A coastal, estuarine and tidal ocean solver in Fortran. It runs on GPUs and CPUs from one source: do concurrent does the parallelism, with a thin layer of OpenACC for reductions and device data. It builds with NVHPC (GPU and multicore), gfortran 15 and ifx, and has a Python interface.

Rakali is under active development and will not be stable until v1.0.0: expect breaking changes to the namelist, the Python API and the output files. It is an experiment, not a production model. See Status before relying on it.

• Grids: structured Cartesian and unstructured triangular, both on an A-grid.
• Equations: 2D shallow water, 3D hydrostatic multilayer, and non-hydrostatic (structured only).
• Numerics: HLL/HLLC on structured grids, Kurganov-Noelle-Petrova central-upwind on triangles, SSP-RK2, and a semi-implicit mode on triangles. Wetting and drying are well-balanced.
• Vertical: sigma, z-sigma and the z-star family, with conservative ALE remap.
• Physics: salinity and temperature plus passive tracers, vertical mixing (PP81, KPP, k-epsilon), Smagorinsky and biharmonic viscosity, wind, pressure and surface fluxes.
• Boundaries: wall, open, tidal, nested, inflow, discharge, clamped, sponge, Chapman.
• I/O: NetCDF (CF-1.8) output, restart, gauges.

docs/CAPABILITIES_AND_LIMITATIONS.md lists what works and what doesn't. docs/CLOSURE_MATRIX.md says which closure runs in which regime.

cmake -B build -S . && cmake --build build
cd build && ctest --output-on-failure      # never pass -j to ctest on a GPU build

Requires CMake, a Fortran compiler (gfortran 15 or newer, ifx, or NVHPC) and NetCDF-Fortran. pic and test-drive are fetched by CMake. tools/build_netcdf_fortran.sh builds NetCDF-Fortran for your compiler on top of an existing NetCDF-C.

./build/rakali validation_examples/coastal/castro_dam_break/castro_dam_break.nml

cmake -B build -S . -DRAKALI_ENABLE_PYTHON=ON && cmake --build build --target rakali_shared
pip install -e ".[all]"

from rakali import Rakali
with Rakali(nx=200, ny=100, dx=10.0, dy=10.0, t_end=3600.0,
            bc_west="wall", bc_east="wall") as sim:
    sim.advance(60.0)
    print(sim.h_phys.max())      # zero-copy numpy view

See python/README.md, including the unstructured API and known limitations.

Most of Rakali's code was written by Claude, Anthropic's AI model, working through Claude Code. The design is Jorge Luis Galvez Vallejo's: the architecture, the API, the memory model and the parallelism model. It claims no novelty.

There is no reference model to check against. The unstructured scheme follows the approach of ANUGA; everything else (the structured solver, multilayer, vertical coordinates, closures) was designed here. Correctness rests on the tests in tests/: analytical cases (lake at rest, Ritter and Thacker, Carrier-Greenspan, seiche convergence, Kato-Phillips, lock exchange) and conservation checks.

Treat these as experimental: k-epsilon, zstar_full (it leaks 1-2% of salt per tidal cycle in intertidal estuaries), the non-hydrostatic solver, and nested boundaries on triangles. Fatal errors inside the Fortran library end in ERROR STOP, which kills a Python session.

• docs/REFERENCE.md: physics and namelist reference
• src/core/coastal/README.md: design contract
• docs/codebase/INDEX.md: repo map and run lifecycle
• FORTRAN_STYLE.md: style and the GPU programming guide
• API docs are generated from the source docstrings with FORD and published by CI.

Apache 2.0. See LICENSE.