Skip to content

vmex 0.11.1

Choose a tag to compare

@rogeriojorge rogeriojorge released this 23 Sep 23:30
· 496 commits to main since this release
750de50

Free-boundary trials are anchored on the coupled root their gradients
differentiate, the exterior field is accurate next to the plasma surface,
optimization results are the states the objective scored, and the
optimization examples were re-tuned against a five-minute budget.

Fixed

  • Free-boundary values and gradients referred to different points (#432).
    VMEC's ftol does not place a free-boundary solve on the coupled
    plasma--vacuum root the adjoint differentiates. Every certifiable solve is
    now Newton-anchored on that root (refine_tol); one that cannot be
    anchored is the new status 3, never a silent pass. A stalled restart from
    the reference gets the iterations the cold reference needed before the
    cold retry of #416.
  • equilibrium_from_x returned the host solve, not the refined state the
    objective read
    (#427). On li383_low_res the mean iota differed by
    1.6e-3 relative; it now matches the objective, and so does the WOUT
    written from it.
  • Un-jitted gradients of state objectives recompiled on every call (#438).
    A concrete solve status now takes its branch in Python instead of a
    lax.cond: an eager repeat of value_and_grad on a Solov'ev deck takes
    0.9-1.7 s with nothing compiled, against 12-15 s and three programs.
  • The exterior field from a free-boundary wout without coil currents had no
    coils
    (#430): VmecExtender.from_wout now raises. The eager derivative
    check no longer warns "up to inf" far from the surface
    (virtual-casing-jax>=0.0.8, the new freeb floor).
  • The QI mirror hybrid example crashed on a private import; the mirror
    and gradient examples follow the example template (#434).

Added

  • The exterior field is accurate next to the plasma surface (#441).
    Eager VmecExtender calls switch each point the direct quadrature cannot
    resolve to a target-graded rule (near_surface="auto", default): 1e-12 in
    B down to 0.01 minor radii on the 2.5 % beta QA deck, where the default
    grid was off by order one. with_graded_quadrature() uses it everywhere,
    under jit too. from_wout/from_state accept project_current (off by
    default, #381).
  • Six single-stage examples (#411): penalty, augmented Lagrangian, least
    squares, 0.5 % beta, and free boundary at zero and 0.5 % beta. Each meets
    every target it states in 129-275 s on the reference laptop.

Changed

  • Cheaper free-boundary anchor and boundary-Schur adjoint (#439). The
    anchor's preconditioner includes NESTOR's edge coupling through a Woodbury
    update, cutting its Krylov iterations 4-6x (1483 -> 310 at the optimized
    coils of the finite-beta single stage) with the anchored values unchanged
    to 10 digits.
  • Optimization example defaults (#426): QA uses helicity [1, 2]; QI and
    QP use [1, 3], with QI at aspect 8 and an aspect weight of at least 0.01;
    finite beta is set through PHIEDGE. 18 scripts were timed under five
    minutes with a converged NS = 71 check; scripts over budget keep their
    previous defaults, and timing the remaining scripts is a follow-up.
  • The implicit max_fsq_ratio default is 1e2, matching every optimize
    entry point (#434); take_gradients.py is merged into
    take_fixed_boundary_gradients.py.
  • Optimization movies colour each frame from a plain forward solve (#436):
    the single-stage movie takes 12.6-21.9 s cold against 57.0 s.
  • Direct-path exterior-field derivatives use closed-form layer kernels
    (#441, same values to 1e-12): first B..gradgradgradB calls 6.0 s ->
    2.9 s, warm gradgradgradB 4-6x faster
    (benchmarks/extender_ab_20260923.json).
  • The summary figure plots bootstrap <J.B> on the force-balance panel,
    computes the effective ripple on 7 surfaces, and no longer writes
    _profiles.png; the README shows two single-stage movies (#442).
  • The mean-iota implicit derivative's 1-3 % gap to re-solve finite
    differences is the m = 1 angle-gauge drift of the solver's path, not a
    missing term (#428).
  • CI: the GPU workflow is retired, the weekly and publish lanes are fixed,
    every manifest lane must run in a workflow or be local-only (#437), and
    parity lane c3d is split into c3 and d (#438). Uncited benchmark records
    are removed (#431); the docs, plan and README are consolidated (#413, #429,
    #433, #435, #440).

Removed

  • VmecExtender.with_near_surface_continuation and near_surface_plan
    (1.6-2.4 % error floor, 18.5 GB to prepare); use with_graded_quadrature().