Large eddy simulations of reacting flows in a dump combustor · 2013-10-15 · Large Eddy Simulations of Reacting Flows in a Dump Combustor' Abstract Large-eddy simulations (LES)

Copyright© 1998, American Institute of Aeronautics and Astronautics, Inc.

Large Eddy Simulations of Reacting Flows in a Dump Combustor'

Abstract

Large-eddy simulations (LES) of turbulentpremixed combustion in a dump combustor that is anaccurate model for an actual gas turbine combustor(General Electric's lean premixed dry low NOx LM-6000) has been carried out to evaluate the potential ofLES for design studies of realistic hardware. A thinflame model for the premixed flame is combined with adynamic model for the subgrid kinetics energy tosimulate the propagation of the turbulent flame in thishighly swirling and high Reynolds number flow field.Comparison of the computed results with experimentaldata indicate very good agreement in spite of relativelycoarse grid resolution employed in the LES. Theseresults provide significant confidence that advancedparallel LES capability for design studies of practicalinterest is feasible in the near future.

Introduction

Time-accurate simulation of turbulent flames inhigh Reynolds number flows is a challenging task sinceboth fluid dynamics and combustion must be modeledand/or resolved accurately. Direct simulations (in whichall scales are resolved and no models are used) are notpractical since the resolution and computational resourcerequirements far exceed the computational capabilityavailable at present and possibly in the near future. Onthe other hand, Reynolds-Averaged methods are not ac-ceptable since they predict the mean motion using a glo-bal model for all turbulent scales and ignore the unsteadydynamics which is critical for accurate predictions. In

Won-Wook Kim*, SureshGeorgia Institute of TechnologyAtlanta, GA, 30332-0150, USA

andHukam C. Mongia**

General Electric Aircraft EnginesCincinnati, OH 45215, US A

the present study, we explore the ability of large-eddysimulations (LES) for simulating high Reynolds numberreacting flows in realistic configurations.

T. Copyright © Kim, Menon and Mongia, 1998, Published byAmerican Institute of Aeronautics and Astronautics by permission.$. Post Doctoral Fellow. Member AIAAtt. Professor, Senior Member AIAA??. Manager, Combustion Technology, Fellow AIAA

In LES, the scales larger than the grid are computedusing a time- and space-accurate scheme, while the unre-solved smaller scales are modeled. Since only the smallscales are modeled it has been suggested that this ap-proach has the potential for studying relatively high Rey-nolds number flows. However, most LES applicationsreported in the past have been limited to simple and/orrelatively low Reynolds number flows. The presentstudy employs LES to study turbulent premixed combus-tion in a combustor which is a part of an operationalhardware. In particular, the device that is investigatedhere is the LM-6000 lean premixed dry low-NOx com-bustor being developed by General Electric Aircraft En-gine Company for gas turbine applications (Hura et al.,1998; Held et al., 1998a,b). Thus, the present demonstra-tion is a focussed "engineering" evaluation of the capa-bility of LES to simulate combustion in realistichardware. A key requirement for "engineering" LES isthat such simulations have to be completed in a reason-able time-frame. This implies that both the accuracy ofthe methodology and the computational effort involvedhas to be evaluated.

The primary focus of this study is on the develop-ment of a new Computational Combustion Dynamics(CCD) tool based on LES for realistic combustor design.To date, this type of design tool has been considered im-practical primarily due to the immaturity of LES method-ology for practical engineering problems and the lack ofadequate computing resources. With the recent progressin subgrid-scale (SGS) modeling techniques and theavailability of massively parallel systems with terraflopcapability, these limitations can be tackled using opti-mized parallel LES codes. This paper explores this capa-


bility and reports on some progress on the developmentof such a simulation tool. The eventual goal is to developa design CCD code that can take advantage of the devel-opments in parallel computing architecture so that it canbe utilized by industry in the near future.

Simulation Methodology

In the LES approach, the scales larger than the gridare computed using a time- and space-accurate scheme,while the unresolved smaller scales are modeled. TheNavier-Stokes equations that govern the conservation ofmass, momentum and energy in a fluid are filtered to ob-tain the LES equations. The filtering operation results interms in these equations that must be modeled. Closureof momentum and energy transport equations can beachieved using subgrid-scale turbulence models. A lo-calized dynamic model based on the subgrid-scale turbu-lent kinetic energy has been developed for the closure ofthe subgrid-scale terms in the momentum and energyequations (Kirn and Menon, 1995; Menon and Kim,1996; Kim and Menon, 1997; Nelson and Menon, 1998).Application of this subgrid model in both incompressibleand compressible turbulent flows shows that this modelis capable of accurately representing the effect of unre-solved terms even when relatively coarse grids are em-ployed. Recently, this model has been compared andevaluated against other contemporary state-of-the-art dy-namic models by Fureby et al. (1997). The details of theLES equations and the subgrid closure employed are giv-en elsewhere and therefore, only briefly summarizedhere.

In a multi-component flow field, the key terms thatrequire closure are the stresses in the momentum equa-

SSS — r'~~^ ** A

tions: T ;- = p[«(-« •-«,-«•] , the energy flux andthe viscous terms in the energy equation:

7, = p tEw.- .EHj-OH.-pw,] ,

y«7 = [ui^ii~ ufiij] anc* me sca'ar correlations:

Xj/£ = TY k — TY k in the equation of state. Here,

hat and bar indicate filtered variables, u-, p , T • • and

E are, respectively, the resolved velocity components,the density, the viscous stresses and the total energy perunit volume. To close these subgrid terms, a modelequation for the subgrid kinetic energy

k = UfUf — u^u- is also solved along with the LESequations (Kim and Menon, 1995; Menon and Kim,1996). The subgrid terms are then obtained using the

subgrid kinetic energy as the characteristic velocityscale and the grid scale as the characteristic length scale.Thus, the eddy viscosity is determined as:

V, = Cvk A where, A is the grid size. Theo o

coefficients in the subgrid kinetic energy equation and inthe eddy viscosity form (e.g., Cv) are computed locally(in space and time) during the simulation using adynamic procedure that has been discussed elsewhere(e.g., Kim and Menon, 1995, Menon and Kim, 1996;Fureby et al., 1997). With k known, the subgrid termscan be closed. Thus, the subgrid turbulent stresses are

modeled as: I'f5 = -2pV,(5j;/-5^8^/3), thesubgrid enthalpy flux is approximated as,

H]85 = -pVf(tf)/Pf, and of?' = urff*.

Here, Pr ~ 1 is the turbulent Prandtl number, S; • isthe resolved rate-of-strain tensor,

Sjj = [("/); +(w/),-]/2, and H is the resolvedtotal enthalpy. Also, all third-order subgrid correlationsare neglected in this closure for simplicity.

In general, simulation of reacting flows (both pre-mixed and non-premixed) requires the solution of thespecies conservation equations. The LES species equa-tions can be written as:

where, Y k and Dk are, respectively, the k th spe-cies mass fraction and the diffusion coefficient. Here,(<j)) f and ((J))^ indicate differentiation with respect to

time and space, respectively. Closure for the species-

velocity correlation = ^(u^Y k- UjY k ] , the

production term (Oj. and the scalar correlations VP k ismore problematic, since, to estimate these terms, small-scale turbulent mixing and species molecular diffusionmust be modeled correctly. Conventional subgrid clo-sure cannot handle these features and, therefore, the newsubgrid combustion models (Menon and Calhoon, 1996;Smith and Menon, 1998) are being developed to addressthis limitation.

For conventional LES simulations of turbulent pre-mixed combustion, a model which circumvents the


above noted closure problem can be used for simula-tions of thin flames (i.e., for flames thinner than thesmallest eddy). For these cases, the flame is modeled asa propagating surface and a variable G is defined that isgoverned by the equation (Kerstein et al., 1998):

(2)

where Up is the local propagation speed. Thisequation describes the convection of a level surface,described by G = constant, by the velocity M whileundergoing propagation normal to itself at a speed Up-.

G = \ denotes the reactant and G=0 the burnt regionwith the thin flame identified by a level surface in the[0,1] range. The effect of heat release is included in thedefinition of the specific enthalpy. For turbulent flames,the propagation speed Up is the turbulent flame speed

Uj and has to be specified. Earlier (Menon and Jou,1991) employed the flame speed model of Yakhot

(1988) uF/SL = exp[u /up]. Here, SL is the

laminar flame speed and «' = */2k/3 is the localturbulence intensity which is known.

In the present study, the G-equation model is againemployed for the simulation. Interestingly enough,although this model was originally proposed for theflamelet regime, the domain of its applicability appearsto be much larger as demonstrated recently by Peters(1998). In the present LES approach, the G evolutionequation is filtered in a similar way as is done forNavier-Stokes equations. Two unresolved terms areobtained as a result of filtering: the transport and thesource terms, respectively. The unresolved transportterm is modeled using a gradient diffusion assumption(Smith and Menon, 1998). For closure of the sourceterm, various models have been developed andproposed, among them is Yakhot's RNG model(Yakhot, 1988) where the turbulent flame speed isanalytically expressed as a function of turbulenceintensity. Yakhot's model has been tested in the presentstudy and it was found that the predicted turbulentflame speed is too small when the turbulence intensityof the LM-6000 is considered (see more discussionbelow). This result may be due to the very high swirlintensity in the present problem and is an issue thatwarrants further investigation. Therefore, to obtain areasonable turbulent flame speed for the regime ofinterest, Pocheau's model (Pocheau, 1994) is employed

in the following form:

„ 7 1/2ur = (1+pV ) (3)

where (5 is an adjustable parameter. In the presentstudy, (3 is calibrated using existing data of weaklyswirling premixed turbulent flames (Cheng, 1995).However, the form of the turbulent flame speed model isan issue for further research and some implications of thechoice is also discussed in this paper.

Parallel Implementation

The technique of data concurrency (i.e., the primarydata space is partitioned and distributed among the pro-cessors) rather than functional concurrency (i.e., theoverall application is decomposed into several distinctparallel computational tasks) was chosen after carefulreview of the type and degree of parallelism inherent inthe numerical algorithm. The data space is partitionedand distributed to the processors so that 1) the distribu-tion of cells to the nodes leads to a nearly balanced loadof communication and computation among all nodes,and 2) the inherent spatial data locality of the underlyingcell structure is maintained so as to minimize interpro-cessor communication. The cell partitioning schemedecomposes the 3D computational domain into logicallycongruent, nearly equal-sized cubes. Maximum concur-rency is extracted to minimize the execution time on agiven number of processors. The overheads associatedwith parallel implementation, such as, (1) load imbal-ance, (2) inter-processor communication, (3) datadependency delays, (4) arithmetic, and (5) memory,were analyzed. While the first four types of overheadslead to performance degradation, the memory overheadlimits the size of the problem that can be executed on afixed number of processors. In practice, simultaneouslyminimizing all these overheads is very difficult.

In the present implementation, the partitioningscheme results in each processor performing computa-tions only on the cells that are held by it. For finite-dif-ferences, each domain contains extra layers of ghostcells along the processor partitions to allow theexchange of boundary cell data. This exchange is car-ried out using a few, relatively long messages. As aresult, the high cost of latency associated with messagepassing is minimized, resulting in a reduced communi-cation overhead even though this data exchange resultsin an increased memory overhead.

The current implementation on parallel systems


employs double precision (64-bit) arithmetic and isbased entirely on Fortran. Performance comparison withand without I/O has been carried out. However, I/Ooverhead is unavoidable since the data generated on thespatio-temporal evolution of the flow field is needed foranalysis. The type, form and frequency of data varieswith the problem and, thus, cannot be standardized. Ingeneral, the 3D flow field is needed for flow visualiza-tion and for restart files. The present approach combinesboth these requirements by making all processors towrite the required data into one file. The location of thisfile depends on system architecture: for example, the filecan reside on the local file system or it has to reside onone processor. In general, if each processor writes itsown local data the I/O overhead is the minimum; how-ever, this capability is not available in many modern dis-tributed parallel systems such as the Cray T3E.

To optimize I/O time on all systems, the flow vari-ables from all processors are written into a temporarybuffer array which always resides on one processor andthen this array is written to a file. This approach resultsin one processor writing a large amount of data insteadof all processors writing small amounts of data (whichwas found to cause I/O bottleneck). Recent studies sug-gest that if a processor is allocated the task of only I/Othen asynchronous I/O can be conducted withoutadversely affecting the computational effort. This I/Oimplementation works reasonably well on all systemsand is considered an optimal compromise to allow flexi-bility in porting the code (and data) to different systems.In addition, this approach allows the simulation to berestarted on arbitrary number of processors. This capa-bility is very useful when the system is heavily loaded.

The timing data for the present 3D LES code is sum-marized in Table 1 for the T3E and the Origin 2000. The3D code achieves 32.33 Mflops per processor on theCray T3E (when 90 processors were used - see furtherdiscussion in the next section). This compares very wellwith some parallel benchmarks reported for this ma-chine. The present LES code also achieves very goodscale-up efficiency on the T3E. However, as is wellknown the scaleup depends upon the problem size andtherefore, the optimum size for a given system configu-ration needs to be determined. Table 1 also shows that al-though the computational speed on the O2000 is better(by almost a factor of 2), the scale-up (for a given prob-lem size) is not as good as on the T3E. This is related tothe problem size; however, in general, it appears that theO2000 architecture is computationally faster than theT3E but does not maintain good scaleup when the num-ber of processors are increased.

Numerical Model and Test Problem

General Electric's lean premixed dry low NOx emis-sions LM6000 gas turbine combustor has been simulat-ed. As shown in Fig. 1, this problem has a complicatedgeometry that is quite challenging for computation. Ahighly swirling jet (the maximum value of tangential ve-locity component is slightly greater than the peak valueof axial velocity component) is injected from a circularinlet under high pressure (P=6.18xl05 N/m2 ~ 6 atmo-spheres) and temperature (T=644 K) conditions. Thecombustor comprises of a rectangular box with twoblocks located at top and bottom surfaces from whichcooling air is blown downstream. The typical maximuminflow axial jet velocity is around 1 lOm/s and the Rey-nolds number based on the maximum velocity is350,000. The premixed fuel-air mixture has a pre-speci-fied laminar flame speed of 0.28 m/ (corresponding to alean methane-air mixture). The estimated flame temper-ature is around 1811 K (based on gaseous emissions).The Reynolds number is high enough such that theflamelet assumption is reasonable for this test case. Esti-mates suggest that the ratio u'/S^ can be locally ashigh as 100 in this combustor.

To carry out a true "blind" comparison (and to dem-onstrate a honest test of the LES model), the present LESwas carried out without any knowledge of the experi-mental data. The only information provided by GEAEwas the configurational details noted in fig. 1, non-di-mensional inflow velocity profiles (axial, radial, and az-imuthal), maximum axial, radial, and azimuthalvelocities, and, the inflow turbulence intensity. LES wascarried out using around 300,000 (80x48x80) and500,000 (96x64x80) grid points. These resolutions areconsidered very coarse and not representative of typicalLES reported in the literature. However, this range ofresolution size was chosen to obtain "engineering" re-sults in a reasonable time frame. The accuracy of such aresolution choice obviously requires confirmation usingboth comparison with data and higher grid resolutionLES. In the present paper, we demonstrate the currentLES capability by direct comparison with data. Higherresolution LES will be carried out in the near future andwill be reported soon (Kim et al., 1999).

The LES model is based on a finite-volume schemethat is fourth-order accurate in space and second-orderaccurate in time. The full compressible LES equationsare solved along with the filtered G-equation and theequation for the subgrid kinetic energy. Both the con-stant coefficient and the dynamic subgrid kinetic energymodels were used and the results are compared. A limi-


tation of employing the G-equation is that it has only twopossible states: G=l denoting premixed cold fuel andG=0 denoting the hot products. In the experiments, coldair is blown through the side blocks (see Fig. 1) primarilyfor cooling the combustor. However, all the combustionprocess is completed upstream of the air ports. There-fore, in the present G-equation approach, we inject thehot product through the injection port using the sameamount of mass flow rate. The consequence of this injec-tion on the combustion process occurring upstream can-not be addressed using the G-equation and requires LESwith multi-species. The current LES code does have thiscapability and we plan to carry out such simulations inthe near future.

The coarseness of the grid employed here impliesthat acceptable resolution in all regions of interest cannotbe achieved. Since we were interested in the flame re-gion, the grid was clustered in the near field of the exitplane at the expense of the resolution near the wall andfar downstream. As a result, the dynamic model has nu-merical problem near the wall. To stabilize the dynamicsimulations, the constant coefficient version is employednear the wall region. The consequence of coarse gridnear the wall is that some of the derived properties arenot likely to be accurate near this region (see the nextsection). Higher resolution simulations are planned to re-visit this issue.

The inflow conditions were specified based on thenormalized profiles provided by GEAE. No data on theincoming turbulence was available or provided (otherthan the incoming intensity of 7%). In the present study,an inflow turbulent field was generated by specifying theturbulence intensity. The outflow was quite problematicinitially, since, due to the high swirl in the flowfield, re-verse flow can occur near the outflow during the initialstart-up transients. To avoid this problem, a buffer regionof 7.6 cm was added in the axial direction and the areawas linearly contracted 25% in the initial 5.08 cm of thebuffer region. Characteristic outflow boundary condi-tions were then imposed at the exit. With contraction theflow near the outflow is positive and causes no problemswith the characteristic outflow boundary conditions. Theeffect of varying the contraction and the length of thebuffer region was also studied. It was determined that thechosen buffer region and the contraction has negligibleeffect on the flame dynamics. This is because due to thehigh swirl the flame is located very close to the fuel exitplane and all the combustion process is completed within1-2 jet diameter distance from the fuel exit plane.

Results and Discussions

The results of the current LES are discussed in thissection. Although both non-reacting and reacting LESwere conducted, only the reacting flow results will bediscussed in detail.

To obtain statistically stationary results for compar-ison with the LDV data provided by GEAE the simula-tions were carried out for over 20 flow-through times andthe data were time-averaged using the last 15 flow-through times. Although analysis suggests that the solu-tion has reached stationary state (comparison of resultsafter 10, 15 and 20 flow-through times shows that manybut not all flow properties have reached stationary state),we plan to continue these simulations further in time toensure this. The key limiting condition for such long sim-ulations is the availability of CPU time. The current sim-ulations were carried out on the Cray T3E and typically90 processors were employed primarily to reduce theturn-around time (for the present problem size, LEScould have been conducted using around 24 processors).As shown in Table I the present LES code does scale upquite well on the Cray T3E but the scaleup efficiency de-grades if the problem size is too small. Furthermore, theprocessing speed of around 33 MFLOPS/processor wasestimated using this problem size on the 90-processorconfiguration. Therefore, it is likely that a larger grid willscale up even better and execute at a much faster speed.Since we plan to simulate this combustion flow field us-ing a much higher resolution (e.g, 1-2 million gridpoints) we will revisit this issue. The 90-processor optionwas chosen at present as an optimum compromise bothfrom scaleup and availability considerations.

A typical simulation using half million grid pointson the CRAY T3E required around 2.2 GB of Memoryand around 1,000 single processor hours per flow-through time. Using 90-processors, it was possible to getone flow-through time overnight. As noted above the to-tal cost of such simulations depend on the requirementsfor stationary state. For example, in the present study, ad-equate results were obtained for comparison with dataafter 10 flow-though times although we did carry thesesimulations out twice as long.

We now discuss some of the characteristic results ofthese calculations. Figure 2 compares the centerlinemean axial velocity decay obtained from non-reactingLES using the two grid resolutions employed so far. Al-though the general trend and magnitudes are similar,there are some differences. For example, in the nearfield, the coarser grid result shows numerical fluctua-tions in the mean velocity. Comparison of the mean and


fluctuating velocity fields at various locations showedsimilar behavior. The higher resolution grid clusteredmost of the points in the near field where mixing andcombustion is expected to occur. Thus, the results weresmoother in the near field. However, in the far field andnear the walls even the finer grid show numerical effects.As mentioned before, higher resolution LES is still re-quired to ensure that such grid effects are not contami-nating the near field predictions.

Figure 3 compares the non-reacting and reactioncenterline mean axial velocity decay using the highergrid resolution. The effect of combustion and heat re-lease can be clearly seen in the reacting flow data as a ve-locity hump. This is the well known flow accelerationbehind a premixed flame that has been seen in many pastexperimental and numerical studies (e.g., Smith and Me-non, 1998). From this figure is can be inferred that thesteady state flame is located within 1 jet diameter dis-tance from the fuel exit plane. This location is far up-stream of the side injection ports (located at 51mm).

Time averaged vorticity magnitude |ft)| contourplots are shown in Figs.4 (a-d) while two arbitrarily cho-sen instantaneous vorticity fields are shown in Figs. 5 (a-d) and Figs. 6 (a-d). The plots include one plane perpen-dicular to z-axis through the combustor centerline andthree planes perpendicular to x-axis at three differentdown stream locations of the fuel exit (x=6 mm, 24 mm,and 78 mm, respectively). The second location at 24 mmis approximately near the steady-state flame location.The swirling incoming premixed jet expands rapidly andresults in a forward stagnation point. The unsteadiness ofthe flow can be clearly seen in the plots (Figs. 5 and 6).Detailed animated visualization of the flow field showsthat the high swirl results in a very complex vortex shed-ding pattern with significant azimuthal structure. Theshear layer is quickly broken up into highly stretched(azimuthally) vortex rings. As the vortices impact on thewall secondary vortices of the opposite sign are also gen-erated. This phenomenon is well known. It can also beseen from the instantaneous cross-plane figures that al-though the vortex rings break down quickly, near theflame (at x=24 mm) there is still significant azimuthalcoherence in the vortex structure.

Figure 7 shows the time averaged (a) and instanta-neous (b and c) flame kernels. The flame zone has beenmagnified for comparison but is actually very small rel-ative to the combustor dimensions. The unsteadiness ofthe shear flow due to swirl also directly affects the flamestructure as can be seen in these figures. Comparison ofthe instantaneous flame and vortical structures show thatthe flame is located very close to the region of high co-herent vorticity. Thus, it is likely that, locally, the flame

may be encountering regions of high strain. This coupledwith the increased flame curvature seen in the LES re-sults would suggest that stretch effects may be importantfor such flames. It is well known that high stretch cancause premixed flame extinction and is an area of consid-erable interest. Although the G-equation model can beused to mimic extinction (for example, by setting theflame speed to zero if the local strain exceeds a pre-spec-ified value) it cannot handle re-ignition (which requiresreduction of strain as well as appropriate temperature).These issues can be addressed using relatively simple(single step) chemistry and will be the focus of future in-vestigation.

Figure 8a compares the present LES results withavailable data on the mean axial velocity decay along thecenterline. Both the constant coefficient and the dynamicsubgrid model predictions are shown in these figures.Also shown is a LES prediction using the constant coef-ficient model but with a reduced inflow mean axial ve-locity (of 103 m/s) similar to the value used by GEAE intheir anchored CCD studies. Comparison with datashows reasonable agreement. There is not much differ-ence between the two subgrid approaches before theflame. However, after the flame location (which isaround 24 mm) there are some observable differences. Itwould appear that the dynamic model agrees with thedata in the far field better than the constant coefficientmodel.

As noted above, earlier LES was carried out using areduced inflow velocity as provided by GEAE (shown inFig. 8a). This axial inflow velocity (at the center) of 103m/sec was anchored by GEAE to carry out their opti-mized CCD calculations. For comparison Fig. 8b showsthe results predicted by GEAE using their AnchoredCCD approach (Hura et al., 1998; Held et al., 1998a,b).These calculations are done using 60x40x28. As can beseen they obtained good agreement by anchoring the in-let velocity profiles. The LES results shown in Fig. 8aalso shows very good agreement without any adjust-ments to the model. (These results can be further im-proved by employing the inlet velocity and turbulenceprofiles which are anchored for LES.) Analysis of thedata shows that the present LES predicts the location ofthe forward stagnation point (59mm from dynamic mod-el LES and 50mm from constant coefficient model LES)in very good agreement with the experimental data(56mm) and the earlier Anchored CCD calculations byGEAE.

Figures 8c and 8d show respectively, the mean axialvelocity profiles in the y-direction at x=5.83mm andx=23.94mm. Agreement is again reasonable at x=5.83mm. At the second location LES predictions are not in


good agreement, even though the magnitude is reason-able. The experiments suggest a peak at y=0 but the LESdata shows a peak further away from the center. The ear-lier calculations using Anchored CCD modeling byGEAE also showed results somewhat similar to thepresent calculations.

Figures 9a and 9b show the mean radial velocity pro-files in the y-direction at x=5.83mm and x=23.94mm, re-spectively. Comparison with LDV data showsreasonable agreement. Both the magnitude and the trendof the velocity profile is accurately predicted by the LES.The lack of resolution in the far field (near the walls)shows up as numerical noise in the far field data. It canbe seen that the dynamic model has more noise than theconstant coefficient model. As noted above, the verycoarse resolution in the far field causes large variation inthe dynamic model coefficient. However, it appears thatin the near field where the grid is reasonably fine, bothsubgrid approaches behave very similar.

Figures lOa and lOb shows the mean tangential ve-locity profiles in the z-direction at x=5.83mm andx=23.94mm, respectively. Again, the agreement with ex-perimental data is considered reasonable considering theresolution employed. There are no appreciable differenc-es between the two subgrid approaches.

These comparisons suggest that at least for suchcoarse grid LES, the constant coefficient model is prob-ably acceptable. Note that, dynamic model is much moreexpensive and therefore, using the constant coefficientmodel will also reduce the overall computational cost. Itis not yet clear if the dynamic model will give improvedpredictions when higher grid resolution is employed.This issue still remains to be addressed.

Many other flow properties were computed andstored during the LES. Such flows properties include allsecond- and third-order correlations. However, theseproperties have not yet been fully analyzed. Here, wecompare some of the predicted results for turbulence in-tensity. There are no experimental data available forthese properties so comparison is not possible. However,given the reasonable agreement of the mean flow data,the predicted turbulence intensity variations are alsolikely to be similar to the experimental case.

Figures 1 la, lib and 1 Ic show respectively, the ax-ial, radial and tangential velocity fluctuation intensityprofiles at the same axial location of x=23.94. Both thereacting and non-reacting data are shown in these fig-ures. The axial velocity fluctuation intensity for the re-acting case shows significant magnitude in the shearlayer region where the flame is located. The radial and

tangential fluctuation intensities are also quite large andcomparable to the axial intensities. Interestingly, the tan-gential intensity (Fig. lie) shows very high values nearthe outer region for the reacting case. However, the accu-racy of the present LES results in the outer regions re-mains to be determined since the current resolution thereis considered too coarse.

Conclusions

This paper has reported on the application of LESto study premixed combustion in a device that is part ofa real combustor. Thus, the geometry and flowconditions are chosen to match actual operationalcondition for the General Electric's lean premixed drylow NOx LM-6000 combustor. The goal of thisexercise was to evaluate the feasibility of carrying out"engineering" LES of actual hardware. To achieve thisgoal, some simplifications such as very coarse gridresolution was needed to reduce the computationaltime. The computational time was also considerablyreduced by using the thin flame model for premixedcombustion. However, analysis of the flow conditionssuggest that the thin flame model is actually a goodapproximation for the real case provided issues such asextinction and re-ignition are not considered. The LESwere carried out using both constant coefficient anddynamic models for the subgrid kinetics energy.Comparison of the computed results with experimentaldata indicate very good agreement in spite of relativelycoarse grid resolution employed in the LES. Theseresults provide significant confidence that advancedparallel LES capability for design studies of practicalinterest is feasible in the near future.

Acknowledgments

This work was supported in part by the Air Force Of-fice of Scientific Research Focused Research Initiativecontract F49620-95-C-0080, monitored by GeneralElectric Aircraft Engine Company, Cincinnati, Ohio andby the Army Research Office Multidisciplinary Univer-sity Research Initiative grant DAAH04-96-1-0008.Computations were carried out under the DoD HPCGrand Challenge Project at NAVO, Stennis Space Cen-ter and ARC, Huntsville.

References

Cheng, R. K. (1995) "Velocity and scalar character-istics of premixed turbulent flames stabilized by weakswirl," Combustion and Flame 101, pp. 1-14.


Fureby, C., Tabor, G., Weller, H. G., and Gosman,A. D. (1997) "A comparative study of subgrid scalemodels in homogeneous isotropic turbulence," Physicsof Fluids 9, pp. 1416-1429.

Held, T. J., Mongia, H. C., and Tonouchi, J. H.(1998) "Application of amicro-Anacro-mixing two reac-tor model to a single cup low-emissions combustor,"ASME Paper No. 98-GT-218.

Held, T. J., and Mongia, H. C. (1998) "Applicationof a partially premixed laminar flamelet model to a low-emission gas turbine combustor," ASME Paper No. 98-GT-217.

Hura, H. S., Joshi, N. D., Mongia, H. C., and Tonou-chi, J. (1998) "Dry low emission premixer CCD model-ing and validation," ASME Paper No. 98-GT-444.

Kerstein, A. R., Ashurst Wm., T. and Williams F. A.(1988) "The Field Equation for Interface Propagation inan Unsteady Homogeneous Flow Field," Physical Re-view A 37, pp. 2728-2731.

Kim, W.-W. and Menon, S. (1995) "A new dynamicone-equation subgrid-scale model for large-eddy simula-tions," AIAA 95-0356.

Kim, W.-W. and Menon, S. (1997) "Application ofa localized dynamic subgrid-scale model to turbulentwall-bounded flows," AIAA 97-0210.

Kim, W.-W. and Menon, S. (1999) "LES of pre-mixed turbulent reacting flows in a highly-swirling com-bustor," submitted to present at 37th AIAA Aerospace

Sciences Meeting, Reno, NV, January 11-14, 1999.

Menon, S. and Calhoon, W. (1996) "Subgrid mixingand molecular transport modeling for large-eddy simula-tions of turbulent reacting flows," Symposium (Interna-tional) on Combustion, 26, pp. 59-66.

Menon, S. and Jou, W.-H. (1991) "Large-eddy sim-ulations of combustion instability in an axisymmetricramjet combustor," Combustion Science and Technology75, pp. 53-72

Menon, S. and Kim, W.-W. (1996) "High Reynoldsnumber flow simulations using the localized dynamicsubgrid-scale model," AIAA 96-0425.

Nelson, C. C. and Menon, S. (1998) "Unsteady sim-ulations of compressible spatial mixing layers," AIAA98-0786

Peters, N. (1998) "The turbulent burning velocity forlarge scale and small scale turbulence," Submitted toJournal of Fluid Mechanics.

Pocheau, A. (1994) "Scale Invariance in TurbulentFront Propagation," Physical Review E. 49, pp. 1109-1122.

Smith, T. M. and Menon, S. (1998) "Subgrid Com-bustion Modeling for Premixed Turbulent ReactingFlows," AIAA 98-0242.

Yakhot, V. (1988) "Propagation velocity of pre-mixed turbulent flames," Combustion Science and Tech-nology 60, pp. 191-214.

Overall Grid | #PEs U T . . II T 11 communication || L compute | * totalCRAY T3E (Sees)

96x64x80 8163264128

1.20985875 9.9847781250.687654344 5.0301440630.478368688 2.5932207810.337139852 1.3553969530.225933906 0.749989961

11.194636885.7177984063.0715894691.6925368050.975923867

SGI ORIGIN 2000 (Normalized by T3E data)96x64x80 8

1632

0.571 0.4880.553 0.4600.841 0.528

0.4970.4710.577

Table 1: Timings for the 3D LES code with one species (G eqn.) on different parallel systems. Note that I/Otiming is not included.


Sida Vlaw

0.051m

End Viau

Swirler Dia - 0.0335m

0.1524m 0.07m

Figure 1: Schematic of the GE low NOx Premixed Combustor.

U

LES (80x48x80)LES (96x64x80)

Figure 2: Comparison of the axial velocity variations along the combustor centerline obtained from LES usingdifferent resolutions.

U

Figure 3: Comparison of the axial velocity variations along the combustor centerline obtained from LES ofnon-reacting and reacting flows.


(b) (c)

Figure 4: Averaged vorticity magnitude contours, (a) Side view, (b) End view (x=6 mm), (c) End view (x=24mm) and (d) End view (x=78 mm).

Figure 5: Instantaneous vorticity magnitude contours, (a) Side view, (b) End view (x=6 mm), (c) End view(x=24 mm) and (d) End view (x=78 mm).

Figure 6: Instantaneous vorticity magnitude contours, (a) Side view, (b) End view (x=6 mm), (c) End view(x=24 mm) and (d) End view (x=78 mm).

(c)

Figure 7: Flame kernel (a) Time Averaged; (b) and (c) Two instantaneous snap shots.

10


(a)

u

120

100

80

60

40

20

0

-20

................. LES (Constant)———— LES (Dynamic)

• Experiment----- LES (Inflow II)

40 80;t (mm)

120 160

(b)

1.30-

\23-

1 103-

! a*-i,1 aaa-

:ts

> 0.00-

-033-

Axial Velocity V3Xy=0rnm, 2=0 ram

i

tYj'V1

2

•***

s.

1 5

^^

!»•»-J

to*CK— i**3 11

**»„

•^.

o 1:

--

•̂̂

D 1

™,

10 11

x.rnm

———— ACC-Anetvredm LDVdaa

— - - - Bef Is 2 • 0 rrmdata——— -Rofie3.anodsbi.VH3

a

Figure 8: Comparison of the predicted axial velocity along the combustor centerline and the experimentaldata, (a) LES and (b) Anchored CCD by GEAE.

(c)

U

120

100

80

60

40

20

0

-20

•••••• LES (Constan;— LES(Dynamij

• Experimentx\- - - - - LES (Inflow I

-40-50 -40 -30 -20 -10 0 10 20 30 40 50

y (mm)-50 -40 -30 -20 -10 0 10 20 30 40 50

y (mm)

Figure 8: Comparison of the predicted axial velocity and the experimental data along the y-axis (c) 5.83mmand (d) 23.94mm downstream of the inlet

(a)30

20

10

0

-10

-20

-30

LES (Constant)LES (Dynamic)

• Experiment .

(b)

-50 -40 -30 -20 -10 0 10 20 30 40 50y (mm)


-60-50 -40 -30 -20 -10 0 10 20 30 40 50

v (mm)

Figure 9: Comparison of the predicted radial velocity and the experimental data along the y-axis (a) 5.83mmand (b) 23.94mm downstream of the inlet

11


(a)120

80

40

o

-40

-80

-120


-40

120(b)

-20 0z(mm)

20 40

80

40

0

-40

-80

-120

- LES (Constant)— LES (Dynamic)• Experiment

-20 0z (mm)

20 40

Figure 10: Comparison of the predicted tangential velocity and the experimental data along the z-axis (a)5.83mm and (b) 23.94mm downstream of the inlet.

LES (Constant)lLES (Dynamic


-50 -40 -30 -20 -10 0 10 20 30 40 50v (mm)

-50 -40 -30 -20 -10 0 10 20 30 40 50V (mm)

----- LES (Constant-—— LES (Dynamic)----- LES (Cold)

. - V

-40 -20 0C(mm)

20 40

Figure 11: Turbulence intensities at 23.94mm downstream of the inlet, (a) axial, (b) radial, and (c) tangentialcomponents.

12

Documents

Large eddy simulations of reacting flows in a dump combustor · 2013-10-15 · Large Eddy Simulations of Reacting Flows in a Dump Combustor' Abstract Large-eddy simulations (LES)