20 Model Availability, Coupling and Recommended Use

This chapter summarizes which model families can be selected together from the GUI and what each combination means for a simulation. Availability depends on the complete case configuration; two individually available models are not automatically a valid combination.

20.1 Carrier-flow families

NeuralFlow supports steady and transient cell-centred finite-volume solutions for compressible materials and constant-density artificial-compressibility flow. The carrier model can be inviscid, laminar, GE RANS, or GE with the SAS extension. The coupled block grows automatically when turbulence, species, dispersed phase or VOF variables are active.

20.2 Homogeneous VOF

The homogeneous VOF model uses physical phases with independent fractions, mixture thermodynamics, bounded/simplex-preserving updates and Upwind, HRIC or transient CICSAM interface transport. Compressible constituents use the compressible carrier path; an all-constant-density mixture uses artificial compressibility.

20.3 Dilute dispersed phase

The separate dispersed-phase model has its own velocity, density/loading, temperature and optional interfacial-area transport. It is intended for dilute particles/droplets with drag/heat-transfer and breakup/coalescence closures, and is different from homogeneous VOF.

20.4 Species and reactions

Species transport uses independent mass fractions. Nonreacting, laminar finite-rate and turbulence-limited finite-rate modes are available. Volumetric reacting species are kept separate from the currently supported VOF configuration.

20.5 Turbulence and wall treatment

GE provides low-Reynolds modification, production limiting, Kato–Launder production, modified SST behavior, curvature correction and configurable turbulent Prandtl numbers. SAS adds a scale-adaptive -source and is intended for transient use with sufficient spatial and temporal resolution.

20.6 Solid thermal and charring

Solid zones support steady/transient heat conduction and optional anisotropic conductivity. Fluid-solid coupling transfers thermal information through coupled walls. A separate one-dimensional charring-material wall response is available for the dedicated wall thermal mode.

20.7 Linear and multigrid solution

The coupled implicit system can use preconditioned Krylov methods, algebraic multigrid, or multigrid-preconditioned flexible Krylov methods. CPU execution is available with Intel MKL and supported GPU acceleration with NVIDIA CUDA. Geometric agglomeration multigrid and FMG provide separate nonlinear/coarse-grid acceleration and initialization controls.

20.8 Parallel execution

Domain decomposition distributes cells across solver processes. Neighboring partitions exchange interface data while global norms/linear operations use distributed reductions. Changing partition count is intended to change wall-clock cost, not the converged physical solution.

20.9 NeuralFlowML

NeuralFlowML is a solver-trained recurrent GNO for the current learned state. It repeatedly proposes local corrections and is trained using NeuralFlow finite-volume residual/Jacobian feedback. VOF, species, turbulence, dispersed and solid variables remain outside the present learned state.

20.10 User qualification rule

A simulation should be considered qualified only when the complete selected combination is accepted by NeuralFlow, converges without persistent physical-limit clipping, satisfies conservation, and gives mesh/time-step/model-insensitive engineering quantities at the required accuracy.

20.11 Where to find the settings

Use GUI Options Reference for the complete user-facing control reference. In particular, see Physics Models, Boundary and Initial Conditions, Agglomeration Multigrid and FMG, and NeuralFlowML Options.