ArWen

ArWen is a GPU implementation of the WRF-ARW model family, with CUDA forecast kernels and Rust data paths for weather inputs and products. The Python package is gpuwm: gpuwm go takes a configuration from analysis inputs through a regional forecast to rendered products. WRF-exact results are demonstrated in strict mode for one 1-hour, 3 km configuration against stock scalar WRF 4.6.1 on glibc 2.39; the verification scope names the grid and physics used.

What it does

• Regional forecasts: gpuwm domain --source selects GFS, HRRR, ERA5, RAP, GEFS and other registered sources; gpuwm sources --json lists their routes.
• Source mappings: the source table and mapped-source guide describe the inputs each route needs.
• Nested grids: gpuwm domain --ladder 12-3 creates a nest; [experiment] feedback = 1 selects the experimental two-way path.
• Moving nests: gpuwm cyclone-setup builds a cyclone configuration with a following nest; the cyclone guide covers track selection and sizing.
• Ensembles: gpuwm go CONFIG --members N --recipe time-lagged uses earlier source cycles; --restart-roster JSON continues the original members from checkpoints.
• Products: gpuwm render RUN --series --products all draws maps with the Rust renderer; gpuwm export-grib2 RUN --out DIR writes surface and pressure-level GRIB2.
• Restart: gpuwm resume CONFIG --outdir RUN finds a valid checkpoint; gpuwm go CONFIG --restart CHECKPOINT continues into a new output folder.
• More than one GPU: gpuwm go CONFIG --devices N splits supported grids into resident slabs; multi-GPU operation describes the supported combinations.

Get ArWen

Download	Requirements
Windows desktop: GUI and TUI	x86-64 Windows; compatible NVIDIA GPU and driver; internet for first setup
Linux desktop: GUI and TUI	x86-64 Linux, glibc 2.39 or newer, such as Ubuntu 24.04; Python 3.11 or newer; X11 or Wayland with working OpenGL
Integration kit	Developer guide, CLI/plan examples, catalogs, and client design notes

The desktop archives contain the applications, native map library and map assets; Windows includes a graphical setup launcher that downloads a private Python and CUDA runtime on first use. The Python package gpuwm (below) carries the engine, native processing tools and TUI in its Windows and Linux platform wheels. The release notes and checksums list the exact artifacts and qualification records.

Get running

Local forecasts need a compatible NVIDIA GPU and driver; the runs below used one RTX 5090 with 32 GiB VRAM. Use Python 3.11 or newer on Linux x86-64 or Windows x86-64. The platform wheels include the native tools; the desktop GUI is a separate download.

Linux

python3 -m venv .venv
. .venv/bin/activate
python -m pip install 'gpuwm[all-cu12]==2.9.8'
gpuwm fetch-tables   # optional: stage Thompson tables ahead of the first run
gpuwm fetch-geog
gpuwm doctor

Windows

py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install 'gpuwm[all-cu12]==2.9.8'
gpuwm fetch-tables   # optional: stage Thompson tables ahead of the first run
gpuwm fetch-geog
gpuwm doctor

Activating the environment puts gpuwm on PATH, and every command the program prints starts with it; gpuwm doctor checks that. gpuwm setup --with-geog runs the same staging (native tools, physics tables and geography) as one resumable command.

fetch-geog sets up the global geography every forecast builds its terrain, land use and soil from, once per computer. It downloads about 2.7 GB and takes about 34 GB of disk once unpacked, including the land-use datasets gpuwm mesh builds MPAS statics from. Every default source is permissively licensed (the soil categories are SoilGrids250m, CC BY 4.0; the deep-soil temperature is built on your machine from ERA5). Add --list to see each dataset and its size without downloading. --datasets wrf stages only what regional forecasts read (about 2.6 GB, about 28 GB unpacked); gpuwm doctor then reports the mesh geography as missing. Without it a forecast is refused at its start.

gpuwm doctor reports each remedy as a command or a # comment explaining the manual step. For CUDA 13, offline installation, source builds, data setup and desktop launchers, use the setup guide.

For a developer installation from a source checkout, use bash install.sh on Linux or powershell -File install.ps1 on Windows. Source installations build the native components and need the corresponding build tools. They make the environment on free-threaded Python 3.14t, using a python3.14t already on PATH or installing one with uv, because multi-card forecasts run about twice as fast on it; --python (-Python on Windows) picks another interpreter. The explicit bash install.sh invocation also works when an extracted source archive has lost executable permission bits; use it if ./install.sh reports Permission denied.

Choose an area and write a three-hour GFS configuration:

gpuwm domain --point 37.5,-90 --point-extent-km 1500 --root-dx 12 --source gfs --cycle latest --hours 3 --out forecast.toml

Run the forecast with one command:

gpuwm go forecast.toml

The command fetches the inputs, prepares the domain, runs the model and writes forecast files, maps and diagnostics in the run folder it prints. Each launch gets a timestamped folder under forecast-go by default. Use gpuwm go forecast.toml --dry-run to inspect the plan before fetching.

Examples

These examples use the configuration above. Replace RUN with the exact run folder printed by gpuwm go; use separate output roots for separate runs.

Three members from GFS cycles

gpuwm go forecast.toml --members 3 --recipe time-lagged --outdir ensemble

This runs three source trajectories over the same forecast window and writes ensemble products. Add --keep-member-files to retain full histories for individual members.

A parent and a 3 km nest

gpuwm domain --point 37.5,-90 --card 32gb --ladder 12-3 --source gfs --cycle latest --hours 3 --out nested.toml
gpuwm go nested.toml

The generated configuration carries the parent and child grids. See configuration for feedback and grid settings. Sub-kilometer storm diagnostics are not resolved tornado dynamics; the scale note explains their interpretation.

Continue an interrupted run

gpuwm resume forecast.toml --outdir RUN

Use the original configuration and run folder. The restart identity check rejects incompatible configurations, and the default retention is one complete checkpoint set.

Render a saved forecast

gpuwm render RUN --series --products t2,refl --out maps

This writes 2 m temperature and composite-reflectivity maps from saved frames. --products all requests the full catalog; the render summary records products that the available fields or forecast times cannot supply.

Export forecast fields

gpuwm export-grib2 RUN --out grib2

The exporter writes surface and pressure-level files. The GRIB2 guide describes fields and definition choices.

Measured runs

These are individual Linux acceptance runs of the 2.9.2 candidate on one RTX 5090, recorded on 2026-10-17. Times are complete command wall times, including products, with geography installed and shared input caches as noted in the measurement record.

Configuration	Forecast length	Command	Wall time
GFS, 12 km, 1500 km extent	3 h	gpuwm go	88.151 s
NCAR CONUS, 12 km, supplied WRF inputs	12 h	gpuwm run --wrfinput	137.915 s
NCAR CONUS, 12 km, supplied WRF restart	Final 1 h	gpuwm run --restart-from	25.428 s

These timings exclude installation and geography download. They describe those configurations and cache states, not a throughput guarantee for other grids.

Whole-step driver runs of the 2.9.3 candidate, recorded on 2026-10-18 and 2026-10-19 (case B on one RTX 4090, no products; case A on both cards; the HRRR-shape case on one RTX 5090 with products on):

Configuration	Forecast length	Wall time
Nested 24 h case A, 1600 parent and 4800 child steps	24 h	111.0 s process, 101.5 s integration on an RTX 5090 (products on cost 0 s); 144.1 s on the RTX 4090, 171.2 s there before the driver
Nested 24 h case B, 4800 steps	24 h	84.8 s (111.8 s before the driver)
HRRR-shape, 796 x 636 x 49 at 3 km, dt 15 s, 5760 steps, products on	24 h	1944.8 s whole command (2092.5 s before), 1495.1 s integration

Documentation

• CLI manual and option reference.
• WRF input and restart files.
• Earth2Studio export: grids, consumer channels and GRIB2 fields.
• Physics choices, hardware and output variables.
• Verification and release changes.
• Desktop and terminal workspace.
• Stage contract for running preprocessing, the simulation and rendering on your own terms.
• Smoke, dust and air quality: wildfire smoke, dust, sea salt, carbon and CAMS air-quality species are an opt-in product, off unless a configuration names chem_sets; the worked example is in CONFIGURATION.md.

Report problems through GitHub issues with the version, platform, command and error text. gpuwm report RUN --output diagnostics.zip collects a diagnostic bundle; review it for private information before sharing, following the reporting guide.

ArWen is research software; use official meteorological services for safety warnings.

The engine is licensed under Apache-2.0. Third-party code and data retain their terms; see NOTICE and provenance.