need the matching geography from WPS (for example the groundwater fields for opt_run=5) and the original MPTABLE.TBL parameters. Surface mosaic (sf_surface_mosaic=1) is not supported. The Noah-MP options guide lists the inputs for each branch.

Lakes, cities, ocean and other surface options¶

The lake model, urban canopy models, the ocean mixed layer, SST updates and terrain-aware radiation are namelist features rather than physics selectors. Their status is on the namelist page.

Earlier releases¶

v0.3.4 and earlier ran a much smaller set of schemes; v0.3.3 and older could silently run the default schemes when a different one was requested. The v0.3.4 support matrix keeps that release's record.