Professional Documents
Culture Documents
S.P. Antal, S.M. Ettorre, R.F. Kunz* and M.Z. Podowski Center for Multiphase Research, Rensselaer Polytechnic Institute (*) Applied Research Laboratory, Pennsylvania State University ABSTRACT
This paper is concerned with the development of NPHASE, a next generation computational fluid dynamics (CFD) computer code to simulate multicomponent multiphase flows of interest in research and industry. The NPHASE code is focused on developing a robust flow solver that is designed from the beginning to handle the complex interactions between numerics and multiphase flow modeling. Selected novel computational features and methods are discussed that distinguish NPHASE from other existing CFD codes. Also, the results of calculations for multidimensional two-phase flows are shown, including comparisons between NPHASE and two other CFD codes: CFX and Fluent. solved in coupled or uncoupled fashion, using frozen coefficient linearizations. The code is fully unstructured and utilizes second-order accurate convection and diffusion discretizations. A key feature of NPHASE is that from the outset the software design has focused on the development of a reliable solver for multiphase flows. Also, the various interactions between the individual fields have been incorporated as an inherent part of the solution algorithm. Thus, the user has available various mathematical formulations of the interfacial models. Only algebraic information for a given situation (such as experimentally-based coefficients, for instance) may require specification via input files. Key features of NPHASE include the following: Use of unstructured grids with arbitrary element types Capability to model an arbitrary number of fields (fluid components and/or phases) Built-in mechanistic modeling, integrated with numerics Improved robustness and numerical convergence Free surface modeling Major characteristics of the NPHASE code are discussed in the following sections.
1. Introduction
Numerous CFD codes have been developed to date, including both commercial codes and special purpose codes. Most commercial codes are multipurpose solvers that encompass a broad range of applications. Furthermore, these codes have been originated as computational tools for multidimensional single-phase problems. Thus, the early versions of general-purpose CFD codes typically included only basic models for two-phase flows. Whereas steady progress has been made resulting in the addition of new models to the existing solvers, more advanced models are still needed to adequately predict complex two-phase flow phenomena over a broad range of conditions. Because of the analytical form of such models, encoding them as a superposition to the already existing numerical solvers (e.g., via user routines) may cause substantial computational difficulties.
2. Overview of NPHASE
NPHASE is a segregated, nominally pressure based code. Individual scalar transport equations are solved for momentum, energy and turbulence quantities for each field. Mixture and field continuity equations are
Currently, the front end supports hexahedron, tetrahedron, prism and pyramid elements. Grid generators that have been used to date in NPHASE simulations included Gridpro [2] for multiblock structured grids and Gridgen [3] to build hybrid meshes and specify boundary conditions. Examples illustrating the use of unstructured grids by the NPHASE code are shown in Figures 1 and 2. Figure 1(a) presents a 189,149 element mesh (108561 tetrahedra, 76940 prisms, 3648 pyramids) grid for flow through a T-junction of two pipes with an internal pole obstruction. The prisms are extruded from surface triangles which define the pipe and pole boundaries. Figure 1(b) shows axial velocity contours predicted by NPHASE.
(along machine axis upstream of nose and in tip clearance region) were employed for this mesh. The calculated velocity vectors for turbulent flow conditions are as shown in Figure 2(b).
Figure 2. Hybrid mesh and predicted velocity vectors for propulsor simulation.
Figure 1. Hybrid mesh and predicted axial velocity contours for pipe junction with an obstruction. Figure 2(a) shows a hybrid grid for flow through a marine propulsor. 133,476 hexahedra and 53 prisms
uid flow shows a strong recirculation pattern near the intersection of the branch and run. .
An important predictive measure for this type of flows is the relative flow split between the run (i.e., flow that travels straight through the Tee) and flow through the branch (i.e., flow that makes the 90o bend). For the ten field case, the flow split of each gas field (i.e, characterized by a specified bubble diameter) is shown in Figure 3. This figure shows that the smaller bubbles mostly flow through the run while the larger bubbles divide more equally between the branch and the run. In fact, for this case, only about 10% of the liquid exits through the branch. Therefore, the 90o Tee is effective for separating the vapor and liquid phases.
0.6
Figure 4. Phasic Flow Split for Ten Field Flow in a 90o Tee.
* * * * * * * * * Figure 5. Liquid Velocity Vectors for Ten Field Flow in a 90 Degree Tee. Figure 6 shows the relative velocity between Field-7 and the liquid Field. This figure shows that large bubbles tend to flow through the branch. It is interesting to note, the even downstream of the intersection, the bubbles are still attracted toward the branch. This is due to the increasing pressure field which slows the bubbles faster than the liquid phase.
Field Number Figure 3. Phasic mass flow Split for a Ten-Field bubbly flow in a 90o Tee (each field represents a different bubble diameter). The vapor fraction profiles and liquid velocity vectors are shown in Figures 4 and 5. As can be seen, the liq-
Figure 6. Relative velocity between Field-7 (bubbles 1.0 mm in diameter) and the Liquid Field for the Ten-Field Flow in a 90o Tee. The pressure field and turbulence quantities are shown in Figure 7. It is also demonstrated in this figure that the NPHASE code insures that the sum of volume fractions over all the fields is almost exactly equal to one everywhere in the solution. This is accomplished through a coefficient normalization during iterations.
(1)
where
r k k k
is the volume fraction of phase k, is the density of phase k, is the velocity of phase k, is the diffusivity of phase k, is the inter-phase exchange coefficient between phases k and l, is the number of phases (or fields),
c kl
(Pa)
(m2/s2)
Pressure Profile
force terms. For various multiphase flow and heat transfer problems the Non-Drag interfacial force terms are needed to accurately predict the volume fraction distribution [4]. Two interfacial forces that are typically included as a very minimum in any multifield model of bubbly flow are the lift force and the turbulent dispersion force. The forces exerted between a continuous liquid field (cl) and a dispersed gas field (dv) can be written as [5] Lift Force
(m2/s3)
L L = M = C ( u u ) u L dv l dv dv cl cl cl
(2)
(3)
where C L and C TD are the lift and turbulent dispersion coefficients, respectively, dv is the volume fracLocal Vapor Fraction tion of the dispersed gas field, l is the liquid density, u is the field velocity, and k cl is the turbulent kinetic energy of the continuous liquid field. As previously shown by Kunz et al. [6], simply adding these terms to the RHS of the momentum equation can lead to numerical oscillations in the volume fraction profiles. A test case to demonstrate this behavior is show in Figures 8 and 9. This case under consideration here is developing turbulent bubbly flow between parallel walls. The numerical oscillations are clearly seen in the volume fraction profiles. To eliminate these oscillations, the turbulent dispersion model is included in the velocity interpolation for the faces of the grid. The velocity interpolation is performed according to
U face = U + B vel ( p p ) + A vel ( )
Distance Between Walls (m) Figure 9. Vapor fraction profile near the exit of a duct. Oscillations are due to lack of numerical coupling.
(4)
The first two terms on the right-hand-side are the standard Rhie-Chow [7] interpolation method. The purpose of the third term [6], is to include the impact of turbulence dispersion on the velocity interpolation.
wall Figure 10. Volume fraction profile for developing turbulent bubbly flow between parallel walls. Oscillations are eliminated through improved numerical coupling.
wall Figure 8. Volume fraction profile for developing turbulent bubbly flow between parallel walls. Vapor fraction oscillation are due to lack of numerical coupling. The case shown in Figures 9 and 10 was re-run using the NPHASE code with the velocity interpolation given by Eq.(4). The results are shown in Figures 10 and 11. As seen in the calculated volume fraction distribution, the profiles are smooth and more accurately reflect the mechanistic modeling for bubbly flows. This treatment is applied to each turbulence dispersion term for two or more fields.
In addition, an algebraic multigrid solver (AMG) is used for the linear matrix solution. The NPHASE code uses a segregated algorithm to solve the governing conservation equations. In this approach, under-relaxation is generally employed to ensure strong diagonal dominance in the resulting algebraic equations. The exception is the pressure (corrector) equation. In the segregated algorithm, the pressure corrector is a multidimensional diffusion equation whose source term is proportional to the sum of the phasic mass imbalances for each point in the grid. Under-relaxation is not employed for the pressure corrector equation in order to achieve a measure of mass conservation during each nonlinear iteration. As as result, the pressure corrector equation typically meets the minimum threshold for diagonal dominance
a p = a nb
(7)
l=1
ckll clk k + Sk
Np Np l=1
(5)
where S l contains all the remaining terms in the discretization. For flows with several fields, the interfield coupling can be very complicated. The importance of this coupling can vary depending on the size of the coefficients in the interfield coupling terms. For large coupling terms, the tighter the numerical coupling, the better the numerical stability. By simple algebraic manipulation, a multifield PEA scheme can be written as
It should be noted that the solution of the pressure corrector equation yields exact mass conservation in single phase flows and conserves mass in an aggregate sense for multiphase flows. The linear solution of the pressure corrector equation is a key factor in the numerical performance of the overall code. For this application, NPHASE utilizes an algebraic multigrid (AMG) solver. Multigrid technology has gained wide acceptance within the CFD community. The basic idea is to represent the error of the solution on successively coarser grids in order to eliminate the long wavelength components. For AMG, the representation on coarse grids is done by agglomerating fine grid nodes and manipulating the corresponding fine grid matrix coefficients and source terms [8]. Coarse grid nodes are comprised of an integral number of fine grid cells. For each coarse grid, a correction term is computed such that the following condition is satisfied
R i fine = 0
(8)
Np
Np
where the summation involves all the fine grid node residuals that comprise a given coarse grid cell. The coarse grid correction is added to each of the associated fine grid nodes.
l=1
Np
( c Ml M ) c lk k + S l
(6)
c Ml
6.2. Implementation
The AMG solver is implemented as an object oriented design using C++. The implementation makes use of the containers and algorithms from the Standard Template Library (STL). The class diagram for AMG solver implementation is shown in Figure 14.
grid_parms Agglomerated Cell Figure 12. Isotropic Coarsening of Grid. CFD calculations often employ stretched grids near walls for example to capture gradients in the flow field. Numerically, the high aspect ratio characteristic of these cells can lead to long wavelength error components in the stretching direction. To address this situation the AMG implementation in the NPHASE code includes a directional coarsening capability based on the cell facial area. This is illustrated in Figure 13. SquareMatrix solver_method
ILU
GS
Matrix_eqn
ICCG
Figure 14. A class diagram for AMG solver. The linear solution for each grid level is governed by the solver_method class. This is implemented as an abstract base class. Three linear solver classes are derived from solver_method: ILU, Gauss-Seidel (GS) and Incomplete Choleski Conjugate Gradient (ICCG). The particular solver chosen is invoked polymorphically. At present, the ILU and GS solvers are utilized with the multigrid scheme; ICCG is restricted to fine grid solutions. In NPHASE, the multigrid solution follows a fixed Vcycle. For a typical problem the number of coarse grid levels is chosen such that the coarsest grid is comprised of ~20 cells. Grid_parms, SquareMatrix and matrix_eqn are stand-alone classes that are used compositionally. Coarsening is handled by the grid_parms class. Experience has shown that for some flow fields the longest wavelength errors as represented on the coarsest grid must be fully eliminated. For these cases the AMG implementation allows a direct solution on the coarsest grid. The SquareMatrix class creates a square matrix representation of the coefficients on the coarsest grid to facilitate direct solution. Coefficients and source terms for all grid levels are housed within the matrix_eqn class. Finally, Figure 15 shows linear solver performance for a single outer loop iteration for typical flow field. The AMG-ILU solver is compared against the ICCG and ILU methods. In addition, the AMG-ILU uses either an iterative or direct solve on the coarsest grid. The results show that the AMG-ILU performance best for the least number of sweeps.
Agglomerated Cell Figure 13.Directional Coarsening of Grid Typically, directional coarsening is implemented based on the relative size of the matrix coefficients with the coarsening performed every iteration. The coarsening strategy used in NPHASE is a more basic approach and is done only once at the outset of the computation. As a final item, the work done to coarsen the grid increases with the use of directional coarsening. This is consistent with other observations [9]. For directional coarsening, the problem complexity relative to the fine grid solution is approximately given by
1 1 1 1 + -- + -- + -- + 2 - - 2 4 8
(9)
(10)
7
ICCG solver ILU solver iccgAMG-ILU iterative solver AMG-ILU direct Summed Residual Error
tum equations. Without this treatment, numerical convergence was very difficult or impossible for many cases. A free surface comparison between the CFX 4.2 [12], Fluent 4.4 [13] and NPHASE code are shown in Figure 16. This case simulates a two phase discharge into a
CFX
Fluent
Figure 16. Comparison of three CFD codes to predict two phase sparger flows in an open pool. large pool. The top boundary simulates the pools surface where the vapor can escape but the liquid is forced to recirculate. Since neither CFX 4.2 nor Fluent 4.4 have explicit free surface models incorporated into the multifield equation solver, the user routines have been used to approximate the actual physical situation near the pool surface. As can be seen in Figure 16, all three codes show similar trends, and the NPHASE predicted void fraction distribution is between those of CFX 4.2 and Fluent 4.4. However, considerable differences have been observed in flow and phase distributions in the upper parts of the pool. The major apparent reason for these differences was the different treatment of the free surface by each code, which, among other factors, affected the recirculating flow within the pool. To evaluate the impact of the (different) numerical approximations of the actual free surface conditions in the CFX 4.2 and Fluent 4.4 codes, on the local prediction of flow near the sparger, the effect of pool height was investigated. In particular, the depth of the pool was doubled and the cases were rerun. As can be seen in Figure 17, the differences between the results of the three codes are smaller than before, especially near the
(11)
where, X , is the distance from the node to the free surface. The local mixture density, mixture , and gravity, g, are also used to extrapolate the surface pressure. The overall change in pressure is small but this treatment was found to be important to balance the momen-
sparger. Some differences are still observed near the pools surface and in the recirculation flows within the pool.
RMS Change in Variable
0.01
0.001
0.0001
1e-005
1e-006
CFX
Fluent
Nphase
1e-007
Iteration Count
Figure 18. Convergence history for NPHASE for two phase flow analysis of sparger flows.
Free Surface
The convergence history for the NPHASE code is show in Figure 18 for the sparger analysis. These recirculating, multiphase flow analyses are very challenging to converge. The free surface modeling can also be used to predict bubble column (i.e., fluid-solid-vapor flows) performance. Figures 19 and 20 show the phasic velocity and volume fraction for a three field (liquid-solid-vapor) prediction. In this analysis the inflow is a small jet of all vapor on the bottom of the bubble column. The solid material are very small diameter particles with twice the density of the liquid. Both the solid and liquid are forced to remain with the column while the vapor can free escape through the pool surface. The velocity plots show the vapor having the highest velocity near the allvapor inlet. The liquid (and solids) are pulled by the vapor and form an effective recirculation pattern in the column. The volume fraction contours show that the solids are not uniformly distributed within the column. This nonuniformity can have a significant impact on the efficiency of the chemical process within the column.
Tank Centerline
Figure 17. Comparison of three CFD codes to predict two phase sparger flows in a deep pool.
Fluid Phase
Solid
Phase
Figure 19. NPHASE Phasic Velocity Plots for Bubble Column Three Field Flows.
References
1. FIELDVIEW 6 Reference Manual, Intelligent Light, 1999. 2. Gridpro User Manual, Program Development Corp.,1998. 3. Gridgen User Manual, Version 13.3, Pointwise, 1999. 4. Kurul, N. and Podowski, M.Z., On The Modeling of Multidimensional Effects in Boiling Channels, ANS Proc. 27th National Heat Transfer Conference, Minneapolis, MN, 1991. 5. Drew, DA, Analytical Modeling of Multiphase Flows, Boiling Heat Transfer Modern Development and Advances, Elsevier Science Publishers, 1992.
Fluid Phase
Solid Phase
6. Kunz, R.F., Siebert, B.W., Cope, W.K., Foster, N.F., Antal, S.P., Ettorre, S.M., A Coupled Phasic Exchange Algorithm for Multi-Dimensional FourField Analysis of Heated Flows with Mass Transfer, Computers and Fluids, 27(7):741-768, 1998. 7. Rhie, C. M., Chow, W. L., Numerical Study of the Turbulent Flow Past an Airfoil with Trailing Edge Separation, AIAA Journal, Vol. 21, p.1527, 1983. 8. Raw M, Robustness of Coupled Algebraic Multigrid for the Navier-Stokes Equations, AIAA 960297, 1996 9 Mavriplis DJ, Multigrid Techniques for Unstructured Meshes, ICASE Report, 1995. 10. Antal, S.P., Nagrath, S. and Podowski, M.Z., Development of Mechanistic Modeling Capabilities for Computer Simulations of Reactor Components and Systems, 2nd Annual Report for USNRC, RPI, 2000. 11. Antal, S.P., Nagrath, S. and Podowski, M.Z., Multidimensional Simulations of Two-phase Flow in Large Volumes with Spargers, Proc. International Conference on Multiphase Flow, ICMF 2001, New Orleans, LA, 2001 (to be published). 12. CFX 4.2 Flow Solver Users Guide, AEA Technology, Computational Fluid Dynamic Services, 1998. 13. FLUENT 4.4 User Guide, 1997.
Figure 20. NPHASE Phasic Velocity Plots for Bubble Column Three Field Flows.
8. Summary
This paper is concerned with the development of NPHASE, a next generation computational fluid dynamics (CFD) computer code to simulate multicomponent multiphase flows of interest in research and industry. The NPHASE code is focused on developing a robust flow solver that is designed from the beginning to handle the complex interactions between numerics and multiphase flow modeling. A key feature of NPHASE is that from the outset the software design has focused on the development of a reliable solver for multiphase flows. Also, the various interactions between the individual fields have been incorporated as an inherent part of the solution algorithm. Key features of NPHASE include the following: Use of unstructured grids with arbitrary element types Capability to model an arbitrary number of fields (fluid components and/or phases) Built-in mechanistic modeling integrated with numerics Improved robustness and numerical convergence Free surface modeling Each of these features was discussed along with several example cases.
10