Release Notes: 2.0.0¶
The v2.0.0 release of the research-grade program tracked in the project
planning notes. This is the first stable release under the DRBX name
(the project was previously developed as jax_drb).
Headline capabilities in 2.0.0:
- end-to-end differentiable drift-reduced Braginskii turbulence on closed and open field lines, in tokamak and stellarator (FCI) geometry;
- Hasegawa-Wakatani drift-wave flagship with gradient-based optimization through saturated turbulence; linear dispersion solver (drift-wave, shear-Alfven, interchange) verified to machine precision;
- rotating-ellipse and island-divertor stellarators, plus imported ESSOS coil and VMEC (VMEX) equilibria, including four-field turbulence on the Landreman-Paul configuration with a sheath-drained scrape-off layer;
- hermes-3 neutral model (packaged AMJUEL rates) with a self-consistent detaching SOL (SD1D rollover);
- structured solves via
solvax(the perpendicular-Laplacian GMRES potential inversion, tridiagonal, Fourier-Helmholtz); the sync-free RHS and solvax GMRES roughly halved the single-CPU four-field step; - multi-device
shard_mapFCI stepping, bit-exact against single-device execution.
Landed (Phase 0):
- CI runs the full fast test suite on Python 3.10-3.12 instead of a seven-file slice.
- Coverage is a single whole-package branch-coverage number over
src/drbx(baseline 86%); the curated closeout/promoted coverage gates were removed. - Dependency metadata corrected: unused
diffraxandequinoxremoved, the perpendicular-Laplacian GMRES now runs onsolvax(the earlierlineaxbackend and its optional extra were removed). - The Alfven-wave benchmark input is a committed fixture instead of a machine-specific absolute path.
- Phase 1 slice 1: the mocked reference-report generators, orphan campaign modules, meeting demos, profiler audits, and one-off diagnose scripts were removed from the package (−10.4k lines).
- Incorporated the FCI operator/sharding stack from PR #3 by Aiken Xie: cell-centered FCI geometry with shard/halo layouts, consistent finite-volume operators, halo exchange, 2-field/4-field/electromagnetic models, an RK4 integrator, and slab/shifted-torus MMS, operator, domain-decomposition, halo, and multigrid verification suites. See the README section "Incorporated FCI/Sharding Stack (from PR #3)" for the exact scope.
Validation¶
The v2 validation program is the benchmark ladder B1-B10 tracked in the project planning notes. In this dev series the previously shipping gates remain in force (operator kernels vs scalar references, MMS convergence order, Alfven-wave phase-speed check); the ladder rungs land phase by phase and are recorded here as they do.
Current Boundary¶
Differentiability claims apply only to pure-JAX paths with derivative tests; host-side SciPy paths are labeled as such.
The 2.0.0 release ships at the end of the plan's Phase 8 with the full native capability matrix, the literature-anchored benchmark ladder, the closed/open tokamak/stellarator example matrix, SOLVAX-backed solvers, and strong-scaling evidence.