You are on page 1of 12

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol.

114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

SL-08-005

Verification and Validation of CFD for the Personal Micro-Environment


Chris N. Sideroff, PhD Thong Q. Dang, PhD

ABSTRACT A detailed Computational Fluid Dynamics (CFD) study of the flow around a Computer Simulated Person (CSP) in an otherwise empty displacement ventilated room is presented. Both Reynolds Averaged Navier-Stokes (RANS) and Large Eddy Simulation (LES) methods are included in this study. The study identifies the requirements of several computational aspects that are needed for accurate CFD simulations of the personal micro-environment, which include issues related to grid size and iterative convergence monitoring, turbulence modeling, and radiation modeling. Using the benchmark test for evaluating CFD in the indoor environment of Nielsen et al. (2003), very good agreement between CFD results and test data was obtained using the standard k-" with wall treatment and a reduced-order radiation model. INTRODUCTION For ventilation systems other than an ideal mixing system, e.g. displacement and personal ventilation systems, the well-mixed assumption can not be applied because spatial gradients of the flow, temperature and pollutant fields can be large in the vicinity of the Personal Micro Environment (PME). Here, the PME is the region around a person that affects the air he/she breathes. In a displacement ventilation system, human exposure to pollutants is influenced by the convective and diffusive transport mechanisms found in the thermal plume around the person. The flow in the thermal plume is dominated by the buoyancy forces arising from the higher temperature of the human body, and this flow field is rather complex even in a situation where the person is standing in an empty room (Clark and Edholm 1985).

The PME and the surrounding environment have a sophisticated relationship that is rarely amenable to simple analytical models. Analytical methods are not without merit as they are powerful tools for understanding the fundamental features (Awbi 1991) of a real scenario but the interactions of these features are not trivial. This concept is elucidated by Melikov and Kaczmarcyzk (2007) who state, The measurement at a point in a room, without a person present, would not define accurately the quality of the air that the person would inhale when present at this location. As a result, sophisticated tools and methods are needed to extract a deeper understanding of the PME. Beyond analytical, the two remaining approaches are experimental and computational. Experiment has traditionally been the most common and reliable method to study the indoor environment. The experimental approach could be characterized by two distinct strategies: field study and detailed measurement. Field study involves placing sensors in a relatively non-intrusive manner within a real environment, statistically analyze the data obtained and making conclusions based on the analysis (Ferro et al. 2003). While this strategy provides the most accurate representation of the actual scenario, the lack of control of input variables makes quantitative inference a dubious task. The alternative experimental strategy is detailed measurements. Again, sensors are used to obtained information about the pertinent quantities, but here the experiment is carried out in a more controlled manner with higher resolution equipment. For the indoor environment and the PME, the types of detailed experimental equipment vary widely depending on the desired quantity. For flow velocity, common techniques are hot-wire, hot-sphere, particle image velocimetry (PIV) and laser-doppler anemometry (LDA), in

Chris N. Sideroff is a technical sales engineer at Pointwise, Inc., Fort Worth, TX. Thong Q. Dang is a professor of mechanical and aerospace engineering, Syracuse University, Syracuse, NY.

!2008 ASHRAE

45

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

order of increasing accuracy and sophistication. For temperature, common techniques are thermal couples, thermistors and infra-red imaging. The drawback of such experimental systems is that they can be expensive, complex to operate and have other limits. Recent examples of detailed experiments with thermal manikins in the indoor environment are Bjrn and Nielsen (2002) to study the interaction of multiple people, Melikov and Kaczmarcyzk (2007) to examine personal ventilation devices on the PME, Cermak et al. (2006) to identify transmission of infectious agents between occupants and Marr (2007) to investigate the influence of body motion. With the advent of high performance computers at commodity prices, computational tools are becoming ever more popular. More specifically, Computational Fluid Dynamics or CFD has been used with success in a variety of areas in the indoor environment community; office buildings (Cheong et al. 2003), homes (Huang et al. 2004), hospitals (Brohus et al. 2006) and aircraft (Zhang and Chen 2001) are examples of a few of these areas. To be able to use CFD as a design tool, one first must have confidence that it can predict the desired quantities with a certain degree of accuracy. To do this, two important questions need to be answered: First, is CFD capable of predicting the flow in question and second, what is needed to do so? Flows around humans in an indoor environment can be particularly difficult to predict because of the complex interaction between many different factors. To properly answer these questions, standard or canonical benchmark cases are needed so the individual issues may be identified along with providing a consistent approach for others researchers to use. The collaborative efforts of Nielsen et al. (2003) have culminated in a benchmark test for evaluating CFD in the indoor environment. They have proposed two canonical building environment scenarios: mixing and displacement ventilated rooms with a centrally situated manikin. Results and discussion of the displacement ventilation case are presented in this work. Verification and validation of CFD for the personal micro-environment (PME) is necessary for it to become a reliable tool. The objective using the benchmark case is to identify the key requirements needed to achieve an accurate prediction of the detailed flow field around the PME of a lifelike Computer Simulated Person (CSP) or manikin. Previous studies of Topp (2002) and Topp et al. (2002) have examined the differences of simplified and detailed CSP. They concluded that while variations between the two were inconsiderable at some distance from the CSP, very near the CSP, and more importantly in the personal micro-environment, the distinctions were much more apparent. They provided examples highlighting how CSP details can affect calculations of contaminant transport, heat-transfer coefficients and view factors need for radiative heat-transfer. Specifically, Topp (2002) provided conclusive evidence that the contaminant distribution in the personal micro-environment of the manikin is a strong function of geometry detail.
46

The CSP geometry used for this study is a digitization of a female manikin in the standing position. The data file was obtained from Katos research group at the University of Tokyo. The surface area of the manikin was approximately 1.48 m2, a value within the known measured range for females. Diminutive body details such as ears, fingers and toes were removed to disburden grid generation while maintaining a sufficient level of surface description. Furthermore, aspects such as hair and clothes were also not included to alleviate further uncertainties (e.g., heat transfer through clothing). Therefore the manikin represents a female of average body surface area situated in the standing posture. The original summary of the scenario setup, boundary conditions and suggested reporting method of the two benchmark cases presented in this paper are outlined in Nielsen et al. (2003). CFD results were extracted and compared at the same locations where experimental data was measured. The experimental data, in addition to further information, for each case is openly available at http://www.cfd-benchmarks.com. COMPUTATIONAL MODEL All CFD simulations were performed using the commercial CFD software FLUENT (version 6.3), where the incompressible Navier-Stokes equations (Equation 1) are solved with the SIMPLEC pressure-velocity coupling method along with the energy equation (Equation 2).
1 Dv -------- = ----- $P + $ * &%$ v ' g & 1 ()T 'k $ * &v'v ' ' #o Dt DT = $ * & +$T ' $ * &T 'v' ' ------Dt

(1)

(2)

The overbar represents averaged quantities whereas primed variables are fluctuating quantities. The second last term in Equation 1 is the Boussinesq approximation for thermal buoyancy effects and the last terms in Equations 1 and 2 are the Reynolds stress tensor and heat-flux vector, respectively. To close the problem the Reynolds stress tensor and heat-flux vector are related to the mean velocities and temperatures through the gradient-diffusion hypothesis
2 &v'v' ' = %T S + -- kI 3 &T 'v' ' = + T $T 1
T

(3) (4)

where S = -- . $v + $ v / is the mean strain-rate tensor and k 2, is the turbulence kinetic energy. The turbulent viscosity, %T , is determined from characteristic velocity and length or times scales. %T = C 0 VL ; %T = C 0 V T
2

(5)

ASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

computed by the various turbulence models.

The turbulent thermal diffusivity, + T = %T 1 Pr T and Pr T = 0.85 . The specification of the characteristic scales are

The numerical convective scheme used for all transport equations was a 2nd order accurate upwind scheme, while diffusion terms are discretized using the 2nd order accurate central difference scheme. Pressure interpolation was achieved with a 2nd order accurate scheme similar to the 2nd order upwind convective scheme. Computational grids were generated using the commercial grid-generation software Gridgen. A Beowulf cluster with 64 1.6 GHz processors (x8664 architecture) and 64 Gbytes of total system memory made this type of analysis possible. Thirty-two processors were typically used, which took anywhere from 48 to 72 hours depending on the grid resolution and turbulence models, the latter include the k-" model with enhanced wall treatment and the % 2 -f model for the Reynolds-Averaged Navier Stokes (RANS) equations, and the Large Eddy Simulation (LES) method. In displacement ventilation systems, clean and cool air is introduced into the lower portion of the room at a low velocity so not to cause draught discomfort. The cool supply air then spreads throughout the occupied zone, i.e. the zone occupied by people. Heat sources present in the room cause the air to rise due to thermal buoyancy. As the warming air rises, contaminants picked up along the way are carried above the occupied zone. The warm, dirty air is then exhausted in the upper portion of the room. Figure 1 illustrates the benchmark displacement ventilation setup used (Nielson et al. 2003). The room has dimensions of 2.5 m wide, 3.5 m deep and 3.0 m high (width, depth and height correspond to the x, y and z coordinates respectively). Air is supplied through a 0.4 m wide by 0.2 m high rectangular hole located on the floor and is discharged through a hole of equivalent shape and size at the ceiling on the

opposite wall. Measured from the top of the head, the CSP is located 1.75 m downstream of the inlet wall, centered in the xdirection and 1.75 m from the floor. Experimental data included PIV data measured in three windows. These three windows are the small gray rectangles indicated in Fig. 1. Within each window, velocity components in the plane of measurement were provided along a 0.2 m long line. These lines are labeled L1, above the head, L2, projecting from the center of the mouth and L3, projecting from the center of the torso. All three horizontal locations are in the x mid-plane. These test data are used for CFD validation. In the CFD calculations, the computational domain consists of the entire room shown in Figure 1. Several computational grids of different mesh size and topology are generated in this work (to be discussed in details later). The following boundary conditions were used for all simulations. The velocity and temperature at the inlet were 0.2 m/s and 22C respectively. A turbulence intensity of 30% and length scale of 0.1 m were experimental values suggested by Nielsen et al (2003). The only variable specified at the outlet was static pressure and was set to zero. The no-slip condition along with zero heat-flux (insulated) was applied to all the room walls. Along with the no-slip condition on the remaining boundary, the manikin surface, a boundary condition for the energy equation is required that approximates a human body. The human body is a dynamic thermoregulatory system (Fiala et al. 1999; Tanabe et al. 2002). It has the ability to adjust heat output in response to global and local changes in skin temperature. The objective of this work was not to predict the thermal response of a human but to evaluate the ability of CFD to predict the flow in the personal micro-environment. As a result, a thermoregulation model was not used to control the heat output of the body. Rather, a constant (spatially and temporally) heat-flux was chosen to mimic the heat exchange of the body with its surroundings. The net heat loss by a human is not only affected by their local environment but also can vary widely depending on gender, age and physiological makeup. The manikin used in this work represents a young adult female with a BMI in the normal range1. A characteristic heat loss for a human of this makeup is around 76 W. Using this value along with the surface of area of the manikin (1.48 m2), a heat-flux of 51.2 W/m2 is obtained. This value is the total heat-flux per unit area, where the fraction due to convection and radiation are determined during the calculation. It is widely accepted that roughly half of human heat loss is due to convective means while the remainder is due to radiation. An often used approach is to ignore radiation to avoid the perceived difficulties associated with radiation calculations. For cases in which only convection was modeled, half of the aforementioned heat-flux per unit area value was assumed (25.6 W/m2).
1.

Figure 1

Displacement ventilation room configuration.

BMI, body-mass index, is a measure of body fat based on height and weight.
47

ASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

VERIFICATION Following the definitions of Roache (1997), verification can be thought of as solving the equations right whereas validation would be solving the right equations. Verification involves quantifying the error induced by solving the chosen equations using discrete approximations. The total error includes those due to coding, discretization, grid convergence and iterative convergence. Because a commercial CFD code was used and access to the source code is not possible, code verification could not be performed. While, the author recognizes this as a potential source of error, due to the large FLUENT user base, it is assumed that any unreported coding errors negligibly affect the solution. A list of known resolved and unresolved errors in FLUENT are available on their supported user website. Solutions to the equations were computed using a well-documented second-order discretization scheme but no verification of the observed order of accuracy was performed. The author recognizes this as a potential source of error but believes it to be reasonable assumption again due to large FLUENT user base that accept the implementation of this discretization scheme to be second-order or close to second-order. The verification of FLUENT performed in this work involves grid and iterative convergence studies. Grid convergence will be demonstrated by comparing relevant quantities on the sequence of grids at strategically chosen locations. A more rigorous approach to determine grid convergence is the Richardson Extrapolation (RE) method (Richardson and Gaunt 1927). Two requirements of RE are constant grid refinement ratio and values at the same location. While the later is possible through interpolation, the former is not for the types of grids used in this work. Furthermore, the reliability of these methods on complex, unstructured grids remains unclear. As such error estimation using the RE method was not used. Iterative convergence for all solutions was performed and an example highlighting the strategies of monitoring convergence will be discussed. Grid Convergence Study Due to the unknown behavior of the global and local flow features in both cases, determining the appropriate grid strategy was non-trivial. Traditionally the number of grid cells has not exceeded several hundred thousands, due in large extent to lack of computational resources (Posner et al. 2003; Xing et al. 2001; Murakami 2004; Zhu et al. 2005; Srebric et al. 2007). In the present study, grids consisting of up to seven million cells were investigated. However, even with a seemingly indispensable amount of processing power, care should be exercised when determining the global resolution (total cells), local resolution (clustering around the CSP) and topology (cell type) for the grids in these scenarios. The approach to creating each grid began with meshing the surface of the manikin. A consequence of the complex
48

manikin surface was a non-uniform distribution of triangles. This was done to ensure all the details of the geometry were represented. Next, the walls, inlets and outlets were meshed uniformly. Finally, the volume grid was created where a growth rate factor (or the rate at which the cell size increases) of 1.2 or less was used to gradually increase the cell size away from the manikin. Some grids included an additional step where several layers of prismatic cells on the manikin were created to sufficiently resolve the boundary layer (called boundary layer grid here). In all cases, the equi-volume skewness values of the interior cells (or the quality of the cells) were monitored to ensure that they fall within the guidelines of the solver FLUENT. To establish grid independent solutions for the displacement ventilation case four grids were used. A summary of the four grids is provided in Table 1. The four grids consisted of two with strictly tetrahedral cells and two with the boundary layer cells added near the manikin. The important parameters that differentiate these grids include the number of triangles used to define the manikin geometry, the average y + value on the manikin surfaces2, and the type of grid (with or without boundary layer prismatic cells). Recall that the growth factor of the cell volume was kept under 1.2 and grid quality as measured by the equi-volume skewness was within the recommended value of the CFD solver. Illustrations of each grid type can be found in Sideroff (2007). The solutions from each grid used for the grid dependency study were obtained using the standard k-" turbulence model with enhanced wall treatment, and only the convective portion of heat-transfer. Since convection alone was included, only half the heat-flux value was applied to the manikin surface. The three measurement stations (head, face, and torso) detailed in Figure 1 are used to compare solutions from the four grids. Besides being located where the data were measured, these stations are meaningful because they penetrate the thermal plume at three distinctly different locations. Comparisons of the vertical velocity were made at the three stations illustrated in Figure 1 for the four grids, and the largest differences between these solutions occurred at the torso (station L3). These results are illustrated in Figure 2, which show that Grid C and Grid D appear to exhibit grid independence while Grid A and Grid B do not. In particular, the steep gradient of the near-wall velocity is difficult to capture when a strictly tetrahedral grid is used, unless of course extremely small tetrahedrons are used. Employing strictly tetrahedral to resolve the boundary layer would result in grids well in excess of ten million cells.
2.

y p u2 Where y + = ---------- : y p is the distance from wall adjacent cellv center to the wall, the friction velocity u2 = kinematic viscosity.
ASHRAE Transactions

3 wal l ----------- and % is the #

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Table 1.
Grid Label A B C D Number of Cells 1.14 (10)6 3.19 (10) 6.80 (10)
6

Displacement Ventilation Grid Summary


Number of Triangles on Manikin 36 (10)3 63 (10)
3

Boundary Layer NO NO YES YES

Average y+ 4.8 3.6 0.5 0.5

4.48 (10)6
6

140 (10)3 140 (10)


3

Figure 2

Vertical velocity at torso comparing grids (station L3.)

Iterative Convergence Study When steady-state RANS simulations are used, convergence is typically done by monitoring the average residuals during the iterative process of the transport equations (mass, momentum energy, and turbulence model equations). Residuals were monitored and typically dropped four orders of magnitude for the continuity equation (two to three orders lower for other equations). The number of iterations required to reach convergence varied but was at least several thousands. For the test case studied here, it was noted that average residuals alone were not good indicators of convergence. Averaging of the residuals can hide regions where the local residual may be many orders of magnitude larger. It was found that along with residuals, velocity magnitude should be monitored at strategically chosen points to give a more accurate indication that the solution had indeed converged. For the present problem, two points were chosen: one in the thermal plume (referred to as point A), and another away from and behind the CSP to monitor the room airflow (referred to as point B). During the iteration process, two distinct flow structures developed: the thermal plume, which was expected, and another not so obvious structure was a rather complex recirculating flow around the room and behind the CSP generated by the inlet vent. The momentum of the inlet vent created a low-speed jet that progressed along the floor, and instead of rising upon hitting the warm feet of the CSP, it proceeded
ASHRAE Transactions

along to the back wall. In fact, the flow circulated several times behind the CSP, three or more times in some instances, before being entrained in the thermal plume. The development of these two flow structures was not independent of each other thus making judgment of convergence non-trivial. As mentioned, FLUENT uses averaged residuals to monitor convergence of the transport equations. In the present problem, the continuity residual typically dropped about four orders of magnitude and leveled out after about 10,000 iterations, and in this case, falsely indicating convergence. It is noted that, since the velocity magnitude is highest in the thermal plume, the average residual is basically a convergence indicator of the flow structure in the thermal plume. When the velocity magnitudes at Points A and B were monitored during the iteration process, it was observed that convergence was not yet reached after 10,000 iterations. Figure 3 shows the velocity in the plume (Point A) declines monotonically, stops changing at about iteration 9,000, and remained constant until about iteration 10,000 where it gradually increases back to about 0.035 m/s from a value of 0.01 m/s. At iteration 14,000, the velocity magnitude began to slowly decline again and eventually leveling out to a value of 0.02 m/s at roughly iteration 27,000. Along the floor (Point B), the velocity magnitude made a couple of large oscillations during the first 7,000 iterations and then began to oscillate at a much higher frequency but at a constant mean value. Eventually the oscillations died away and the velocity became steady near iteration 27,000. During the first several thousands iterations (<9000), the plume and recirculating flow appear to develop independently. While the plume initially appeared to stabilize, the recirculating flow was still undergoing changes, as seen by the oscillations at point B in Figure 3. As the floor velocity continued to oscillate, the plume began to change again and thickened up, i.e., the increase in the velocity of point A in Figure 3 at iteration 10,000. Consequently, if only the development of the plume had been monitored, one would have incorrectly assumed that the solution had converged at around iteration 10,000th. Because two distinct interdependent flow structures developed, care must be taken when evaluating convergence. Along with residuals, velocity should be monitored during convergence in both the plume and ventilation flow, i.e., floor jet, to ensure a steady-state solution is obtained. It is noted that if the velocity around the CSP is the only quantity of interest, then the error may be small. However, if one is interested in the transport of contaminants into the CSPs breathing zone, then the error can
49

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Eddy-viscosity RANS model are the most popular class of turbulence model utilized for indoor environment simulations. They are attractive because they offer the best balance between accuracy, complexity and computational cost. The most popular of these, the standard k-" (Jones and Launder 1972), was used in this work. The transport equations for the turbulence kinetic energy, k , and turbulence dissipation rate, " , respectively are as follows:
%T Dk = $ * . . % + -----/ $k/ + 2% S 2 + " + P -----T b ,, Dt 4k - C 1" %T " D" ------ = $ * . . % + -----/ $"/ + -------- & 2%T S 2 + C 3" P b ' C 2" -,, - T T Dt 4"

(6)

(7)

Figure 3

Convergence history of velocity at Points A and B.

be very large if the sources of contaminants are located behind the CSP in the vicinity of the re-circulating flows. VALIDATION It is validation when one wants to determine how well the computational model can represent reality. Through verification, one gains confidence that the CFD code can produce solutions to the discrete equations of known error but provides no guarantee these equations sufficiently captures the physics of the problem. To validate CFD codes, experimental data must be used because experiments themselves do not selectively exclude the relevant physics. When validating a computational model, the user must now account for not only the error associated with the computational model but also the error associated with the experiment. Therefore error due to measurement uncertainty and error to facility biasing should be also accounted for. Katos research group at the University of Tokyo has obtained experimental data of the displacement ventilation case studied here. The setup of Katos experiment is similar to the computational model shown in Figure 1 where the manikin is suspended slightly above the floor. Two velocities components were measured with a particle image velocimetry (PIV) system near the manikin and values for each were provided along the lines indicated in Figure 1. Velocity magnitude, turbulent intensity and temperature were measured at several locations away from the manikin including the inlet. Unfortunately no estimates of uncertainty or error were provided. Turbulence Modeling Due to our inability to compute the entire range of flow scales for all but the most trivial problems, modeling of turbulence has become a research science of its own. The most ubiquitous approach to modeling turbulence is the RANS method while the more robust but computationally demanding approach is the LES method. Two RANS models as well as an LES model were investigated in this work.
50

5T where %T = C 0 k is the turbulent viscosity, Pb = g+T ( ----- is " 5z k is the time thermal buoyancy induced production and T = " scale. Complete details of these equations, including the model constants, can be found in many other books on turbulence (Pope 2000; Tennekes and Lumley 1972).

To extend the applicability of the standard k-" model, FLUENT has made available an enhanced wall treatment option. The enhanced wall function is a near-wall modeling approach that enables the standard k-" equations to be integrated all the way to the wall. Essentially, this provides the standard k-" with the ability to resolve the boundary layer through the buffer layer into the viscosity affected sub-layer without the need for an explicit wall function. To allow this, a two-layer model is used to specify " and % t in the viscous sublayer. Providing the standard k-" with the ability to resolve the boundary layer imposes requirements on y + , i.e. y + less than one. Complete details of the enhanced wall treatment are found in the FLUENT manual. A state-of-the-art RANS model, dubbed the % 2 -f , with some noteworthy improvements over the standard k-" , was also used in this work. Regardless of the ability to full resolve the boundary layer, it is well known that the standard k-" model over-predicts the production of turbulent kinetic energy near solid walls. The performance of the k-" definition of the eddy viscosity degrades rapidly as solid walls are approached primarily due to incorrect scaling of turbulent kinetic energy. The % 2 -f model of Durbin (1991) changes the definition of the 2 velocity scale to & v' 2 ' 1 1 2 primarily because v' scales as y 4 as it approaches solid walls whereas k scales as y 2 (Durbin 1991); thus providing stronger damping as the wall is approached. Furthermore, an elliptic relaxation function, f , is included to account for the non-local damping effects of solid boundaries. Through this approach the elliptic relaxation function can account for anisotropy present near solid boundaries. There have been various modifications to the % 2 -f model and the versions of the v 2 and f equations incorporated in FLUENT are identical to Model 3 of Sveningsson (2003). Along with the Equations 6 and 7, the remaining equations that constitute this version of the % 2 -f are as follows:
ASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

%T Dv' 2 = $ * . . % + -----/ $v' 2/ + 6v' 2 " + kf ----------,, k Dt 4k 2 -- v' 2 1 k 2%T S 2 5v' 2 1 k 3 2 $ 2 f = & C 1 ' --------------------- + C -------------- + --------------fL f1 f2 T k T

(8)

Kolmogorov length ( 6 ) and time ( 3 ) scales are be expressed in terms of the kinematic viscosity, %.
%3 1 1 4 6 = . ---- / , "; % 112 3 = . --/ , "-

(10)

(9)

where % t = C 0 v' 2 T is the turbulent viscosity. The time scale, T, defined to allow adjustment near walls and model constants can be found in Sveningsson (2003). While the standard k-" employs ad-hoc wall treatments, the % 2 -f model instead handles the near-wall behavior through a more physical approach: the v 2 and f equations. Without the need for wall functions, the % 2 -f turbulence model requires that the boundary layer be fully resolved and y + should be less than one. Eddy-viscosity RANS models, such as the standard k-" and % 2 -f , are known to be deficient for low Reynolds numbers flow. The majority of the assumptions on which they are based can be traced to the fully turbulent approximation. Furthermore, turbulence production from thermal gradients is a particularly challenging issue for typical RANS models. While there has been considerable effort to extend the capabilities of traditional RANS models (Murakami et al. 1996; Kenjere et al. 2002, 2005), these new models tend to be overly complex requiring additional equations, more coefficients and typically introduce numerical instability. Despite their achievements, these models undoubtedly fall well short of being reliable for a wide range of indoor environment flows. With this in mind, it seems reasonable to consider models beyond the RANS approach. LES is a more robust method of approximating complex turbulent flows than RANS methods. In contrast to RANS modeling, Large Eddy Simulation (LES) computes the large-scale motions of the flow directly. The small-scale, dissipative motions of turbulence tend to more amenable to modeling because of their more uniform character, whereas the large-scale motions contain the majority of the energy and anisotropy. As a result, LES is expected to be more accurate, particularly in complex flows where the assumptions inherent to RANS models rarely exist. The drawback is that LES simulations are always three dimensional, unsteady one. Provided the boundary layer is resolved and y+ is less than one, a linear stress-strain relationship is assumed for LES. If however the first cell does not lie well within the viscous sub-layer, it is assumed the first cell lies within the logarithmic layer and the traditional law-of-the-wall is used. LES typically does not suffer the drawbacks associated with RANS models provided the energy-containing length scales are sufficiently resolved. The question is how to ensure this happens without a priori knowledge of the flow. Because LES computes the large scale motions and models the small scale ones, knowledge or even approximation of the small scales would be beneficial. From known values of the length and time scales, estimates of the smallest scales, the Kolmogorov scales (Tennekes and Lumley 1972), can be made. The
ASHRAE Transactions

In determining 6 and 3 , only an estimate for e remains. Following Tennekes and Lumley (1972), an inviscid estimate for " can made through the relationship " = u' 3 1 l . The fluctuating velocity, u' , is determined by the relation u' = UI . Here, U would be the peak plume velocity, 0.3 m/s. The turbulent intensity, I = 30% , and turbulent length scale l = 0.03 m of the plume were obtained from Marr and Glauser (2006) and Marr (2007). Using these values, " = 0.0243 m2/s3. From this, the estimated Kolmogorov scales of the thermal plume are 6 = 6.1 & 10 ' 4 m = 0.61 mm and 3 = 0.025 s . Now that approximations of the smallest scales are available, determination of the appropriate grid spacing and time step for LES can be made. If a DNS of this flow was performed, a grid spacing and time step would be chosen to match the Kolmogorov values. However, since LES models the small scales a larger grid spacing and time step can be used. Meyers et al. (2003) recommends the LES sub-grid scale cutoff, i.e., grid spacing, to be around 1020 Kolmogorov scales. This would suggest a grid spacing of about 6 mm would be sufficient for LES. The time step would be determined in a similar manner but numerical concerns must also be considered. The recommendation of Meyers et al. (2003) applied to the time step would yield 0.25 s but the flow time for a 6 mm cell with a velocity of 0.3 m/s would be 0.02 san order of magnitude smaller than 103 . It is unlikely that the maximum velocity would occur in the smallest cell, since the smallest cells are located adjacent that walls. The appropriate time step size would lie somewhere between 0.02 and 0.25. The dynamic Smagorinsky sub-grid scale model of Germano et al. (1991) was used in this work. The sub-grid scale viscosity and length scale are defined as
% sgs = L s 2S S L s = min(7 &8p 9 C s ) '

(11) (12)

where the over-tilde denotes spatial average (as opposed to temporal averaging for RANS models) and the filter width, ) = :1 1 3 . Complete details along with the model constants cell can be found in Germano et al. (1991). As previously mentioned, results from the turbulence models were compared with the experimental data at the locations indicated in Figure 1. Upon examining the vertical velocity profiles in Figure 4 above the head, significant differences are revealed between not only the turbulence models but also the experimental values. The general profile shapes appear reasonable, but only a small the portion of the profile near the head (<0.03 m) is captured by the % 2 -f model. Apart from this, both RANS models undeniably over-predict the data. At the last data point, the standard k-" exceeds the data by 0.05 m/s, and the
51

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Figure 4

Vertical velocity above head (station L1).

Figure 5

Vertical velocity at face (station L2).

% 2 -f by 0.1 m/s. It is evident from Figure 4 that LES provides

marginal improvement. Near the head, the LES profile is consistent with the slope of the data. Beyond the near-wall region, the LES profile compares better to the test data than either RANS models, but the improvement is marginal. The LES results overpredict the magnitude by as much as 0.05 m/s. At the face shown in Figure 5, the RANS results yielded considerably different trends. Overall, the k-" model underpredicts the thickness of the thermal plume, while the % 2 -f model over-predicts it. Near the wall (<0.02 m), the standard k-" predicts a peak velocity of 0.1 m/s and has a more dramatic change than the data shows. The % 2 -f profile peaks at the highest value (~0.13 m/s) and maintains velocity further away than the standard k-" predicts or the data shows. Finally, the LES results appear to yield marked improvement over the RANS models, although the comparison is off in the near-wall region (<0.02 m). At the torso station shown in Figure 6, the vertical velocity profiles show similar patterns observed at the face. The k-" model under-predicts the thickness of the thermal plume, while the % 2 -f model over-predicts it. Both RANS models predict a maximum velocity of about 0.25 m/s, whereas the experiment shows a peak value of about 0.16 m/s. Upon examination of the LES results in Figure 6, the anticipated advantage is once again less clear. The LES peak value is overpredicted but not to the level of the RANS models. The trend in the LES profile beyond the peak region appears to be a blend of the two RANS profiles. Radiation While the LES results showed some improvements over the RANS results, comparisons were still not satisfactory. Often attributed to the effort required to compute radiant energy exchange, a common approach is to neglect radiative heat transfer and assume half the heat loss to convection. Due to the nonlinear coupling of radiant and convective surface energy transfer to the fluid dynamics equations, it becomes apparent that simply neglecting half the heat loss may be erro52

neous. In this case, the inclusion (or exemption) of radiative heat-transfer warranted further investigation. To examine the impact of radiation, the standard k-" , % 2 -f and LES calculations were run that included surface-tosurface radiation modeling. The surface-to-surface radiant energy flux is computed from a system of N ; N equations where N is the number of participating surface elements in this case the number of triangles on the interior surface mesh. This system of equations also requires N2 view factors which are solely dependent on the geometry and can be computed prior to the simulation. The radiant energy fluxes are computed every time the flow equations are computed. A concise description about how the heat fluxes and temperatures are computed when radiative heat transfer is included can be found in Sideroff (2007). As mentioned earlier, it is generally accepted that roughly half of the human heat loss is due to convection and half is due to radiation (in cases where latent heat transfer is negligible). To confirm this assumption, the actual fractions of each were computed. When the total heat-flux is specified the contribution from each mode is calculated a posteriori. Using the standard k-" result from grid D, the fraction due to convection was calculated to be 38% while the remaining 62% was due to radiation. With this result, it is intuitive then to consider whether using a value of 38% of the total heat loss, instead of the 50% previously assumed, would yield a more reasonable result for the convection only calculation. To answer this, one has to consider the coupling of the surface-to-surface radiation with the flow more thoroughly. The inclusion of radiation has consequences beyond the balance of the heat transfer modesradiation also affects the surrounding surface temperatures and in particular the room walls. Temperature contours on the surfaces shown in Figures 7 and 8 make evident that the inclusion radiation significantly changes the surface temperatures. Without radiation (Figure 7), the manikin has noticeably larger regions of high temperature as opposed to the radiation case (Figure 8), which has a much more uniform distribution. Additionally, the temperatures of
ASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Figure 6

Vertical velocity at torso (station L3).

Figure 7

Surface temperatures without radiation; LES model on grid D.

without radiation included. Here, it could be argued the LES provides no considerable benefit over the standard k-" or % 2 -f . All numerical profiles show roughly the same peak value albeit higher than the data0.125 m/s for the % 2 -f , 0.11 m/s for LES and standard k-" versus 0.08 m/s for the data. However, beyond about 0.03 m the numerical profiles follow the data remarkably close. The final location at the torso indicates similar trends as seen at the face. As shown in Figure 11, the peak value is overpredicted by both models but beyond there they follow the data reasonably well. The standard k-" does not predict the change in slope seen in the data or the % 2 -f and LES but overall yields acceptable results. Reduced-Order Model for Radiation Figure 8 Surface temperatures with radiation; LES model on grid D. The calculations with radiation provided evidence that radiation modeling cannot be ignored, but the inclusion of radiation modeling did not come without a cost. For surfaceto-surface radiation calculations, the view factors are needed. Although purely a pre-processing step, computing the view factors can be computationally expensive (Sideroff 2007). It would be beneficial if the effects of radiation could be included without actually performing the radiation modeling. If the surface temperatures were somehow known a priori, the effects of radiation can be included without the difficulties of actually computing radiative energy transfer. Unfortunately, determining a detailed spatial (or temporal) description of the surface temperature from experiment is exceedingly difficult or simply not possible, nor it is the case in practice. It was postulated that perhaps a reduced-order or loworder description of the surface temperature would be adequate. And if a low-order description of the surface temperature were used, how would this affect the flow in the PME? To answer this question, an examination of the impact of a reduced surface temperature description was carried out. The simplest way is to reduce the surface temperature of each surface to an average value.
53

side walls and floor without radiation are several degrees lower than with radiation, while the converse is true on the ceiling. Indubitably, the incorporation of radiation has consequences on not only on the individual contribution of heat transfer modes but also on the surface temperatures. To evaluate the effect of radiation more definitively, results were compared to the test data at the three locations indicated in Figure 1. Upon inspection of the results of Figure 9 above the head, it is immediately obvious the influence radiation has on the thermal plume. The over-prediction of the magnitude seen in Figure 4 without radiation appears to be remedied by the inclusion of radiation. The LES profile is in excellent agreement with the data, and the standard k-" and % 2 -f yield more than satisfactory results. The results at the face with radiation included shown in Figure 10. While not in agreement to the degree seen above the head, these results still provide improvement over predictions
ASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Figure 9

Vertical velocity above head with radiation (station L1).

Figure 10 Vertical velocity (station L2).

at

face

with

radiation

Figure 11 Vertical velocity (station L3). Table 2.

at

torso

with

radiation

Figure 12 Vertical velocity at torso comparing radiation modeling and averaged surface temperatures (station L3).
k-" with the unaltered surface temperatures from radiation modeling, along with the test data. Comparing the solutions at the three stations where test data are available, it was found that excellent agreement were obtained at station L1 (above the head) and station L2 (at face). At station L3 or the torso shown in Figure 12, more noticeable differences are observed. The peak value of the profile with the reduced-order model is marginally lower than with the full radiation modeling included. However, the shapes of the profiles are very similar.

Average Surface Temperatures


Average Temperature (K) 303.9 297.9 298.5 298.3 298.3 298.4

Surface manikin floor ceiling side walls front wall rear wall

Using the solution from the standard k-" on grid D, an area-weighted average of the temperature distribution was computed for each surface. The resulting average surface temperatures and those subsequently used for boundary conditions are tabulated in Table 2. A calculation using the standard k-" on grid D was performed using the average surface temperatures from Table 2 and compared against the results from the standard
54

CONCLUSIONS A detailed verification and validation study using a commercial CFD code (FLUENT) of the flow around a Computer Simulated Person in a displacement ventilation room was carried out. Following the guidelines of the benchmark displacement ventilation case of Nielsen et al. (2003) several recommendations concerning the verification and valiASHRAE Transactions

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

dation of CFD for the PME were elucidated by examining grid dependency, iterative convergence, turbulence modeling and radiation modeling. Four grids of increasing resolution and varying topologies were constructed to demonstrate grid independence. Due to the complex geometry of the CSP, the grids used were of the unstructured type. It was found that achieving grid independent solutions while maintaining an acceptable cell count using a strictly tetrahedral topology was exceedingly difficult. The two finest grids included a technique where several layers of thin triangular prism cells were created by extruding the surface triangles away from the CSP. These layers of prismatic cells allowed full resolution of the boundary layer without significant increase in the number of cells. To achieve grid independent solutions, approximately 100,000 surface triangles are needed on the CSP, at least several layers of prismatic cells around the CSP to achieve y + values less than one and a maximum growth rate of 1.2 for the remaining tetrahedral cells around the CSP. To achieve convergence for steady-state (RANS) calculations, quantities other than average residual needed to be monitored. Because two distinct flow structures developed, averaged residual was not a sufficient convergence parameter. Velocities were monitored at strategically chosen pointsone in the plume and the other away from the CSP in the recirculating room flowduring the iteration process. While the velocity in the thermal plume appeared to reach steady-state at approximately the same number of iterations as the residuals this was not so. The velocity of the recirculating flow near the floor continued to oscillate for several thousand iterations beyond the apparent steady state of the plume. Nearly 30,000 iterations were needed to achieve convergence. It was noted that if the velocity in the PME of CSP were of interest, the error incurred by incomplete convergence may be inconsequential. However, if, for example, contaminant transport from a source away from the CSP in the recirculating flow were of interest, iterative convergence error may be important. Due to the perceived effort required to include it, radiation modeling is commonly avoided by assuming half the CSP heat loss for convection only. This approach requires using a heat-flux (first derivative) boundary condition. Through careful examination of the coupling of surface-tosurface radiation equations to the flow equations and results that include radiation modeling it was found that neglecting radiation modeling when heat-flux boundary conditions are used is erroneous. While the actual radiative to convective heat transfer ration was closer to 60/40 (not 50/50), it was concluded that the significant change in wall temperatures due to radiation caused the differences and not the lower convective component. If somehow the actual surface temperatures were known a priori, then the effects of radiation could be included without actually including a radiation model. The actual surface temperatures would need to be obtained from experiment because reality does not disclude radiation. However, obtaining a detailed description of the surface temperatures experimentally would be exceptionally difficult or simply not possible. By applying spatially averaged values of surface
ASHRAE Transactions

temperatureobtained from the calculations with radiation includedit was found that velocities in the PME were relatively insensitive to this approach. This suggests that measuring the surface temperature at few points on all the participating walls and using those as temperature boundary conditions for CFD calculations would yield satisfactory solutions. Turbulence models tend to be the most unreliable aspect of CFD, particularly for low Reynolds, thermally buoyant flows encountered in the PME. Results from turbulence models were validated with high-resolution PIV data supplied by Katos research group at the University of Tokyo. The standard k-" with an enhanced wall-treatment, the % 2 -f model of Durbin (1999) and dynamic Smargorinsky LES turbulence models were utilized. Without radiation (or the effects) included none of the three turbulence models provided satisfactory results. Only when radiation (or the effects) were included did any turbulence yield reasonable comparison to the data. Regardless of the innovative improvements of the % 2 -f , it did not yield improvement over the standard k-" . LES does not suffer from the drawbacks of RANS models but despite its wider applicability did not offer any benefit over the standard k-" . When an enhanced wall treatment was used with a grid that resolves the boundary layer, the standard k-" model was found to provide solutions as accurate as the % 2 -f or LES. ACKNOWLEDGMENTS This work was supported by the US Environmental Protection Agency (EPA) through the STAR Center for Environmental Quality Systems (www.eqstar.org) and the Center of Excellence in Environmental Systems (www.coees.org) at Syracuse University. REFERENCES Awbi, H.B. 1991. Ventilation of Buildings. E & FN Spon Publishing. Bjrn, E. and Nielsen, P.V. 2002. Dispersal of exhaled air and personal exposure in displacement ventilated rooms. Indoor Air 12(2): 147-164. Cermak, R. Melikov, A. Forejt, L. and Kovar, 0. 2006. Performance of personalized ventilation in conjunction with mixing and displacement ventilation. International Journal of HVAC Research. 12: 295-311. Cheong, K.W.D., Djunaedy, E., Pho, T.K., Tham, K.W., Sekhar, S.C., Wong, N.H. and Ullah, M.B. 2003. Measurements and computations of contaminants distribution in an office environment. Building and Environment. 38: 135-145. Clark, R.P. and Edholm, O.G. 1985. Man and His Thermal Environment London, UK: Edward Arnold. Durbin, P.A., 1991. Near-wall turbulence closure modeling without damping functions. Theoretical and Computational Fluid Dynamics 3: 1-13. Fenske, J.D., Paulson, S.E. 1999. Human breath emissions of VOCs. Journal of Air and Waster Management 49: 594598.
55

2008, American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions Vol. 114, Part 2. For personal use only. Additional reproduction, distribution, or transmission in either print or digital form is not permitted without ASHRAEs prior written permission.

Ferro, A., Kopperud, R.J., and Hildeman, L. 2004. Source strengths for indoor human activities that resuspend particulate matter. Environmental Science & Technology 38(6): 1759-1764. Fiala, D., Lomas, K.J. and Stohrer, M. 1999. A computer model of human thermoregulation for a wide range of environmental conditions: the passive system. Journal of Applied Physiology 87(5): 1957-1972. Germano, M., Piomelli, U., Moin, P. and Cabot, W.H. 1991. A dynamic sub-grid scale eddy viscosity model. Physics of Fluids A 3: 1760-1765. Huang, J.M., Chen, Q., Ribot, B., and Rivoalen, H. 2004. Modeling contaminant exposure in a single-family house. Indoor and Built Environment 13: 5-19. Jones, W.P. and Launder, B.E. 1972. The prediction of laminarization with a two-equation model of turbulence. International Journal of Heat and Mass Transfer 15: 301-314. Kenjere, S., Hanjalic, K. and Gunjaro, S.B. 2002. A TRANS/VLES approach to indoor climate simulations. Proceedings of ASME 2002 Fluids Engineering Division Summer Meeting, Montreal, CA. Kenjere, S., Gunjaro, S.B. and Hanjalic, K. 2005. Contribution to elliptic relaxation modelling of turbulent natural and mixed convection. International Journal of Heat and Fluid Low 26:. 569-586. Marr, D.R. and Glauser, M.N. 2006. Length scale propagation along a joint inlet and thermal buoyancy driven Flow. AMSE International Meeting Paper 98532. Marr, D.R. 2007. Length scale propagation along a joint inlet and thermal buoyancy driven flow. PhD Thesis, Syracuse University, Dept. of Mechanical and Aerospace Engineering. Melikov, A. and Kaczmarcyzk, J. 2007. Measurement and prediction of indoor air quality using a breathing thermal manikin. Indoor Air 17: 50-59. Meyers, J. Geurts, B.J. and Baelmans, M. 2003. Database analysis of errors in large-eddy simulation. Physics of Fluids 15(9): 2740-2755. Modest, M.F. 2003. Radiative Heat Transfer. San Diego, CA, USA: Academic Press. Murakami, S., Kato, S., Chikamoto, T., Laurence, D. and Blay, D. 1996. New low-Reynolds number k-e model including damping effects due to buoyancy in a stratified flow field. International Journal of Heat and Mass Transfer 39(16): 3483-3496. Murakami, S. 2004. Analysis and design of micro-climate around the human body with respiration by CFD. Indoor Air 14(Suppl. 7): 144-156. Nielsen, P.V., Murakami, S., Kato, S., Topp, C. Yang, J-H. 2003. Benchmarks test for a computer simulated person.

Aalborg University, Indoor Environmental Engineering (see http://www.cfd-benchmarks.com). Pope, S.B. 2000. Turbulent Flows. Cambridge, UK: Cambridge University Press. Posner, J.D., Buchanan, C.R. and Dunn-Rankin, D. 2003. Measurement and prediction of indoor air flow in a model room. Energy and Building. 35: 515-526. Richardson, L.F. and Gaunt, J.A. 1927. The deferred approach to the limit. Philosophy Transactions of the Royal Society of London, Ser. A. 226: 299-361. Roache, P.J. 1997. Quantification of uncertainty in computational fluid dynamics. Annual Review of Fluid Mechanics 29: 123-160. Sideroff, C.N. 2007. Detailed examinations of the human micro-environment by CFD. PhD Thesis, Syracuse University, Dept. of Mechanical and Aerospace Engineering. Srebric, J., Vokovic, V., He, G. and Yang, X. 2007. CFD boundary conditions for contaminant dispersion, heat transfer and airflow simulations around human occupants in indoor environments. Building and Environment Article in Proof. Sveningsson, A. 2003. Analysis of the performance of different -f turbulence models in a stator vane passage flow. Ph.D Thesis, Chalmers University of Technology ISSN 1101-9972. Tanabe, S., Kobayashi, K., Nakano, J., Ozeki, Y., and Konishi, M. 2002. Evaluation of thermal comfort using combined multi-node thermoregulation (65N) and radiation models and computational fluid dynamics (CFD). Energy and Buildings 34: 637-646. Tennekes, H. and Lumley, J.L. 1972. A First Course in Turbulence. Cambridge, MA, USA: MIT Press. Topp, C. 2002. Influence of geometry of a computer simulated person on contaminant distribution and personal exposure. ROOMVENT 2002. 1: 265-268. Topp, C., Nielsen, P.V., and Srensen, D. 2002. Application of computer simulated persons in indoor environmental modeling. ASHRAE Transactions 108(2): 1084-1089. Xing, H., Hatton, A. and Awbi, H.B. 2001. A study on the air quality in the breathing zone in a room with displacement ventilation. Building and Environment 36: 809820. Zhang, T. and Chen, Q. 2007. Novel air distribution systems for commercial aircraft cabins. Building and Environment 42: 1675-1684. Zhu, S., Kato, S., Murakami, S. and Hayashi, T. 2005. Study on inhalation region by means of CFD analysis and experiment. Building and Environment 40: 1329-1336.

56

ASHRAE Transactions

You might also like