La interfaz está localizada; las descripciones y documentos de las Skills se traducen automáticamente con IA a su idioma. El código, los comandos, rutas y términos técnicos se mantienen en el original, y prevalece el documento del autor.
Explorar Skills
Productividad · SKILL.md

advanced-solvers

advanced-solvers

Route OpenMC's native-backed solvers, variance-reduction controls, C API lifecycle, and optional geometry or physics integrations with explicit build and verification limits.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentación de la Skill

GitHub

name: advanced-solvers description: "Route OpenMC's native-backed solvers, variance-reduction controls, C API lifecycle, and optional geometry or physics integrations with explicit build and verification limits." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Advanced solvers and native integrations

Use this route when the task involves random ray, CMFD, weight windows or weight-window generation, openmc.lib, a C/C++ build or CTest, MPI/DAGMC/ libMesh/NCrystal, photon or charged-particle transport, or restart/specialized native interfaces.

Route

  1. Check the native boundary first. Run scripts/check_native_features.py with --help or no arguments for a safe diagnostic; add explicit --executable, --library, and/or --build-dir paths when inspecting a particular build. Read references/native-and-optional-builds.md for CMake flags, dependency gates, and safe native test selection.
  2. Choose the solver workflow. Read references/advanced-api-reference.md for random-ray settings and restrictions, CMFD lifecycle, weight-window shapes/generation, and openmc.lib sequencing.
  3. Diagnose instead of inferring support. Read references/troubleshooting.md whenever a shared library, feature flag, optional dependency, data prerequisite, or solver input is missing. A configured option or Python class is not proof that the native capability is available in the current build.
  4. Verify at the smallest feasible level. Keep the base import openmc gate separate from the executable and libopenmc gates. Prefer import and XML/API smoke checks first, then a bounded native test only when the executable/shared library and required data are present. Do not download external data or claim an unavailable optional backend.

Boundaries

  • Route routine materials, geometry, sources, and settings construction to ../model-geometry/SKILL.md; this route only covers advanced solver-specific settings and native constraints.
  • Route ordinary statepoint, tally, track, and result analysis to ../tallies-results/SKILL.md.
  • Route package installation, executable orchestration, MPI/OpenMP environment setup, and cross-section path setup to ../setup-runtime/SKILL.md.
  • Route nuclear-data conversion, MGXS production, and depletion semantics to ../nuclear-data-depletion/SKILL.md.
  • openmc.lib is an optional native binding: base import openmc can succeed while import openmc.lib fails because the packaged libopenmc shared library has not been built or is not loadable. Treat that as a native prerequisite failure, not as proof that the Python package is unusable.
  • Do not promise GPU/CUDA support. The documented build contract is CPU/C++17 with optional OpenMP and explicitly gated native integrations.

References and bundled checks

Skills guardadas