Process Metallurgy

CFD Benchmark for a Single Strand Tundish (Part I)

SMS Demag AG, R&D Division, Dsseldorf, Germany Johannes Kepler University Linz, Institute for Fluid Dynamics and Heat Transfer, Linz, Austria 3) RHI AG, Technology Center, Computer Modelling and Simulation, Leoben, Austria A Computational Fluid Dynamic (CFD) benchmark for the water model of a single-strand continuous casting tundish was performed by ten members of the newly founded working group "Fluid Mechanics and Fluid Simulation" of the German Steel Institute VDEh. A critical comparison is drawn between laser-optical velocity measurements and residence time measurements on the one hand and CFD simulations using different CFD programs, turbulence models, boundary conditions, proposed solutions, etc., on the other hand. The validation criteria used include, among others, the turbulence distribution, the position of the recirculation center and the maximum backflow velocity in the tundish which is induced by the recirculation, as well as the residence time distribution. The results show that the flow and turbulence structure can be computed on the basis of the Unsteady Reynolds averaged Navier-Stokes (URANS) equations with a good degree of accuracy. The relative positional deviation of the recirculation center is -12.5% < x/L1 < 5.0%. The characteristic times min, max, 20% and 5% , which describe the residence time distribution, are established with a variation of 15%. The benchmark yields important results for the sensible use of today's commonly used numerical CFD models and contributes to further improving the reliability of CFD simulations in metallurgical process engineering. DOI: 10.2374/SRI08SP163; submitted on 18 October 2008, accepted on 27 January 2009

Introduction CFD simulations are an indispensable tool in the design of steelmaking plant and equipment and in the development of new technologies. They serve to improve the understanding of metallurgical processes while increasing the operational reliability and availability of a plant. The relevant questions almost always concern turbulent, time-variable and additionally hot, multi-phase flows of molten metal and/or gases whose simulation is extremely complex and requires an intensive computational effort. In contrast to flow-related applications in the design and construction of aircraft, automobiles or turbines, the majority of metallurgical plant and equipment is not flow-optimized, since the focus here is on robust, affordable structures and not on minimal pressure losses or non-separated flows. This gives rise to flow phenomena representing an extreme challenge to turbulence modeling. Computational domains involving several million cells can be handled by today's standard computers. During the past 15 years alone, the size of typical CFD problems rose by a factor of 20. Nonetheless, even with a further rise in computing capacity it will take another 20 years before complex turbulent flows for technical applications can largely be exactly solved also by the numerical method of Direct Numerical Simulation (DNS). Owing to the simplifying character of the available models and the lack of material data at high temperatures, some limits presently still exist for simulations of metallurgical processes. In addition, the classical parallelization by the decomposition of the computational domain, as employed, for example, in formula 1 racing technology, will not always result in a clear reduction of the computing time in metallurgical applications. This is due to the fact that the effectiveness in the parallelization
264

of many connected chemical-physical models is much worse for small grids than for large grids comprising few, yet simple models. In metallurgical flows, the coupling of solutions in partial computational domains will also in the foreseeable future be a limiting factor in the consideration of all relevant aspects. Moreover, the question of the reliability of the numerical results always arises. Turbulence effects strongly affect the fluid flow, but can only be modeled by semi-empiric approaches. As regards their validity, only vague criteria exist, so that especially in the case of complex flows only a comparison with measured data will yield information about the usability of the CFD results. On the one hand, a multitude of empiric turbulence models are available in CFD programs, on the other hand, many numerical parameters also exist which not only influence convergence but also the result. The Committee on Metallurgical Fundamentals of the German Steel Institute VDEh has set up the working group "Fluid Mechanics and Fluid Simulation" in 2007. This group is to critically scrutinize CFD simulation, which in recent years has been expanding heavily into the field of metallurgy, in terms both of the approaches in fluid physics as well as the solution methods of the equation systems set up, and to assess its potential for the development of metallurgical process engineering. To this end the current methods of physical simulation and their benefit in the validation of CFD results are also assessed. The acceptance of metallurgical fluid mechanics and in particular of CFD is to be promoted and the corresponding system expertise in metallurgy is to be disseminated further. The working group currently comprises around 30 members from 11 industrial companies and 10 research institutes. Examples of current topics include turbulence modeling, electromagnetic flow measurement and flow control, CFD simulation of the heat and material transport
steel research int. 80 (2009) No. 4

Process Metallurgy

in induction systems, modeling in disperse multiphase flows or the modeling of macro-segregations in metallurgical processes. The following presents results of a CFD benchmark study on the water model of a continuous-casting tundish. The factors that influence the accuracy of a CFD simulation that are discussed in the next paragraph demonstrate the urgency with which a benchmark is needed in the field of metallurgy. The tundish was chosen because on the one hand the dominating flow structures are readily comprehensible and on the other hand reliable experimental data are available. Ten members of the working group, six from industrial companies (ANSYS Germany; Corus - R&D Technology Depart-ment; RHI AG - Technology Center; SMS Demag - R&D Division (two participants); ThyssenKrupp Steel -Department for Physical Technology) and four from universities (Johannes Kepler University Linz - Institute for Fluid Mechanics and Heat Transfer; RWTH Aachen University - Institute for Industrial Furnaces and Heat Engineering; Martin-LutherUniversity Halle-Wittenberg -Institute for Mechanical Process Engineering; University Bergakademie Freiberg Institute for Mechanics and Fluid Dynamics), took part in the benchmark. The intention here was to perform the CFD simulation as blind test for which only the tundish geometry, the volumetric flow-rate at the shroud and SEN (Submerged Entry Nozzle) as well as the number of 500,000 grid cells were defined. All other solution strategies and parameters were to be chosen freely by the participants. In addition to the flow structures, the RTD (Residence Time Distribution) curve was also to be calculated. After this, the results were to be compared with each other and the models assessed critically. Factors Influencing the Accuracy of CFD Simulations Owing to the large number and the complex way in which the influencing factors of a CFD simulation are interlaced, it is up to the user's experience and intuition to prepare a good grid and make the best possible settings. Best practice guidelines such as those of ERCOFTAC [3] may provide helpful recommendations in many areas. The influencing factors arise in various levels of the simulation procedure. The uppermost level of influence is represented by the mathematical models which describe the flow field by partial differential equations. These are the Reynolds averaged Navier-Stokes (RANS) equations in combination with a turbulence model. Since all turbulence models are of semi-empiric nature, even an analytic, i.e. exact solution of these equations would not exactly reproduce the real flow field in almost any case. The turbulence models are classified by the number of turbulence equations. The most widespread models belong to the k- class as a subset of the two-equation models. They are based on the eddy viscosity hypothesis as formulated by Boussinesq in 1877, i.e., the effect of the turbulence in the Reynolds averaged flow field is taken into account as turbulence-dependent, increased viscosity. The best-known model of this class is the standard k- model [10] as presented by Launder and Spalding in 1972.
steel research int. 80 (2009) No. 4

Since non-physical, negative Reynolds normal stresses may arise here, it is today increasingly replaced by the realizable k- model [18]. Increasingly popular are the two-equation models of the k- class. In contrast to the k- models, no modeling of the near-wall turbulence is required. Unfortunately, the model responds sensitively to the turbulence data of the boundary conditions at the inlet. The Shear Stress Transport (SST) k- model combines the k- and k- models, with a blending made from one model to the other depending on the wall distance. This is to combine the benefits of both approaches [7]. Even more elaborate are Reynolds Stress Models (RSM) which are not based on the Boussinesq hypothesis and in which seven equations for the components of the Reynolds stress tensor and the or equation are solved. This allows certain flow effects to be reproduced significantly better than in two-equation models [8,9]. Still the RSM is based on empirical assumptions and may therefore not be regarded as universal. One method that is now being increasingly used is the Large Eddy Simulation (LES). This provides for the large, anisotropic turbulence eddies, which transport the main amount of mass, momentum and energy, to be computed directly and without any simplification. A simple, algebraic model takes into account the interaction between the small, isotropic turbulence eddies which are not dissolved by the grid and the large, anisotropic eddies [14]. Those simulations require fine grids and small time-step sizes. The results usually depend very much on the grid and on the boundary conditions [6]. While all RANS models describe turbulent mixing as an isotropic diffusion process, LES simulations are capable of resolving anisotropic mixing due to large turbulence structures and thus of reproducing mixing and the resulting RTD curve more accurately. The second level of influence is discretization, i.e., the transformation of the partial differential equations into algebraic equations by the respective approximation of space and time variables. The grid cells must be small enough to sufficiently resolve local flow effects such as those prevailing in the free shear layer below the shroud so that it is the skilful local adaption of the cell size that matters rather than the total number of grid cells. This can be realized by local grid refinement (grid adaptation). In the mostly unstructured grids, the cell types used (hexahedrons, tetrahedrons) may vary strongly from the ideal, regular shape (skewness), which has a negative effect. Thus, special criteria are used to assess the cell quality. The Finite Volume Method (FVM), which is predominantly used for CFD, interpolates the values at the cell surfaces from those at the cell center. Depending on the number of interpolation points, a distinction is made between 1st order, 2nd order and higher-order upwind methods (e.g. QUICK scheme). A common property of all upwind methods is that they cause an artificial, so-called numerical viscosity and diffusion, respectively. Although the 1st order upwind method may cause significant errors, it is quite popular due to its damping and stabilizing effect. The chosen upwind method and the formulation of the spatial derivations, i.e., cell or node based gradients, thus have a direct influence on the computed flow quantities.
265

Process Metallurgy

Since the discretized equations are solved with floatingpoint numbers of finite accuracy the used number accuracy represents a third level of influence. Most CFD programs provide both single- and double-precision solver, i.e., the resolvable difference in size is about 1:107 or 1:1014, whereas it is recommended to generally use the doubleprecision solver [3]. The last level of influence refers to the iterative solution of the discretized equation system. Ideally, the numerical solution should approximate the real solution with increasing iteration steps. The damping of the iterative process by introducing underrelaxation factors is necessary to achieve and accelerate the convergence of the iterative process. The proper adjustment of the underrelaxation factors and the monitoring of both residuals and characteristic flow variables are important and should always be used as an indicator for the quality of the solution. In the best case only the convergence speed is decreased due to non-optimum numerical settings. In the worst case the computation diverges or no stable solution is produced. In the latter case a popular approach is to change from a steady to an unsteady formulation. Owing to the non-linearity of the governing equations, this approach may be meaningful despite the steady-state boundary conditions. Under the term Unsteady Reynolds Averaged Navier-Stokes (URANS), the transient simulation is increasingly establishing itself in practice, although to be correct, the time-dependency of the solution needs to be considered in the evaluation. Further approaches include, among others: - Sequential or simultaneous solution of equation systems (segregated/coupled solver), - Mandatory incorporation of continuity equation in the iterative process (pressure-velocity coupling), - Cyclic transition to coarser grids in order to improve convergence (multigrid option).

Boundary Conditions of the Tundish CFD Benchmark The term benchmark refers to a comparative analysis with a defined reference value. Numeric simulations use operating data or the results of physical simulations for this purpose. The subject of the present CFD benchmark is the water model (scale 1:1.7) of a stopper-rod-controlled, steady-state operated 16-ton single-strand tundish used for the production of stainless steels. The shroud and SEN are each approx. 1.0 m long. Table 1 shows the predefined process data of the original tundish and of the water model. More information on the geometry can be found in [1,2,12]. The reference values used for the benchmark are results of residence time and three-dimensional Laser Doppler Anemometry (LDA) measurements. The RTD curve for steady-state operation was found by dye injection through a pulse input. The response signal was detected in the SEN by means of a photocell and the test repeated three times for the purpose of averaging. At the time tmin a rise of concentration was first noted on the shroud, at time tmax the maximum concentration was detected. The 3D LDA measurements were made in two xz longitudinal planes of the water model. To this end 90 x 25 measuring points were scanned in each plane, with N = 2000 samples detected at each measuring point. From the components of the mean velocity u i and the velocity of fluctuation u i ' the local turbulence intensity is computed by

Tu =

ui '2 3 ui
2

2 3

at i = 1 to 3.

(1).

ui

Table 1. Process data of the single strand tundish. Notation Liquid flow volume in the tundish Length of the tundish bottom Width of the tundish bottom Tundish filling level for steady-state casting condition Length of the shroud Inner diameter of the shroud Length of the SEN Inner diameter of the SEN Mass flow rate during steady-state casting Mean flow velocity inside the shroud Theoretical mean-flow velocity through the tundish Maximum back-flow velocity in the tundish Theoretical residence time of the fluid in the tundish Some characteristic residence times Reynolds number Froude number Symbol V in m
3

Original 2.275 3.140 0.780 0.800 1.0 0.068 1.0 0.070 38.0 1.49 0.008 420 10380 2.310-3

Model 0.463 1.847 0.459 0.471 1.0 0.040 1.0 0.040 3.68 2.92 0.015 0.07 126 5, 60, 225, 345 10380 5.810-3

L1 in m B1 in m H in m Lsh in m Dsh in m LSEN in m DSEN in m

& msh , SEN in kg/s

u in m/s
u in m/s ttheo in s tmin, tmax , t20%, t5% in s Re Fr

266

steel research int. 80 (2009) No. 4

Process Metallurgy

This is the ratio of turbulent kinetic energy k = 1 ui '2 and 2 mean kinetic energy of the main flow. The measuring system has an error whose amount can be estimated at approx. 5% [12]. It has to be noted that the u-component in x- and the w-component in z-direction, due to the relatively small LDA measuring volume, were established with greater accuracy than the v-component in y-direction. The greatest uncertainty, however, is in the fabrication tolerances of the water model and the alignment of the shroud. Even minute alignment errors induce an asymmetric flow. This had been already recognizable in the measured data as a flow transverse to the symmetry xz-plane. The relative error in the mean flow velocity due to the finite number of individual measurements can be computed from the standard deviation i of the measured velocities. With a probability of 95.5% the relative error of the averaged measured velocity is

a) Central section (y/B1 = 0)

b) Near sidewall section (y/B1 = -0.5) Figure 1. 3D-LDA velocity and turbulence intensity distribution in the water model tundish (scale 1:1.7),
& & V sh = V SEN = 3.68 l/s, Re = 10380, Fr = 5.81 10 3

u i
ui

<

2 i N ui

with i = u i ' 2

(2) Results Table 2 (see p. 274) provides an overview of the CFD parameters that were freely selected by the working group members. Although the tundish has a vertical symmetry plane, all participants modeled the full geometry. This suggests that today the available computing capacity for grids with a size of around 1 to 2 million cells no longer presents any restriction. Gambit (5x), ICEM CFD (4x) und CFX-Mesh (1x) were used for the generation of the grid. Despite the specified requirement, the number of grid cells varied between 384,000 and 660,000. User 6 neglected the geometry of the stopper rod. To check the grid quality, commonly used criteria were applied such as aspect ratio, equiangle skewness, equisize skewness, etc. [5]. The CFD software applied was FLUENT (6x), OpenFOAM (2x), CFX (1x) and Fastest3D (1x), the latter being a proprietary development from a university environment. One half of the users employed a doubleprecision solver for the solution of the equation system. Differences were found mainly in the selection of a steady or unsteady solution procedure, the turbulence model, the wall as well as the free surface treatment and the discretization method. It is these parameters that have the greatest influence on the result. For the discretization the 1st order upwind method was often selected. For the turbulence equations it might still be argued that these are mainly source-term influenced equations and therefore the convective terms, i.e., the terms affected by the discretization scheme, have no influence. A greater influence may be expected to arise in the momentum equation and in scalar transport [6]. The length of the shroud had been specified as 1 m. However, the majority of users modeled the shroud and SEN arbitrarily with lengths between 0 and 1 m. For the inlet boundary
267

This locally variable error still needs to be added to the error of the measuring system. Figure 1 shows the results for the steady-state casting & & operation at Vsh = VSEN = 3.68 l / s . The center of the recirculation area is located at x/L1 = 0.52 and z/H = 0.20. The maximum backflow velocity in the tundish is above this position approx. at x/L1 = 0.49 and z/H = 0.39 and is u = 0.07 m/s. A short-circuit flow prevails along the sidewalls. The maximum component here is u 0.2 m/s. The shroud flow arising produces a double-vortex system which prevails all the way through the tundish and whose intensity decreases with rising x-coordinate [1,2,12]. The turbulence intensity Tu rises with increasing tundish length because the mean velocity is reduced to a larger extent than the turbulent fluctuation. Significant turbulence can be found in regions of flowmechanical instabilities such as inside the shear layers of the recirculation area. A turbulent area also exists at the center of the tundish below the free surface, where the stagnation line of the double-vortex system rotating in opposite directions is located. There, the fluid material of these two vortex-rolls meets and is diverted into the interior of the tundish, i.e., downward. High values of turbulence intensity are found in the vortex to the left of the shroud as well. The (x,y,z) position of the center of the recirculation region and the maximum backflow velocity serve as a criterion of quantitative validation for the CFD simulations.
steel research int. 80 (2009) No. 4

Process Metallurgy
Central section (y/B1 = 0) Near sidewall section (y/B1 = -0.5)

User 2: FLUENT, unsteady, RSM, enhanced wall treatment, QUICK scheme

User 3: Fastest3d/Lag3d, steady, standard k- model, standard wall function, 2nd order upwind scheme

User 5: CFX, steady, SST k- model, automatic near wall treatment, 2nd order upwind scheme

User 8: OpenFOAM, steady, realizable k- model, standard wall function, 2nd order upwind scheme Figure 2. Fluid flow structure in the central (y/B1 = 0) and near side wall (y/B1 = -0.5) section.

User 5: CFX, unsteady, Smagorinsky-LES model, automatic wall treatment, central difference scheme. Figure 3. LES simulation of User 5: fluid flow structure in the central (y/B1 = 0) and near side wall (y/B1 = -0.5) section.

268

steel research int. 80 (2009) No. 4

Process Metallurgy

condition on the shroud, only the flow rate had been specified; all users chose a constant velocity profile (velocity or mass flow inlet). Likewise, the turbulence parameters at the inlet could be chosen arbitrarily as well, with the turbulence intensity Tu and hydraulic diameter Dhydr being the parameters that were used most. The turbulence intensity inside the shroud is very low in comparison to the turbulence generated in the free shear layers of the shroud jet [6]. The turbulence boundary conditions thus have almost no influence on the result as long as they are chosen in the right order of magnitude. For the free surface, either a symmetry condition, a no-slip wall or a frictionless wall was used as boundary condition, with the latter reflecting the physical conditions best. As regards the number of iterations and the convergence criteria, no information was regrettably available. In addition to the options presented in Table 2, furthergoing work was conducted by Users 5, 6, 8 and 9. With CFX, User 5 additionally carried out an LES simulation using the Smagorinsky model and around 4.67 million grid cells, which alone required a computing time of around 14 days. User 6 employs different turbulence models (standard k-, realizable k-, RSM) for OpenFOAM. User 8 took both FLUENT and OpenFOAM and reviewed the influence of the standard k- and of the realizable k- model. User 9 employed FLUENT and the realizable k- model as well as the RSM to investigate the influence of the wall conditions, e.g. standard wall and non-equilibrium wall. No user changed the standard constants of the respective turbulence model. Accordingly, it may be assumed that no experience-based information were available for the influence of these constants, but that users instead relied on the software maker's recommendations. As an example Figure 2 shows the flow structure predicted by the various CFD programs at the central section and at the side wall of the water model. Important program settings like the type of CFD software, steady or unsteady treatment of the flow problem, the turbulence model, the law of the wall and the discretization scheme of the momentum equation, are listed. All results were interpolated to the 90 x 25 measuring grid, more details may be obtained from Table 2. The velocity vectors are characterized by a quadratic color distribution, the position of the recirculation center is marked as well. In terms of quality, all flow structures computed, including those which are not shown here, agree well with those of the 3d LDA measurement from Figure 1. Merely the (x,y,z) position of the recirculation are varying from (0.735, 0, 0.100) m for User 7 and (1.065, 0, 0.095) m for user 9, while the measurement is (0.961, 0, 0.094) m. All previous results were based on (U)RANS solutions. Figure 3 shows the result of a more elaborate LES simulation whose computation alone took 14 days. This was generally a time-dependent simulation, however Figure 3 is no momentary image but shows the timeaveraged flow status, which is quite similar to that of the (U)RANS simulations. The benefit of LES is that relevant flow phenomena can be clearly better resolved in terms of space and time and therefore, for example, convective mixing (RTD curve) is covered with greater accuracy.
steel research int. 80 (2009) No. 4

By way of analogy, Figure 4 shows the Tu distribution according to Equation (1) in the central section. Although in terms of quality the calculated turbulence fields agree with the LDA measurement, major discrepancies can be recognized especially in the outlet region of the tundish. Those users (1,4,7,8,9) who took a symmetry boundary condition at the free surface, computed a clearly too high turbulence directly below the free surface, whereas users (2,3,5,6,10) who chose a frictionless wall, found a slightly lower turbulence than in the LDA measurements. The discrepancies between CFD simulated turbulence and LDA data were lower for the wall boundary condition of the virtual free surface. Figure 5 shows the Tu distribution calculated by user 5 with the LES model. Due to the large number of grid cells even fine turbulence structures, e.g., inside the shear layers of the shroud jet and the recirculation area, are clearly resolved and the consistency with Figure 1a is excellent. Figure 6 shows the relative positional deviation x/L1 and z/L1 of the vortex center with reference to the center measured by LDA; the data are given in percent. All users computed symmetric flow structures with the vortex center at y/L1 = 0. In horizontal direction, the relative positional variation is -12% < x/L1 < 5%, in vertical direction it is even as little as -2% < z/L1 < 1%. The choice of the CFD program, of the turbulence model and of the solution algorithms has only a minor influence on the position of the recirculation area, i.e., all CFD programs safely predict the vortex center. The most accurate results are supplied by User 2 (FLUENT, unsteady, RSM, standard wall), User 3 (Fastest3d/Lag3d, steady, standard k-, standard wall) and User 10 (FLUENT, steady, realizable k-, standard wall). There is a tendency for the position of the recirculation region to be predicted more accurately in simulations using a frictionless wall as boundary condition rather than a symmetry boundary condition. Further, what needs to be observed is that the accuracy in determining the position of the recirculation center depends on the size of the grid cells there. Where characteristic areas in the CFD simulation are to be investigated, the grid must be locally refined there. Figure 7 shows the relative deviation u/u of the computed maximum backflow velocity in the recirculation area, referred to the LDA measured value of u = 0.07 m/s. Using the measured standard variation of i = 0.032 m/s for Equation (2), the total measuring error of the LDA mean velocity value is around 7%. This error bandwidth is sketched in Figure 7 as a gray bar. For User 1 (FLUENT, steady, realizable k- standard wall), User 2 (FLUENT, unsteady, RSM, standard wall), User 4 (FLUENT, steady, realizable k- standard wall) and User 8 (OpenFOAM, steady, realizable k- standard wall) the variation is within the LDA error tolerance. Both the realizable k- model and the RSM obviously predict the separation line at the tundish bottom and the associated backflow structure better than the standard k- model. The selection of a higher-order discretization method, too, is needed for the momentum equation. Since the computed backflow velocity at a certain point can be determined only with uncertainty, the mean
269

Process Metallurgy

User 1: FLUENT, steady, realizable k- model, stand. wall, symmetry

User 2: FLUENT, unsteady, RSM, enhanced wall, wall shear = 0

User 3: Fastest3d/Lag3d, steady, standard k- model, stand. wall, wall shear = 0

User 4: FLUENT, steady, realizable k- model, stand. wall, symmetry

User 5: CFX, steady, SST k-, automatic wall, wall shear = 0

User 6: OpenFOAM, steady, RSM, standard wall, wall shear = 0

User 7: FLUENT, steady, RSM, enhanced wall, symmetry

User 8: OpenFOAM, steady, realizable k- model, stand. wall, symmetry

User 9: FLUENT, steady, RSM, non-equilibrium wall, symmetry Figure 4. Turbulence intensity Tu in the central section (y/B1 = 0).

User 10: FLUENT, steady, realizable k- model, stand. wall, wall shear= 0

270

steel research int. 80 (2009) No. 4

Process Metallurgy

horizontal u-velocity profile is plotted in Figure 8 versus the vertical z-component. For each user, the velocity profile was evaluated on a vertical cut-section through the center of the corresponding recirculation area. In some tundish regions major deviations between the numerical and the experimental data can be seen. The URANS simulation of User 2 reflects the 3D LDA measurement best. In the case of Users (1,4,7,8,9), who selected a symmetry boundary condition at the surface, velocities below the free surface are too high, whereas in the case of Users (2,3,5,6,10), who used a frictionless wall, velocities are clearly closer to the measured values. The reason for this may also be that due to the different boundary conditions, the x-position of the respective velocity profile varied between users. Figure 9 contains the comparison between the measured and computed RTD curves. The current time t has been normalized with the theoretical residence time & ttheo = V / msh = 126 s ( = t / ttheo) and the local concentration c with the maximum concentration c*. The validation variables used are the minimum residence time min and the maximum concentration time max as well as the two points in time 20% and 5%, at which the RTD curve dropped to 20% or 5%, respectively, of its maximum level. Particularly min and max are variables which in chemical reactor engineering are used to define the plug flow, the mixing flow and the dead-water flow volumes and which are often regarded as a measure for a flow-optimized design [11]. User 6 did not carry out an RTD simulation. Besides, the RTD curve shown by User 5 is from the LES simulation. All in all there is good agreement between the CFD curves and the measured values. Figure 10 shows the deviation of the dimension-less points in time min, max, 20% and 5% in the general form of

Figure 5. LES simulation of User 5: Turbulence intensity Tu in the central section (y/B1 = 0).

Figure 6. Relative positional deviation x/L1 and z/L1 of the recirculation centre (y/L1 = 0).

min = min,CFD min,LDA .

(3)

While the minimum residence time min is longer in almost all simulations than in the measurement, there are deviations in both directions for max and 20%. For time 5%, Figure 7. Relative deviation u/u of the maximum backflow velocity between a one-sided late attainment of the 5% level is CFD simulation and LDA measurement. then more likely to be computed again. There is obviously no recognizable correlation It is generally striking that especially the minimum between the deviations and the chosen CFD models. It is residence time min, which is preferably used for possible that the time-step size t used for the computation computing the plug-flow volume vplug, involves systematic, of the RTD curve influences the result due to numeric minor deviations in the CFD simulation. The more the diffusion, but t has regrettably not been acquired. Worth time progresses, the larger the inaccuracies become. This noting are the times predicted by User 5, which are based must be noted especially when computing the dead-flow on LES simulations and indicate minor deviations volume vdead at > 2 [1,2,16]. throughout.
steel research int. 80 (2009) No. 4 271

Process Metallurgy

influence of various solution strategies and CFD parameters on the flow structure in the water model of a single-strand tundish. From these studies, guidelines were derived for the influence of the CFD software, turbulence models (k- model, k- model, RSM, LES), near wall treatment, discretization methods, etc., on the calculated flow field. All commercial and self-developed CFD programs used in the benchmark predicted the global flow and turbulence structure inside the tundish and, based on it, also the RTD curve with a good degree of accuracy. (U)RANS simulations generally already provided good approximations of the flow in the water-model tundish that had been measured by laserFigure 8. Horizontal u velocity component in a vertical section through the optical equipment. The LES simulation, which center of the circulation region. clearly required more computing time, produced a more accurate distribution of the turbulence and a slightly higher accuracy in the RTD curve. The choice of the boundary condition for the free surface, i.e., the use of a frictionless wall instead of a no-slip wall or a symmetry condition had an influence on the position of the recirculation center and in addition on the flow and turbulence situation. The flow structure with recirculation regions and a double-vortex roll running through the entire tundish is relatively complex, so that compared with flow-optimized applications such as turbine design, where the accuracy is frequently within a range of a few per cents, a somewhat reduced accuracy had to be expected for the CFD benchmark. All in all, the deviations of the simulations, both relative to each other and to the measurement, were indeed within an acceptable range. In the Figure 9. Comparison between measured and calculated RTD curves. following part II, the current CFD benchmark will be extended to the original 16 t tundish The total time load as shown in Figure 11 comprises the with molten stainless steel (X5CrNi18-10). times for preprocessing, computing time and postTo sum up, it can be said that a CFD simulation for processing, and is an average of 50 hours and varies from metallurgical flows is necessary wherever the real process 18 hours to 108 hours. The computing time very much defies direct observation and the expected change in the depends on the computer type, parallelization procedure objective criterion is large. Parameter studies in which an and CFD model, see Table 2. With the work effort for the optimization causes only very small changes in terms of supplementary simulations of some users left unconsidered, flow and heat, involve larger computing errors. What must an acceptable approximation of the actual tundish flow can be borne in mind is that such small changes are hardly be presented after a few working days. A greater amount discernible in the real steel production process due to the of work effort, for example by making a transient highly fluctuating process conditions (e.g. steel computation, an LES simulation and a successive grid composition, inclusion concentration, temperature). Yet adaptation with respect to the flow structure in advance, this might provide a stimulus to reflect about more flowwill in this case result in a simulation with more accurate oriented process modifications, which means e.g. for the details (see, e.g. Users 2, 5). tundish to consider alternative designs beyond installation of simple baffles such as dam/weir systems. Summary As part of a CFD benchmark carried out by the working group Fluid Mechanics and Flow Simulation of the German Steel Institute VDEh, ten members with an industrial and university background investigated the
272

References
[1] Blling, R.: Numerische und physikalische Simulation der stationren und instationren Strmung im Stranggieverteiler, Doctoral Thesis, RWTH Aachen, 2003.

steel research int. 80 (2009) No. 4

Process Metallurgy
[2] Blling, R.; Odenthal, H.-J.; Pfeifer, H.: Transient fluid flow in a continuous casting tundish during ladle change and steady-state casting, Steel Research International, 76 (2005), No. 1, 71-80. [3] Casey, M.; Wintergerste, T.: Best practice guidelines, European Research Community on Flow, Turbulence and Combustion. Special Interest Group on "Quality and Trust in Industrial CFD", Fluid Dynamics Laboratory Sulzer Innotec, Version 1.0, 01.2000. [4] Ferziger, J.H.; Peric, M.: Computational Methods for Fluid Dynamics. 2nd ed., Springer Verlag, 1999. [5] FLUENT 12.0 Users Guide, 2008. [6] Javurek, M.: Strmung von Flssigstahl und Transport von Einschlssen in einer Stahl Stranggussanlage. Doctoral Thesis, Johannes Kepler Universitt Linz, 2006. [7] Langtry, R.B.; Menter, F.R.: Transition modeling for general CFD applications in aeronautics. AIAA-2005522, 43rd AIAA Aerospace Sciences Meeting and Exhibition, Reno, USA, 10-13 Jan 2005. [8] Launder, B.E.: Second-moment closure and its use in modeling turbulent industrial flows. Int. J. Numerical Methods in Fluids, 9 (1989), 963-985. [9] Launder, B.E.: Second-moment closure: present... and future? Int. J. Heat Fluid Flow, 10 (1989) 4, 282-300. [10] Launder, B.E.; Spalding D.B.: Lectures in mathematical models of turbulence. Academic Press, London, 1972. [11] Levenspiel, O.: Chemical reaction engineering, 3rd ed., Wiley/VCH Weinheim, 1999. [12] Odenthal, H.-J.: Physikalische und numerische Strmungssimulation kontinuierlicher Gieverfahren der Hochtemperaturtechnik, Habilitation Thesis, RWTH Aachen, 2004. [13] Oertel, H.; Laurien, E.: Numerische Strmungsmechanik, 1. Aufl., Springer Verlag, 1995. [14] Pope S.: Turbulent Flows, Cambridge University Press, 2003. [15] Sagaut, P.: Large Eddy Simulation for incompressible flows, an introduction, 2nd ed., Springer Verlag, 1998. [16] Sahai, Y.; Emi, T.: Criteria for water modeling of melt flow and inclusion removal in continuous casting tundishes, ISIJ Int.,36 (1996) No. 9, 1166-1173. [17] Schfer, M.: Numerik im Maschinenbau, 1. Aufl., Springer Verlag, 1999. [18] Shih T.-H.; Liou, W.W.; Shabbir, A.; Yang, Z.; Zhu, J.: A new k- eddy-viscosity model for high Reynolds number turbulent flows - model development and validation. Computers Fluids, 24 (1995) No.3, 227-238. [19] Versteeg, H.K.; Malalasekera, W.: An introduction to computational fluid dynamics, the finite volume method. 1st ed., Longman Group Ltd., 1995. List of Symbols B1 m Width of the tundish bottom c Tracer concentration c* Maximum tracer concentration Dhydr DSEN Dsh Fr H k L1 LSEN Lsh & m N Re SEN m m m m m2/s2 m m m kg/s Hydraulic diameter Inner diameter of the Submerged Entry Nozzle Inner diameter of the shroud Froude number Tundish filling level for steady-state casting condition Turbulent kinetic energy Length of the tundish bottom Length of the Submerged Entry Nozzle Length of the shroud Mass flow rate Number of LDA samples per grid point Reynolds number Submerged Entry Nozzle

Figure 10. Relative deviation of the calculated RTD data min, max, 20% and

5%.

100

80 Time in hours

60

40

20

1 2 3 4 6 6 7 8 9 10

1 2 3 4 6 6 7 8 9 10

1 2 3 4 6 6 7 8 9 10

1 2 3 4 6 6 7 8 9 10

Total time load

Preprocessing

Calculation time

Postprocessing

Figure 11. Personal time load of the CFD simulation.

sh t5% tmin,max ttheo Tu u ush u V

s s s m/s m/s m/s m3 m3/s m m2/s3 m/s -

Shroud Residence time for c/c* = 5% Minimum, maximum residence time Theoretical residence time of the fluid in the tundish & (ttheo = V / Vsh , SEN ) Turbulence intensity Maximum back-flow velocity in the tundish Mean flow velocity inside the shroud Theoretical mean-flow velocity through the tundish Fluid flow volume in the tundish Volumetric flow rate Specific dead flow volume Specific plug flow volume Tundish coordinates Dissipation rate Standard deviation Dimensionless residence time ( = t / ttheo)

vdead vplug x,y,z

steel research int. 80 (2009) No. 4

273

Table 2. Chosen CFD boundary conditions for the tundish benchmark

Process Metallurgy

274

steel research int. 80 (2009) No. 4