Roadmap.
Numra is pre-1.0. This page describes the directions we're
actively working on. It's deliberately silent on dates — the
single-maintainer release cadence makes calendar promises
unreliable, and a vague schedule is worse than none. For the
most current signal of where work is happening, see the
open issues
and the
CHANGELOG.
Active directions
Expand DAE capabilities
Today Numra's stiff solvers (Radau5, BDF) handle index-1
differential-algebraic equations with constant mass matrices.
The infrastructure for higher-index DAEs — Pantelides
structural analysis, symbolic differentiation, and consistent
initialisation — already ships, but it's not yet wired into
the solver call path. The work here is to close that gap so
higher-index problems are practical to solve, with sparse
Jacobian support and constraint stabilisation as related
extensions.
Expand PDE capabilities
Method of lines on 1D, 2D, and 3D grids is solid today —
parabolic problems (heat, diffusion, reaction-diffusion,
advection-diffusion) work with all four boundary-condition
types via the second-order stencils that ship in the MOL
pipeline. The growth areas are multi-component /
coupled-PDE systems (so hyperbolic and wave equations
become first-class), an elliptic static solver path, wiring
the existing fourth-order stencil helpers through the MOL
assemblers, and more discretisation choices beyond finite
differences.
Build numra-plot
Numerical results without a comfortable plotting story are
half-finished. numra-plot will be a Numra-native
plotting crate aimed at the
equation-class-aware visualisations Numra users actually need
— phase portraits, time-series with confidence bands,
method-of-lines surface plots, convergence diagrams — without
forcing a Python detour.
Numra-native linear algebra; remove the faer dependency
Iterative Krylov solvers (CG, GMRES, BiCGSTAB, MINRES) and
preconditioners (Jacobi, ILU, SSOR) are already native code
in numra-linalg. The dense direct path —
LU, QR, Cholesky, SVD, eigendecomposition — is currently a
wrapper around faer. The long-term goal is for
Numra to own its complete linear-algebra surface. This is the
largest item on this list and is intentionally framed as a
direction, not a milestone — the work spans BLAS-equivalent
primitives, dense factorisations, and refactors across every
downstream crate that uses linalg trait bounds.
What you can rely on today
- Stability commitments. See
/stability for MSRV, semver, and
deprecation windows.
- Solver behaviour in any released version.
Covered by the regression suite documented in
docs/audit/correctness-test-map.md.
- License terms. Free for non-commercial
academic and research use; commercial use requires a
separate license. See
/license.
How to influence the direction