Native plasma-physics kernels with explicit fidelity boundaries. Fast controller-support models and research solver contracts — production-parity lanes stay fail-closed until external same-case evidence exists.
| Domain | Capability |
|---|---|
| Grad-Shafranov equilibrium | Fixed and free-boundary solve; SOR and multigrid; vacuum field; GEQDSK I/O; native same-case profile-source metrics |
| Transport (1.5D) | Radial transport with QLKNN-10D and FNO neural surrogates; operator-source checks |
| Gyrokinetic (three-path) | External-code bridges, native GK research solver, and neural surrogates — see below |
| MHD stability | 7 criteria: Mercier, ballooning, Troyon, kink, RWM, tearing, vertical stability |
| Neoclassical transport | Bootstrap current, neoclassical conductivity and flux contracts |
| Disruption chain | Disruption prediction and mitigation chain; pellet/SPI injection models |
| Edge & pedestal | ELM, MARFE, and L-H transition models |
| Runaway electrons | Runaway-electron contracts (fail-closed against DREAM kinetic parity) |
| Impurity transport | Impurity contracts (fail-closed against independent Aurora/STRAHL validation) |
| Plasma-wall interaction | PWI erosion and wall-interaction contracts |
| 3D equilibrium | 3D equilibrium research operators (external coupling required for full nonlinear MHD) |
Three transport lanes are exposed with explicit fidelity limits. Lane A defers to external codes; lane B is a bounded native research solver; lane C uses neural surrogates.
| Lane | Scope | Performance / fidelity |
|---|---|---|
| A: External codes | GENE, CGYRO, GS2 same-case bridges | Full nonlinear 5D parity stays fail-closed until public same-case evidence exists |
| B: Native GK | Linear transport plus nonlinear 5D operator contracts | ~0.3 s/surface for linear eigenvalue; bounded NumPy/JAX research solver with explicit invariant benchmarks |
| C: Surrogates | QLKNN-10D transport, FNO turbulence | QLKNN test rel_L2 = 0.094; FNO val rel_L2 = 0.055 |
The public operator-source GEQDSK gate passes and the FreeGS public-example vacuum comparison passes. Native same-case profile-source metrics, finite signed-q sanity, and machine-readable strict threshold checks are published, and free-boundary strict parity is accepted for the public FreeGS same-case lane. EFIT-grade profile-source / free-boundary reconstruction remains an open evidence gate with row-level debug traces.