virtual_casing_jax

virtual_casing_jax is a JAX implementation of the virtual casing principle for computing magnetic-field contributions from plasma currents using high-order singular quadrature. It is based on the C++ reference implementation in hiddenSymmetries/virtual-casing and on the SIMSOPT virtual-casing interface in hiddenSymmetries/simsopt.

Documentation is available at virtual-casing-jax.readthedocs.io.

Install the latest release from PyPI:

python -m pip install --upgrade virtual-casing-jax

VMEX users should install its free-boundary dependency with the same Python interpreter that runs VMEX:

python -m pip install --upgrade "vmex[freeb]"
python -c "from vmex.core.freeboundary_diff import have_virtual_casing_jax; assert have_virtual_casing_jax()"

Or install from a local source checkout:

git clone https://github.com/uwplasma/virtual_casing_jax.git
cd virtual_casing_jax
python -m pip install -e .

The SIMSOPT-compatible wrapper can be used as a drop-in virtual-casing calculation when SIMSOPT is installed:

from virtual_casing_jax import VirtualCasing

vc = VirtualCasing.from_vmec(
    "wout_example.nc",
    src_nphi=32,
    trgt_nphi=32,
    trgt_ntheta=32,
    filename="auto",
)

B_external_normal = vc.B_external_normal

For lower-level JAX workflows, use VirtualCasingJAX directly after preparing surface coordinates and magnetic-field arrays:

from virtual_casing_jax import VirtualCasingJAX

vc_jax = VirtualCasingJAX()
vc_jax.setup(digits, nfp, stellsym, Nt, Np, gamma, Nt, Np, Nt, Np)
B_external = vc_jax.compute_external_B(B_total)

compute_external_B and compute_internal_B are differentiable in the source field directly. They are also differentiable in the surface coordinates once the geometry-dependent precision selection has been frozen:

plan = vc_jax.plan_precision(digits=4)

def loss(surface_coord):
    vc = VirtualCasingJAX()
    vc.setup(digits, nfp, stellsym, Nt, Np, surface_coord, Nt, Np, Nt, Np)
    return objective(vc.compute_internal_B(B_total, precision=plan))

grad = jax.grad(loss)(surface_coord)

precision=plan reuses concrete quadrature sizes and singular-patch indices, while the numerical surface geometry remains differentiable. Recreate the plan when geometry changes are large enough to alter the appropriate quadrature.

Performance features:

• Source/target tiling with auto-tuned chunk sizes.
• Rematerialization hooks for GradB singular correction.
• Optional target-scan mode to reduce GradB peak memory (scan_targets).
• Mixed-precision POU/patch tables with float64 outputs.
• Bundled Quas3/LHD/W7X geometry assets (converted from SCTL .mat).

SIMSOPT compatibility: The package ships a SIMSOPT-compatible VirtualCasing class that mirrors simsopt.mhd.virtual_casing.VirtualCasing while using the JAX backend. Import it as from virtual_casing_jax import VirtualCasing. See docs/using_simsopt.rst and the examples in examples/ for full scripts.

Reference test data: The default suite uses generated analytic cases. Upstream C++ and SIMSOPT parity data are kept outside git so a clone stays small. To run those scheduled checks locally, download the checksummed release archive and run the marked tests:

python tools/fetch_reference_data.py
pytest -m "large or reference"

The finite-beta VMEC input in examples/inputs/ remains small enough to ship with the source.

Sphinx documentation lives in docs/ and is configured for ReadTheDocs. It includes the equations, numerics, implementation details, and validation strategy. Run locally:

pip install -r docs/requirements.txt
sphinx-build -b html docs docs/_build/html

Use the profiling harness to capture JAX traces and inspect performance:

JAX_ENABLE_X64=1 python tools/profile_vc.py --case case_vc --op B --jit \
  --repeat 5 --trace-dir /tmp/vc_trace

tensorboard --logdir /tmp/vc_trace

For the new tuning knobs:

JAX_ENABLE_X64=1 XLA_FLAGS="--xla_dump_to=/tmp/vc_xla --xla_dump_hlo_as_text" \
  python tools/profile_vc.py --case case_vc_large --op GradB --jit \
  --chunk-size auto --target-chunk-size auto --pou-dtype float32 --patch-dtype float32 \
  --interp-block-size auto --remat --donate \
  --repeat 2 --trace-dir /tmp/vc_trace_case_vc_large_GradB

tensorboard --logdir /tmp/vc_trace_case_vc_large_GradB

This writes JAX traces under /tmp/vc_trace_* and HLO dumps under /tmp/vc_xla_*. See docs/performance.rst for detailed interpretation.

virtual_casing_jax can wrap VMEC boundary data as an EXTENDER-like exterior field. The current downstream integration is VMEX, whose vmex.core.freeboundary_diff module builds VmecSurfaceFieldData from a wout file or VMEX state:

from vmex import read_wout
from vmex.core.virtual_casing import surface_field_data_from_wout
from virtual_casing_jax import ExteriorFieldConfig, VirtualCasingExteriorField

wout = read_wout("wout_circular_tokamak.nc")
surface = surface_field_data_from_wout(wout, nphi=32, ntheta=32)
field = VirtualCasingExteriorField(surface, ExteriorFieldConfig(digits=8))

points = [[1.8, 0.0, 0.0]]
B_plasma = field.B_plasma_xyz(points)

# Reuse accurate on-surface data for many targets close to the LCFS.
near = field.plan_near_surface(digits=4)
B_near = field.B_plasma_near_surface_xyz(points, near)

# Large target batches can use every visible JAX device.
B_total = field.B_xyz_sharded(points)

plan_near_surface is a first-order local continuation. Bound field-line traces by distance from the LCFS, and use a converged direct off-surface schedule before interpreting farther targets or magnetic topology.

The explicit field functions and their derivatives are JAX differentiable. VMEX owns the user-facing magnetic-field object and SIMSOPT-compatible stored-point methods.

For targets outside the VMEC boundary, the plasma-current contribution uses the internal virtual-casing branch because the plasma currents are inside the LCFS. The external branch means currents outside the VMEC surface, not targets outside it.

The legacy surface_field_from_vmec_jax bridge remains available for the historical vmec_jax package name and requires that package to be importable. This field wrapper is not a self-consistent SOL or edge-MHD solver.