Skip to content

Hazards ​

This page collects the PRISM inputs that are refused, and the ones that run but behave differently than their key names suggest. Every other PRISM page links here instead of restating these rules. The keys themselves are in the input reference.

Two stop points appear below. Load is the parsing of the INI file. Initialization is the backend set-up after it, before the first step. A few checks fire later, at the first solve that needs them; the tables say so.

Divergence correction ​

The GLM scalars φ (for D) and ψ (for B) exist only under [numerics] divergence_correction = hyperbolic. constrained_transport selects which of them are carried. physics sets their slots and the variable names from one branch (#11, #61):

constrained_transportdivergence_correctionResult
nononv = 9
D, B, DBnoconstrained transport ignored with a warning, nv = 9
D or Bhyperbolicnv = 10 (φ or ψ in slot 7)
DBhyperbolicnv = 11 (φ in 7, ψ in 8)
nohyperbolicrefused at load: there is no scalar to carry
anypoissonrefused at load: no backend implements the projection

PIC models add rho, one more slot, up to NV_MAX = 12 (state vector). The cleaning equations are on the models page.

Seam div(B) ​

On a uniform grid the centred curl and divergence commute, so ∇h⋅B stays at round-off (numerics). At a 2:1 AMR face the two sides difference at different resolutions and the identity fails. The seam injects a ∇⋅B source in the grid-oscillation modes.

  • The source is truncation-order, O(hp): it converges under refinement.
  • At fixed h it grows without bound in time, independent of Δt. On rmf-amr-fd-pulse (source-free pulse, 1 rank, ni = 16) the seam max |∇⋅B| grows from 6.9 to 1002 in 100 steps; the uniform-grid control stays at 0.
  • ‖B‖ stays bounded (−3 % over the same run): this is a constraint violation, not an energy instability.
  • The same source appears with fd_centered and fv_centered, and on octrees and quadtrees.

Five local fixes were ruled out (#29). The resolution is accept-truncation. An upwinded or dissipative GLM coupling at the seams is open as #99; a staggered-B route is parked as #30. The measurements are on the forest page.

Key ([IO])DefaultEffect
seam_divB_tol-1 (off)When > 0, warn when the global max |∇⋅B| outside the absorbing layers exceeds it. The maximum is not seam-local; on a grid without 2:1 faces it stays at round-off and the monitor stays silent.
seam_divB_error.false..true. stops the run instead of warning.

Mitigation. divergence_correction = hyperbolic with constrained_transport = B or DB and a damping rate [physics] alpha > 0 delays the growth. It does not cure it.

Inter-realm seams. A 1:1 inter-realm seam is divergence-free only with coupling_cadence = stage_coincident (β); under α the stages after the first read stale seam ghosts (#31). The cadences are on the numerics page.

PML constraints ​

BackendAcceptedChecked at
CPUPML_type CLASSIC, CLASSIC_DIRECT, BERMUDEZ, CFS; scheme_space = fd_centered; scheme_time = runge_kutta with SSP-11/22/33/54; divergence_correction = no (hence no constrained transport)initialization
FNLthe same PML_type kinds; fd_centered; runge_kutta with SSP-11/22/33/54; no hyperbolic correction, no constrained transportinitialization, before any device allocation

Any other combination is fatal. The [PML] section itself is also checked at load:

  • width > 0 and the six face flags are required; every flag .false. is fatal.
  • Each PML_type has its own required parameters (the [PML] table). A missing one is fatal, not defaulted. An unknown PML_type is fatal.
  • CLASSIC_DIRECT is not a Runge–Kutta PML. It skips the auxiliary variables and their integration, and damps the tangential components directly. It is a testing instrument, not the production path (absorbing layers).

Elliptic boundary map ​

The t0 solves (numerics) run on FLAIL's FFT solver with PRISM's own operator D(G) of order fdv_order. Each face kind maps to an elliptic kind (build_elliptic_bc_types), then to a FLAIL face kind (solve_time_zero):

[bc_*] typeElliptic kindFLAIL face kind
periodicperiodicFLAIL_BC_PERIODIC
DirichletDirichletFLAIL_BC_ZERO_GHOSTS (every ghost layer zero)
PECPECFLAIL_BC_DIRICHLET (φ, tangential A), FLAIL_BC_NEUMANN (normal A)
Neumann, Silver_Mullerexact/openFLAIL_BC_FREE_GHOSTS (the free-space potential of the source in the ghosts, summed over the whole domain)
extrapolation—fatal, at the first t0 solve

The map is total, but it runs only when a solve first touches the faces. An input with an extrapolation face passes load and initialization, and stops at t0 if a PIC field or a t0 current needs a solve. A null axis takes the periodic kind.

Refused combinations ​

InputWhere it stops
[numerics] scheme_time = leapfrog, blanes_moan, commutator_free_magnusinitialization, both backends (#55)
[PIC] scheme_time = leapfroginitialization, both backends
[numerics] divergence_correction = poissonload
divergence_correction = hyperbolic without constrained transportload
[bc_*] type = radiativeload (#57)
[initial_conditions] type = riemann-problem, magnetic_nozzle, rmf_magnetic_nozzleload (#56)
[external_fields] external_fields_applied = magnetic_nozzle, rmf_and_magnetic_nozzleinitialization
[numerics] scheme_space = weno on FNLinitialization, after the device allocation
runge-kutta-yoshida on FNLinitialization
a non-SSP Runge–Kutta scheme with PIC or the PMLinitialization
a non-SSP Runge–Kutta scheme on the staged path (2:1 faces, forests)when the forest first queries the stages
[PIC] current_weighting_model = conserving on the CPU backend with more than one block or one rankinitialization, at the t0 conserving deposit (the first step on a restart)
[PIC] current_weighting_model = conserving on the staged path (2:1 faces, forests)the first staged step, both backends
[PIC] current_weighting_model = conserving with periodic faces and current_conserving_solver = londrillo or esirkepov_tail_cleanup = .true.initialization, both backends (PIC)
a refined grid that enters a t0 elliptic solvethe solve (#84 E7: the wide operator across 2:1 faces is not defined)
fdv_order = 10 with a t0 elliptic solvethe solve (FLAIL accepts 2, 4, 6, 8)
an extrapolation face with a t0 elliptic solvethe solve (above)

Every refusal prints a message naming the combination. src/tests/prism/regression/refusals.sh keeps them refused, one leg per refusal (cookbook).