20 Implementation Status and Model Compatibility
Source basis: uploaded CMPS source archive, 13 August 2026. This page replaces older audit notes that no longer reflect the current VOF and solver implementation.
20.1 Core coupled flow
The production solver contains steady and transient cell-centred finite-volume paths for compressible flow and constant-density artificial-compressibility flow. The coupled block can include carrier flow, GE turbulence variables, dilute dispersed-phase variables, species and homogeneous VOF variables according to the active case.
20.2 Homogeneous VOF
VOF is now a first-class fully coupled model with N-phase/N-minus-one composition, mixed supported EOS constituents, Upwind/HRIC/CICSAM interface transport, exact physical-time flow–VOF Jacobian coupling, simplex-preserving nonlinear updates, distributed partition consistency and restart compatibility checks.
- Compressible VOF: AUSM carrier flux.
- All-constant-density VOF: artificial-compressibility path.
- HRIC: steady or transient; physical Courant correction only when a physical time step exists.
- CICSAM: transient only.
- Agglomeration multigrid, split volumetric reactions and non-adiabatic VOF wall coupling: intentionally locked in the current production readiness gate.
20.3 Dilute dispersed phase
The separate dispersed-phase solver remains available with its own velocity, density, temperature and optional interfacial-area transport. Momentum and heat exchange, multiple particle fluxes, coalescence and breakup mechanisms are represented independently from VOF.
20.4 Species and reactions
Species transport uses an N-minus-one mass-fraction formulation. Laminar and turbulent finite-rate volumetric chemistry modes are present with stiff local source integration. This reaction operator is currently kept separate from the active VOF path.
20.5 Turbulence and near-wall treatment
The active turbulence family is GE k–omega with configurable low-Reynolds modification, production limiting, Kato–Launder option, modified SST production option, curvature correction, free-stream modification, wall treatments and an SAS extension.
20.6 Solid thermal and charring
Solid zones support steady/transient conduction and anisotropic thermal conductivity. Fluid–solid walls can couple heat transfer to the carrier-flow thermal closure. One-dimensional charring response is available as a specialized wall/material model. The current homogeneous VOF path uses adiabatic wall compatibility only.
20.7 Linear and multigrid solvers
The scalable block solver layer includes preconditioned Krylov methods, algebraic multigrid as a solver, and AMG-preconditioned flexible Krylov methods. CPU execution is documented with Intel MKL; accelerator execution is documented with NVIDIA CUDA. The separate geometric/agglomeration multigrid hierarchy supports correction, FAS and FMG strategies.
20.8 Parallel CPU/GPU execution
MPI domain decomposition is the distributed-memory model. Interface states and gradients are exchanged by neighboring ranks, while distributed global norms and linear-solver operations are collective. GPU-enabled ranks retain the same domain decomposition and execute supported local algebra on NVIDIA CUDA devices, allowing multi-GPU and multi-node GPU-cluster operation.
20.9 NeuralFlow solver-trained recurrent GNO
CMPS also contains the NeuralFlow solver-trained recurrent GNO path. NeuralFlow learns local iterative corrections from CMPS states, mesh/boundary information and finite-volume residual/Jacobian feedback; it does not replace the conventional residual definition or numerical flux laws. The current VOF readiness gate restricts VOF to conventional CFD mode.
20.10 Qualification rule
A capability is considered supported only when its complete combination passes the solver's configuration/readiness checks. The existence of two individual modules does not imply that their combination is valid. Explicit compatibility locks are therefore documented as part of the numerical formulation rather than hidden as implementation details.
20.11 Coordinates, rotating frames and body forces
Planar and axisymmetric coordinate systems are supported. Static and rotating cell-zone frames are represented; rotating zones use the absolute-velocity/grid-relative-transport formulation described in the reference-frame chapter. Compressible rotating carrier flow currently requires AUSM, and the dilute dispersed phase is locked out with rotating zones. Gravity sources are available for carrier and dilute dispersed momentum.
20.12 Source-resident versus selectable paths
The manual distinguishes files that exist in the project tree from models reachable by the active solver configuration. The current selectable carrier-flow set is inviscid, laminar and GE k–omega, with optional SAS. Source-resident helper/legacy modules are not advertised as active production models unless they are connected to the current option/readiness path.
20.13 Nonlinear and low-Mach acceleration status
The current source includes both stretched-grid pseudo-time acceleration and the isentropic-Mach preconditioning accelerator. The former modifies the local pseudo-time spectral radius through the implicit cell aspect ratio and is suppressed on mass-flow and velocity inlet faces. The latter uses an inlet-derived reference total pressure to prevent the low-Mach reference velocity from falling below the velocity implied by strong isentropic expansion. These are convergence accelerators; neither changes the target steady conservation equations.
NeuralFlow: current production limits
- Learned state schema: carrier \([u_1,\ldots,u_D,p,T]\) only.
- Training: serial-only in the audited source; inference supports CPU and NVIDIA CUDA.
- Learned boundary families: far field, pressure outlet, ordinary wall and symmetry.
- VOF and expanded carrier equation sets: conventional CFD mode.
- Frozen-Jacobian training metrics: current implementation uses the steady spatial Jacobian; transient time-Jacobian terms are not part of those learned-training metrics.