Dynamical Core

wrf_gpu integrates a nonhydrostatic, split-explicit ARW dynamical core on the GPU: the same equations and numerical methods as WRF's Advanced Research WRF core, rewritten in JAX. Each release is compared hour by hour with the original CPU-WRF from the same inputs; results are close to WRF but not bit-identical (see validation). The result path has no hidden clamping or nan_to_num repairs.

Time integration: RK3 + acoustic sub-stepping¶

The core uses WRF's split-explicit time scheme. A third-order Runge-Kutta (RK3) large step advances the slow (meteorologically significant) modes, while the fast acoustic and gravity-wave modes are integrated with smaller acoustic sub-steps inside each RK3 stage. Only rk_order=3 is supported.

Native single-root and live-nested runs honor time_step_sound. Zero or omission derives WRF automatic sound steps from dt/grid spacing (4 on the tracked grid). The legacy CPU-history replay driver has a separate historical default of 10; it is not the native default.

Per-domain advection controls¶

The requested moist_adv_opt and scalar_adv_opt are used for each domain, in native runs and in replay runs. The values of the default suite are validated; other values run as requested but are not separately validated (namelist page).

Precision: WRF's single precision by default¶

Since v0.3.0 the dynamical core runs WRF's own single precision (fp32) by default, with double precision only where WRF itself uses it — as in WRF's standard single-precision build. The fp32 path is checked against original CPU-WRF on the 24 h release gate. GPUWRF_FAST_DEFAULTS=0 selects the earlier fp64 path.

fp32 is the default since v0.3.0

The default path runs the validated fp32 kernels (WRF's own single precision; double precision only where WRF uses it), checked against CPU-WRF on the 24 h release gate. GPUWRF_FAST_DEFAULTS=0 turns the fast paths off. The older opt-in mixed-precision acoustic mode is described under Performance.

Horizontal diffusion¶

The diffusion settings are used exactly as requested. gpuwrf run refuses a diff_opt/km_opt pair it cannot run, instead of running without diffusion.

Closure	Setting	Status
2-D Smagorinsky (horizontal)	`diff_opt=1`, `km_opt=4`	Default — WRF's recommended real-data setting, including the terrain-slope and map-factor deformation terms
No explicit horizontal diffusion	`diff_opt=0`	Runs as requested; not separately validated
3-D TKE	`diff_opt=2`, `km_opt=2`	Experimental — passed a 24-hour comparison; not with every microphysics scheme or inside nests
SMS-3DTKE	`diff_opt=2`, `km_opt=5`	Experimental — passed a 24-hour comparison; needs `bl_pbl_physics=0` and no or Noah-MP land
Constant K	`diff_opt=2`, `km_opt=1`	Python API only
3-D Smagorinsky	`diff_opt=2`, `km_opt=3`	Not supported on the GPU

Other numerics settings: diff_6th_opt 0/2 (2 = WRF's monotonic sixth-order filter), damp_opt 0/3 (3 = upper-level Rayleigh damping of w) and w_damping 0/1.

Gravity-wave drag¶

Orographic gravity-wave drag with flow blocking (gwd_opt=1) is wired and default-ON on the nested 1 km path. It requires the sub-grid orography statics in wrfinput. Set GPUWRF_GWD_NESTED=0 to force it off for a memory-tighter config. The GSL drag suite (gwd_opt=3) is not wired.

Lateral boundaries and free-running stability¶

Limited-area runs are forced through lateral-boundary relaxation from the boundary tendencies (wrfbdy for the root; the parent supplies each nested child's boundary live — see Nesting). This is the validated operational path.

Free-running open boundaries can go unstable

Running free (no lateral-boundary relaxation) on wide domains (nx ≈ 160+) can go non-finite beyond approximately forecast hour 14. The validated operational path uses boundary forcing; keep boundary relaxation on for wide-domain runs.