Solver Backends

Zoomy’s symbolic layer compiles to several numerical backends. The same Model → SystemModel → NumericalSystemModel definition drives all of them, so switching backend is a change of solver import, not a change of model.

Backend	Package	Container	Status	Page
NumPy	library/zoomy_core	zoomy_numpy	reference implementation	NumPy
JAX	library/zoomy_jax	zoomy_jax	verification backend; CPU + GPU	JAX
AMReX	library/zoomy_amrex	zoomy_amrex	block-structured AMR	AMReX
OpenFOAM	library/zoomy_foam	zoomy_openfoam	3-D coupling via preCICE	OpenFOAM
Firedrake	library/zoomy_firedrake	zoomy_firedrake	finite element	Firedrake

NumPy is the reference. Every other backend is ported from it, and a disagreement with NumPy is treated as a bug in the port. JAX is the verification backend for numerical checks.

Backends are deliberately simple loops. Well-balancing, positivity, truncation and every other scheme decision lives at the symbolic level (Numerical System Model) and is emitted through the code printers — so no backend carries its own numerical constants, and they cannot drift apart.

Note

PETSc/DMPlex and FEniCSx are early prototypes: the packages are skeletons with no test coverage and only placeholder CI containers, so their pages sit under Work in progress. Do not plan work against them yet.

Which one should I use?

• Starting out, or a 1-D/2-D case that fits in memory → NumPy.

• You need speed, GPUs, or gradients → JAX.

• Large domains needing adaptive refinement → AMReX.

• Coupling a depth-averaged model to a resolved 3-D free surface → OpenFOAM via preCICE.

Installation for each is on the Installation page — the container path is the supported one for the compiled backends.