GLISS

GLISS (Global Linear Ideal Stability Solver) computes the linear ideal-MHD stability of three-dimensional toroidal equilibria with nested flux surfaces. It solves the energy-principle eigenvalue problem K x = omega^2 M x with Fourier harmonics in the angles and spline finite elements in the radius, reads equilibria from the GVEC CAS3D export, and is built for differentiability: verified assembly kernels carry Enzyme-generated derivative actions. The public equilibrium-parameter to spectrum gradient chain and optimization loop remain under construction.

Version 0.0.2 supports production fixed-boundary FEEC spectra and energies, Mercier diagnostics, and symmetric GVEC or VMEC equilibrium input. Selected free-boundary operators remain research components; they do not form a public physical plasma-vacuum API. The TERPSICHORE FORT.23/24 entry points reproduce that code's stored discretization for validation and are labeled compatibility paths throughout the API and documentation.

Python

The Python package is the primary user interface. Install it with python -m pip install gliss; version 0.0.2 provides reusable Equilibrium and fixed-boundary StabilityProblem contexts with typed, certified lowest-eigenpair results, opt-in full spectra with per-pair diagnostics, deterministic full-spectrum run containers, and atomic versioned equilibrium export. The API also exposes the shared two-component marginality operator through an explicit general 3-D mode table and the axisymmetric family used for the pinned Solov'ev comparison with DCON. A separate CAS3D2MN phase-envelope entry point translates the ordered carrier/envelope table and calls the same production assembly and eigensolver. The default physical-L2 norm canonicalizes coincident Fourier modes. The explicit Schwab coefficient norm instead pulls that physical operator back to every labeled envelope coefficient, retaining the exact redundant zero-stiffness directions and evaluating inertia on the physical quotient. Paired TERPSICHORE FORT.23/FORT.24 files from a MODELK=0 pressureless- pseudoplasma run can be solved through the same public Python package, with the stored TERPSICHORE mode available for direct diagnostic comparison. This dense same-basis compatibility path is a validation tool; it is not the production physical plasma-vacuum interface. See the Python guide for examples, conventions, input and output contracts, direct VMEC conversion, and the optional SIMSOPT adapter.

Release 0.0.2 provides a manylinux x86-64 wheel and a source distribution. macOS wheels, asymmetric or precomputed BOOZ_XFORM input, the production free-boundary solve, and the complete equilibrium-to-spectrum derivative chain are tracked as future work.

Build

Requires CMake, Ninja, a Fortran compiler, BLAS/LAPACK, PkgConfig, and the NetCDF C library. A clean single-config build defaults to the optimized Release configuration, enables OpenMP assembly, and prefers threaded OpenBLAS. If OpenBLAS is not installed, CMake falls back to another available BLAS/LAPACK provider. GLISS does not set a thread count: the OpenMP and BLAS runtimes use their default thread counts.

cmake -S . -B build -G Ninja
cmake --build build
ctest --test-dir build --output-on-failure

Before a release, audit the committed tree for compiler-generated Fortran array temporaries and run the complete test suite under the audited -O3 build:

./ci/array_temporary_audit.sh

The script uses a detached temporary worktree and a private fo cache, so it does not reconfigure the normal build tree. Set GLISS_AUDIT_TMPDIR to place the temporary build on a large or fast filesystem.

The Enzyme gradient gate needs matching Flang, opt, llvm-link, and LLVMEnzyme versions:

cmake -S . -B build-enzyme -G Ninja \
  -DCMAKE_Fortran_COMPILER=flang-new \
  -DGLISS_ENABLE_ENZYME=ON \
  -DENZYME_PLUGIN=/path/to/LLVMEnzyme-22.so
cmake --build build-enzyme
ctest --test-dir build-enzyme -L enzyme --output-on-failure

Formulation and provenance

The formulation follows the CAS3D energy-principle programme published by Carolin Schwab, later Carolin Nuehrenberg (one author): the 1991 dissertation and the 1993 formulation paper appeared under her maiden name, the capability papers from 1996 on under her married name. Further methods derive from Bernstein et al. (1958) for the energy principle, Newcomb (1960) and Suydam (1958) for the cylindrical gates, Mercier (1960) and Landreman and Jorge (2020) for the interchange criterion, and Anderson et al. (1990) for eigenvalue counting by matrix inertia. PROVENANCE.md maps each module to its sources.

License

MIT. See LICENSE.