Glossary
Short definitions of the WRF/NWP terms and the wrf_gpu-specific concepts used throughout this guide. If you already run WRF, the first table should feel familiar; the second covers vocabulary particular to this GPU re-implementation.
WRF & NWP terms¶
| Term | Definition |
|---|---|
| ARW | Advanced Research WRF — the nonhydrostatic, mass-coordinate WRF dynamical core that wrf_gpu re-implements on the GPU. |
met_em | WPS metgrid output files: meteorological forcing interpolated onto the model grid. In the current public CLI they are part of the upstream preprocessing chain that produces wrfinput/wrfbdy; the run path consumes those WRF real-data artifacts. See Input Data. |
wrfinput | The initial-condition file for a domain — the atmospheric and land state at forecast hour 0. The current standalone path reads it directly. |
wrfbdy | The lateral boundary-condition file for the outermost domain: time-varying values that force the domain edges through the forecast. |
wrfout | The WRF history output file (NetCDF) holding model fields at each output time. wrf_gpu writes a WRF-compatible wrfout. See Output. |
wrfrst | A WRF restart file — a full model-state checkpoint used to resume a forecast from a saved time. gpuwrf has an internal WRF-named serializer, but the public forecast command does not yet read or write operational restarts. |
namelist.input | The WRF run-configuration file (domains, timestep, physics options, output cadence). wrf_gpu reads the same file; see Namelist Compatibility. |
| RK3 | The third-order Runge-Kutta integrator that advances the slow, meteorologically significant terms within each ARW timestep. |
| acoustic sub-step | A short sub-step inside each RK3 stage that advances the fast acoustic and gravity-wave modes; many acoustic sub-steps run per RK3 stage. |
| split-explicit | The ARW time-integration strategy: slow modes on the large RK3 step, fast acoustic modes on smaller explicit sub-steps — avoiding a globally tiny timestep. |
| nonhydrostatic | A dynamical core that solves the full vertical momentum equation (no hydrostatic approximation), required for realistic kilometre-scale flow. |
| nesting (one-way / two-way) | Running a finer child domain inside a coarser parent. One-way: the parent forces the child only. Two-way: the child additionally feeds its solution back onto the overlapping parent cells. See Nesting. |
| MYNN | The MYNN-EDMF planetary-boundary-layer scheme (turbulent vertical mixing) — one of the GPU-operational PBL options. |
| Thompson | The Thompson bulk microphysics scheme for cloud and precipitation, including an aerosol-aware variant. |
| RRTMG | The RRTMG longwave-plus-shortwave radiation scheme (with topographic shading and slope effects in wrf_gpu). |
| Noah-MP | The Noah Multi-Physics land-surface model, evolving soil and vegetation state (sf_surface_physics=4). See Physics Options. |
wrf_gpu concepts¶
| Term | Definition |
|---|---|
| cell-identity proof | Evidence that wrf_gpu's field values match CPU-WRF cell-for-cell to a stated tolerance — the identity claim for the dynamics/thermodynamics core. See Validation. |
| oracle | The trusted reference that judges correctness — CPU-WRF v4 output, or an analytic solution — that wrf_gpu is validated against. wrf_gpu is validated against WRF as an oracle; it is not derived from the WRF source. |
| cold compile | The one-time XLA compilation of a given geometry on its first run (minutes). See Performance. |
| warm cache | Reuse of a previously compiled executable on later runs, skipping the cold compile. |
| AOT cheap-key | The default-on ahead-of-time warm-start: a metadata-only key computed without lowering locates and loads the matching compiled executable across processes in microseconds. Fail-open — any miss falls back to a normal compile. |
| fused runtime | The default nested execution path that lowers a parent and its child domains into one XLA program for higher GPU throughput; bit-identical to the eager per-domain path. |
| batched ensemble | Running B independent, same-geometry forecasts at once under one outer jax.vmap to fill an otherwise launch-bound small-grid GPU. Opt-in via GPUWRF_BATCH_ENSEMBLE. |
| fp32 (default since v0.3.0) | WRF's own single precision. Since v0.3.0 the default path runs fused fp32 kernels, with double precision only where WRF uses it; GPUWRF_FAST_DEFAULTS=0 turns them off. Before v0.3.0 an opt-in fp32 acoustic-step mode existed for capability and lower VRAM. |
| fail-closed | When an unsupported option is requested, the run stops before any compute with a named reason — it never silently substitutes a different scheme. See Boundaries. |
| reference-only | A scheme present in the source for reference or comparison but not GPU-operational; it is refused (fail-closed) on the run path rather than executed. |
| tolerance manifest | The declared per-field acceptance tolerances that a validation run must meet against the oracle. |
| proof object | A test result, profiler artifact, ADR, fixture comparison, or reproducible command that supports a claim. Every claim in this project has one on disk. |
| whole-state GPU residency | The design property that high-frequency model state stays on the GPU across timestep loops, with no host round-trips inside the loop. |
| WRF-compatible | Compatible where it matters — variables, fixtures, namelist mapping, and validation behavior — with WRF, without being a Fortran port. |
| physics-identical-enough | Statistically and physically consistent with trusted WRF or analytic references under documented tolerances. |
See also
Cross-references: Concepts & Architecture, Validation against WRF, Performance & GPU Modes, and the Environment Variables that toggle several of the concepts above.