What the solver actually does

Inviscid flow at all Mach numbers, from subsonic to hypersonic.

Full viscous stress and heat conduction, temperature-dependent transport.

Species transport with detailed kinetics, mixture-averaged diffusion, variable specific heats.

Spalart–Allmaras, k–ε, k–ω SST.

Smagorinsky and dynamic Smagorinsky subgrid closures.

Reacting large-eddy and transonic reacting configurations, same core.

Interface state from local particle distributions, splitting into a non-dissipative equilibrium part and a gated non-equilibrium part.

HLLC and hybrid central–HLLC formulations for reference and control.

Evaluated at the face from the two reconstructed states, inside the flux kernel. No stencil-wide reduction, no pre-pass.

Built from already-stored face fluxes. Orders-of-magnitude shock/smooth separation with no extra Jacobian or least-squares work.

MUSCL with limiters, and first-order paths for robustness studies.

Green–Gauss on arbitrary polyhedra.

Explicit Euler and higher-order Runge–Kutta, entirely on device. Adaptive CFL-limited stepping.

Rosenbrock-2 and -4 with analytic Jacobians, constant-volume LSRK, explicit RK4.

A local explicit coupled update, so a burning cell vents rather than accumulating constant-volume overpressure. One correction iteration suffices.

Heat release as an explicit source, so refinement interpolation cannot manufacture temperature.

Hexahedral 1→8 subdivision on a runtime sensor, adapting throughout the run.

Bey 1→8 subdivision with shortest-diagonal octahedron splitting.

Z-order addressing for coalesced access, with sort-and-scan primitives replacing serial tree walks.

Level jumps constrained with terminating cancellation, so no cell oscillates between flagged and protected.

Mass and energy to one part in 10¹³, checked every adapt cycle.

Faces between two unrefined cells are frozen; non-hexahedral cells carried through verbatim.

Space-filling-curve repartitioning, per-rank halos rebuilt after each adapt.

Reads the standard open polyhedral format directly, including decomposed cases and all patch types.

Multi-block structured generation with shared-face vertex deduplication.

Region refinement against triangulated surfaces, cell removal by ray-cast parity, per-surface patches.

Boundary vertices projected by nearest-point queries on a BVH, with quality safeguards.

3-D Morton space-filling-curve partitioning with correctly paired processor patches.

Hexahedra, tetrahedra, prisms, wedges, general polyhedra. The kernels are shape-agnostic.

330 call sites, all unsequenced-parallel, inside the algorithm set an offloading compiler accepts.

The full reacting path, chemistry included, from the same source — bit-identical to the CPU path.

Thread-parallel on the same source, used as the correctness floor.

Halo exchange at a few per cent of step time, so the solver is compute- not communication-bound.

Standard parallel algorithms compiled to a portable GPU IR, reaching hardware current toolchains have dropped.

Explicitly laid-out fields from a pooled allocator. An unused model costs nothing per cell.

Order-of-accuracy studies, Riemann problems, smooth-flow transport, shock–turbulence interaction, strong-shock reflection.

Steady detonation structure, ignition against an independent chemistry library, laminar flame speed, cross-checks against a reference code.

The standard battery for contacts between dissimilar gases, where conservative multicomponent schemes misbehave.

Positivity, mass-fraction closure and geometric closure asserted continuously, refinement included.

Refactors verified byte-identical on adaptive reacting cases, so a regression is attributable to its cause.

Hermetic and content-addressed, so a result ships with everything needed to regenerate the binary.