vmec-jax 0.0.18

JAX implementation of VMEC2000 with differentiable fixed-boundary and branch-local free-boundary research paths.

vmec-jax

End-to-end differentiable JAX implementation of VMEC2000 for fixed-boundary workflows, with free-boundary support, VMEC2000-compatible mgrid workflows, and direct-coil research paths. Full adaptive free-boundary solve adjoints remain in development.

Runtime Snapshot

This full bundled single-grid fixed-boundary matrix compares VMEC2000, vmec_jax cold/warm CPU runs, and VMEC++ where VMEC++ converges cleanly. The performance docs keep the detailed CSV/JSON provenance, WOUT-parity rows, memory columns, and current-vs-main regression classifications. In short, warm vmec_jax is faster than VMEC2000 on most bundled rows, while cold tiny rows still pay Python/JAX/XLA setup cost.

Differentiation Evidence

The current differentiable diagnostics agree with central finite differences for fixed-boundary geometry/profile scalars, QS/QI residuals, DMerc, D_R, and branch-local direct-coil free-boundary scalars. The free-boundary rows are same-branch/fingerprint-gated evidence only; arbitrary adaptive branch changes are still an explicit research lane. Detailed commands and tolerances are in the validation guide.

Install

pip install vmec-jax

The plain package includes plotting support and booz_xform_jax; no separate extra is needed. If bare pip does not install into the Python you intend to use, check that pip --version and python -m pip --version agree; use the matching python -m pip form only if bare pip points at the wrong interpreter.

pixi add vmec-jax
conda install --channel conda-forge vmec-jax

Developer install from source:

git clone https://github.com/uwplasma/vmec_jax
cd vmec_jax
pip install -e .

Large optional validation assets stay out of git; inspect released bundles with python tools/fetch_assets.py --list.

Quick Start

For a first run after pip install vmec-jax, use the bundled test case:

vmec --doctor
vmec --test

vmec --doctor prints Python, pip, package, and JAX backend diagnostics. vmec --test copies input.nfp4_QH_warm_start, runs with FTOL_ARRAY = 1e-12, writes WOUT and plots under vmec_jax_test/, and prints equivalent manual commands. The canonical executable is vmec; old aliases remain supported.

To run the same workflow manually with an input downloaded from the repository:

curl -L -O https://raw.githubusercontent.com/uwplasma/vmec_jax/main/examples/data/input.nfp4_QH_warm_start
vmec input.nfp4_QH_warm_start

vmec --plot wout_nfp4_QH_warm_start.nc
vmec --plot wout_nfp4_QH_warm_start.nc --outdir figures/

Run Boozer coordinates with bundled booz_xform_jax; by default --booz uses mbooz = 32, nbooz = 32, and all VMEC surfaces:

vmec --booz input.nfp4_QH_warm_start
vmec --booz --plot input.nfp4_QH_warm_start
vmec --booz wout_nfp4_QH_warm_start.nc
vmec --plot boozmn_nfp4_QH_warm_start.nc

Use the Python API:

import vmec_jax as vj

run = vj.run_fixed_boundary("input.nfp4_QH_warm_start")
wout_path = "wout_nfp4_QH_warm_start.nc"
vj.write_wout_from_fixed_boundary_run(wout_path, run, include_fsq=True)
vj.plot_wout(wout_path, outdir="figures/")
boozmn = vj.run_booz_xform(wout_path, mbooz=32, nbooz=32)
vj.plot_boozmn(boozmn, outdir="figures/")

Profile-polynomial, spline, finite-beta, and free-boundary examples are in examples/ and documented in the performance, validation, and free-boundary guides.

Direct-Coil Free-Boundary Research Lane

The direct-coil free-boundary lane samples differentiable Biot-Savart coils directly while keeping the existing mgrid path for VMEC2000 compatibility. Current coil-only examples validate complete-solve acceptance plus same-branch, fingerprint-gated branch-local derivatives; arbitrary adaptive host-controller branch differentiation remains a research lane.

ESSOS direct-coil, generated-mgrid, finite-beta scan, and coil-only QS optimization commands are documented in docs/free_boundary_coil_optimization.rst.

Backend Selection

vmec_jax follows the selected JAX backend. If CPU-only JAX is installed, runs use CPU. If GPU-enabled JAX is installed and selected, runs use the accelerator; vmec_jax does not silently force those runs back to CPU. Install or upgrade GPU-enabled JAX using the official JAX installation matrix: https://docs.jax.dev/en/latest/installation.html

python -c "import jax; print(jax.default_backend()); print(jax.devices())"
JAX_PLATFORMS=cpu vmec input.nfp4_QH_warm_start
JAX_PLATFORM_NAME=gpu vmec input.nfp4_QH_warm_start
JAX_PLATFORMS=cuda vmec input.nfp4_QH_warm_start

From Python, leave solver_device unset to inherit JAX's default backend, or pass solver_device="cpu" / solver_device="gpu" explicitly.

Optimization Examples

Editable optimization examples live in examples/optimization/. Start with examples/optimization/README.md, then use docs/optimization.rst, docs/optimization_sweep_results.rst, and docs/piecewise_omnigenous_plan.rst.

The compact panels show QA/QH/QP common-minimal-seed runs and QI NFP1/2/3/4 minimal-seed examples. For QI, the public per-NFP scripts start from the same circular-torus-like input.minimal_seed_nfp* decks, first build a QP basin, and then switch the objective to QI. Full numeric tables, caveats, LASYM panels, and artifact-promotion rules live in the docs, with historical readme_best_optimization_qa.png, readme_best_optimization_qh.png, readme_best_optimization_qp.png, and readme_best_optimization_qi.png archived there.

Reproduce the minimal-seed optimization rows with:

PYTHONPATH=. JAX_PLATFORMS=cuda python3 examples/optimization/generate_minimal_seed_showcase.py \
  --cases qa_nfp2,qa_nfp3,qh_nfp3,qh_nfp4,qp_nfp2,qp_nfp3 \
  --backend-label gpu \
  --solver-device gpu --worker-jax-platforms cuda --policy continuation --max-mode 5 --ess on \
  --max-nfev 70 --continuation-nfev 20 --inner-max-iter 550 --inner-ftol 1e-10 \
  --trial-max-iter 550 --trial-ftol 1e-10 --ess-alpha 1.2 --case-timeout-s 7200 --rerun
PYTHONPATH=. python examples/optimization/render_minimal_seed_showcase.py --publication-matrix
PYTHONPATH=. python examples/optimization/render_qi_readme_cases.py

Run individual editable examples with python examples/optimization/QA_optimization.py, QH_optimization.py, QP_optimization.py, or QI_optimization.py. The public simple QI examples are QI_optimization_nfp1.py through QI_optimization_nfp4.py; each file exposes the seed, objective tuples, QP stage, QI stage, saved outputs, and plots directly. The seed-3127 preset is retained as a diagnostic stress case, not a README promotion row. Full provenance and artifact-promotion rules live in the docs: docs/optimization.rst and docs/optimization_sweep_results.rst.

Performance, Validation, Release

• Performance notes: docs/performance.rst; validation, coverage, and release gates: docs/validation.rst, docs/testing_strategy.rst, and docs/release_checklist.rst.
• Latest repository release tag: v0.0.16.

CLI Reference

vmec input.*           run the equilibrium solver and write wout_*.nc
vmec --plot wout.nc    generate VMEC diagnostic plots from a WOUT file
vmec --booz wout.nc    run booz_xform_jax and write boozmn_*.nc
vmec --plot boozmn.nc  generate Boozer contour and spectrum plots
vmec --parity input.*  force the conservative VMEC2000-style loop
vmec --solver-mode memory input.*  choose the lower-peak-memory parity path
vmec --help            show the full option list