Professional Documents
Culture Documents
Abstract: Aerosol dispersion in living spaces especially bioaerosols, due to accidents or deliberate acts, is of significant current interest.
Computational fluid dynamics 共CFD兲 provides an accurate and detailed platform to study the influence of different parameters on aerosol
distribution in indoor spaces. The simulations however are time consuming and site-specific. The work here introduces an approach
toward addressing this challenge. During emergencies, an accurate, quicker, and more general model is required to give rapid answers to
first responders. Significant parameters influencing aerosol behavior in an office room were identified and through dimensional analysis,
nine dimensionless groups were developed. Fractional factorial design was used to build sixteen scenarios to explore the design space.
These scenarios were then simulated using a comprehensive CFD model. Large eddy simulation with the Smagorinsky subgrid scale
model was applied to compute the airflow. Aerosols were modeled as a dispersed solid phase using the Lagrangian treatment. The
influence of the dimensionless groups on the temporal variation of the number of aerosols in the room and the spatial distribution of the
particles in the room was analyzed. The results showed that all the identified dimensionless groups were significant. Multiple linear
regression models were developed for the prediction of the number of aerosols in the room and their spatial distribution as a function of
the significant parameters influencing aerosol transport. The linear models accurately predicted the data on which they were based but did
not predict the results of the independent tests as well. The limited predictive ability of the model showed that the relationships between
the dimensionless groups are nonlinear and a higher level of experimental design will have to be applied to better explore the design
space.
DOI: 10.1061/共ASCE兲EE.1943-7870.0000201
CE Database subject headings: Air flow; Computational fluid dynamics technique; Design; Dimensional analysis; Regression
models.
Author keywords: Aerosols; Air flow; Computational fluid dynamics; Design; Dimensional analysis; Regression models.
Fig. 1. Geometry of the representative system The 3D indoor turbulent flow was modeled here by large eddy
simulation 共LES兲. LES is a technique between direct numerical
simulation 共DNS兲 and Reynolds averaged Navier-Stokes 共RANS兲.
tern of the influence of the independent variables on the result. DNS is the most accurate of the three methods but it requires high
The process requires time and accuracy to ensure that the design computing power 共high-speed and large capacity computer兲 for
space is explored efficiently and the results can be applied for airflow simulation in a building 共Beghein et al. 2005兲. RANS
predictive purposes. To determine which parameters influenced modeling applies different turbulence models to calculate the
the ultraviolet germicidal irradiation 共UVGI兲 air disinfection sys- Reynolds stresses and close the system of Navier-Stokes equa-
tems, Kowalski et al. 共2003兲 applied dimensional analysis to tions. It generates the mean velocity field and not the instanta-
group the parameters and used experimental design to decide on neous velocity field, which influences particle dispersion
the number of computer simulations to be conducted. significantly. LES generates instantaneous flow information re-
In the present work, comprehensive CFD simulations are ap- quired for particle simulation 共Beghein et al. 2005; Tian et al.
plied to explore the multidimensional design space of the particle 2007兲. In LES, small-scale turbulence is filtered out from the
dispersion system in an office room. The study develops a de- Navier-Stokes equations, and a model is used to evaluate small
tailed CFD model which is applied to simulate multiple scenarios scales. The size of the filter is equal to the smallest grid size. The
of the problem. The different conditions for each simulation were filtered Navier-Stokes equations are solved for the large-scale mo-
decided by applying fractional factorial design principles. Dimen- tion, which is responsible for most of the momentum and energy
sional analysis was used to obtain dimensionless groups or factors transport. The small scales are from subgrid-scale 共SGS兲 models,
from the multiple parameters. The study identifies the important which in turn influence the large eddies. For incompressible flow
dimensionless groups and determines the significant factors influ- of a Newtonian fluid the filtered Navier-Stokes equations are
encing particle transport and removal. Finally a multiple linear
regression model is developed.
ūi
=0 共1兲
xi
Methods
To develop the model, the following steps were performed: ūi 1 p̄ ij 2ūi
+ 共ūiū j兲 = − − + t 共2兲
1. Development of the CFD model; t xj xi x j xj xj
2. Dimensional analysis;
3. Experimental design: Fractional factorial design; and The effect of the small scales appears through the SGS stress
4. Simulation steps. term, ij 共Piomelli 1999兲. The small scales are more uniform and
Geometry of the three-dimensional 共3D兲 room is shown in Fig. isotropic than large-scales. The most commonly used SGS stress
1. Fig. 1 is a representative scenario of the problem where aero- model is the Smagorinsky model which is an eddy-viscosity
sols are released at a point location within a ventilated room. The model of the form, ij − 共1 / 3兲kk␦ij = 2tS̄ij. The SGS stresses are
room is bare of any furnishings or partitions. Air is flowing in related to the large-scale strain rate tensor by S̄ij = 共ūi / x j
through one end of the room and is exiting at the opposite end. + ū j / x j兲 / 2. The eddy viscosity, t is obtained from the algebraic
Infiltration and exfiltration of air is assumed negligible. To define model, Eq. 共3兲, where ⌬ is the grid size and Cs is the Smagorin-
the system, thirteen dimensional variables were selected. “L” 共m兲, sky constant
“Hr” 共m兲, and “W” 共m兲 represent the length, height, and width of
the room. The vertical distance between the inlet and outlet of
area “A” 共m2兲 is denoted by “I” 共m兲. I can be either negative or t = Cs⌬2兩S̄兩S̄ij 共3兲
positive. A positive value of I indicates that the inlet is above the
outlet and a negative value indicates the reverse case. The inlet air
velocity is “U” 共m/s兲. The initial number of the particles released
兩S̄兩 = 共2S̄ijS̄ij兲1/2 共4兲
in the room is “No” 共particles/ m3兲. The final number of particles
remaining in the room after time 共s兲 is denoted by N Near solid boundaries or in transitional flow this needs to be
共particles/ m3兲. The properties of air considered are density, “ f ” adjusted and has been accomplished by ad hoc corrections such as
共kg/ m3兲 and dynamic viscosity, “” 共N / m2 · s兲. The aerosols are Van Driest damping 共Piomelli 1999兲. The Smagorinsky constant
characterized by the diameter, “d p” 共m兲 and density, “ p” 共Cs兲 is 0.16. In the Smagorinsky model the eddy-viscosity coef-
共kg/ m3兲. ficient is maintained constant for the entire flow domain.
1
FG = d3p共 p − f 兲g
6
共7兲 冉 冊 冉
N
No
=
L L I WHr Q f L p
, , , , , , ,
W Hr Hr A L d p f
冊 共10兲
In Eq. 共10兲 the first four groups take into account the dimensions
The drag force is represented by the Eq. 共8兲 where, CD = drag
of the room, the locations of the inlet and outlet and the area of
coefficient, uជ = velocity of air flow, f = density of the fluid, and
the inlet and outlet. The fifth group which is very similar to the
Cc = Cunningham correction factor
Reynolds number defines the flow characteristic in relation to the
room dimensions. Q 共m3 / s兲 is the volumetric flow rate of the air.
FD = 共1/8Cc兲d2p f CD兩uជp − uជ 兩共u
ជp − uជ 兲 共8兲 The sixth group is the ratio of the particle tracking time, 共s兲, to
the theoretical residence time of the room, 共s兲,
The Cunningham correction factor is a function of the Knudsen = V共m3兲 / Q共m3 / s兲. It is the same as the mean residence time for a
number and is determined from Cc = 1 + 共2 / d p兲共1.257 CMFR 共completely mixed flow reactor兲 共Nazaroff and Alvarez-
+ 0.4e−1.1dp/2兲 where d p is the particle diameter. For particle mo- Cohen 2002兲. Similar to the CMFR, air would reside for different
tion with Reynolds’ number less than 0.1, CD = 24/ Re, when the length of times at different locations in a room. Some of the air
Reynolds’ number is from 1,000 to 2 ⫻ 105, CD ⯝ 0.44. In be- would flow out at a time period shorter than the residence time of
tween, the correct value of CD is calculated from the following the room while another portion of the air would reside for much
expression: CD = 共24/ Re兲共1 + 0.16 Re2/3兲共Hinds 1999兲. longer than the residence time. The second last group compares a
particle property to the length of the room while the last group is
the ratio of the density of the particles released to the density of
Numerical Scheme and Boundary Conditions
the fluid.
The control-volume method was employed to discretize the gov-
erning equations. A second order upwind scheme was applied for
the convective-diffusive terms and the Euler first order implicit Experimental Design: Fractional Factorial Design
scheme was used for the temporal terms in the governing equa-
tions. The outlets were designated as pressure boundaries, with Based on Eq. 共10兲 the independent variables or factors are the
Dirichlet conditions specified for pressure and Neumann condi- eight dimensionless groups: 共L / W兲, 共L / Hr兲, 共I / Hr兲, 共WHr / A兲,
tions 共i.e., zero normal gradients兲 specified for all other dependent 共Q f / L兲, 共 / 兲, 共L / d p兲, and 共 p / f 兲. The dependent variable is
variables. The boundary conditions at the walls of the room rep- the dimensionless group, 共N / No兲. To determine the significance of
resent no-slip conditions. The particles were initialized over a flat these eight factors and to estimate the magnitude of their indi-
surface slightly above the floor to avoid the influence of the vidual effects as well as their interacting effects on the response
boundary layer. Particle trajectory calculations are terminated variable, a series of simulations have to be conducted. Factorial
when the particles reach the exit of the room. When a particle designs are efficient exploratory tools applied to achieve these
strikes a surface it is assumed that it loses half its energy as it objectives with as few simulations as possible. For the present
bounces back. Following assumptions were applied: case a two-level full factorial design would require 28 = 256 simu-
1. Incompressible and isothermal conditions; lation runs or cases. Two levels are typically enough to determine
2. No heat and mass transfer between the air and the particles; if a factor is significant or not 共Bingham 1997兲. Since a single
3. The coefficient of restitution was set to 0.5 to account for simulation would require at least 24–48 h to run on a single PC
some energy loss when the particle strikes the surface; the full factorial design is currently impractical. Instead a 16-case
4. Spherical, solid and neutral particles; fractional factorial design, 28−4 was developed 共Box et al. 1978;
5. Airborne particles has no influence on the surrounding air Montgomery 2005兲. Fractional factorial designs take advantage of
flow field, i.e., one-way coupling; the fact that higher order interactions often tend to become neg-
6. Negligible particle-particle interactions; and ligible. A common criterion based on which fractional factorial
7. Negligible resuspension. designs are generated is “resolution” 共Ankenman 1999; Box et al.
1978兲. Resolution seeks to reduce confounding between lower minimum value of the flow rate Q was chosen slightly below the
order effects. For the Resolution IV design, no main effect is ASHRAE standard of 0.009 m3 / s person. The residence times,
confounded with any two-factor interaction but two-factor inter- of the room was calculated from the air changes per hour 共ACH兲
actions are confounded with other two-factor interactions. The range of 1–10. This range accommodates the required ventilation
8−4
design chosen here is a Resolution IV design, 2IV . A higher res- rate for houses, offices, and hospitals. The inverse of the resi-
olution design could have been chosen but at the expense of more dence time of a room, gives the number of ACH. Minimum
simulation time 共64 cases兲.
particle tracking time, was equal to the minimum residence time
Table 1 illustrates the design matrix. Each of the dimensionless
group is labeled with a coded variable: A, B, C, D, E, F, G, and H. of the room while the maximum was chosen to be five times the
The design generators chosen are E = +BCD, F = +ACD, G = maximum residence time. Based on these values the correspond-
+ABC, and H = +ABD 共Montgomery 2005兲. For A, B, C, and D a ing dimensionless groups were evaluated and the highest and low-
low 共⫺1兲 and a high 共⫺1兲 level are assigned independently. For est values chosen. Table 3 shows the high and low values of each
E, F, G, and H the level is determined based on the design gen- group.
erators. Depending on the level of each coded variable the actual
minimum or maximum value of the dimensionless group is iden-
tified.
The maximum and minimum values of each dimensionless
group were determined by assigning a maximum and minimum Table 3. High and Low Values of the Dimensionless Groups
value for each parameter 共Table 2兲. The minimum values for each Dimensionless High value Low value
individual parameter were assigned by assuming an office room Factors group 共+兲 共⫺兲
for a single occupant while the maximum values were assumed
by considering a room capable of holding up to 30 people. The A
冉冊L
W
3.3 0.4
L m
Maximum
10
Minimum
3
C
冉 冊
WHr
A
280 15
Hr
W
m
m
4
7
3
3
D
冉冊
5 1
I
A
m
m2
2.5
0.8
⫺2.5
0.1
E +BCD
冉 冊
Q f
L
5033 50
冉 冊
s 18,000 360 F +ACD 3.3 0.8
L
Q m3 / s 0.24 0.008
Hr
冉冊
s 3,600 360
G +ABC L 1 ⫻ 107 6 ⫻ 105
f kg/ m3 1.1614
kg/m s 1.85⫻ 10−5 dp
p
dp
kg/ m3
m
1,000
5 ⫻ 10−6
800
1 ⫻ 10−6
H +ABD
冉冊p
f
861 689
Each of the 16 cases was repeated once to obtain duplicate re- The room was partitioned into two zones with a partition located
sults, i.e., in total 32 cases were simulated. For each case, first the at the center of height 0.15 m. The square inlet and outlet area
dimensions of the room, the inlet air velocity, and physical prop- was 0.01 m2. The velocity of the uniform inflow was 0.235 m/s.
erties of air and particles were determined from the levels of each Based on the inlet dimensions the Reynolds’ number was 1,600.
dimensionless group or factor as given in Table 1. The length of The Reynolds’ number in the room could range between laminar
the room was initially assumed to be a certain value and the and turbulent depending on the length and velocity scales. This
corresponding dimensions calculated from the high and low val- scenario was adopted and used with the current numerical model.
ues shown in Table 3. The ACH was then calculated and the For the numerical results the velocity data was averaged over the
length adjusted until the ACH was within the range of 1–10. Next last 60 s. The vertical mean velocity along the midpartition height
the air flow is developed for one residence time. After one resi- obtained from the numerical simulations and experimental data
dence time the airflow achieves a pseudo steady state condition. has been compared in Fig. 3. The plot shows that the numerical
Fig. 2 shows the velocity profile in the x direction over time for a and measured results are in very good agreement.
sample case 共Case 10兲 with a residence time of 405 s. The Lagrangian particle model was verified against experi-
After developing the airflow, 100,000 particles were released mental data obtained by Lu et al. 共1996兲. The experiment was
and tracked for the required amount of time based on the level of conducted in a two zone room with a connecting door. The di-
D, 共 / 兲. Particles were initialized over an area of 0.01 m2. The mensions were 5 m共x兲 ⫻ 2.4 m共y兲 ⫻ 3 m共z兲. The inlet and outlet
duplicate case was simulated by randomizing the positions of the dimensions were 0.15共x兲 ⫻ 0.5共y兲 ⫻ 1.0共z兲 m3. The dimensions of
particles over the same area when released from the floor of the the door were 0.95 m共y兲 ⫻ 0.7 m共z兲. To simulate the scenario,
room. Randomizing the particle locations took care of the fact airflow was first established in the room for one residence time
that if particles were released from the same region again they and then particle trajectory calculation was started assuming the
would not occupy the same exact location. For all the simulation particles were uniformly mixed in the first zone. Fig. 4 shows the
cases, nonuniform grids were generated. The smallest mesh size change of the concentration of the particles over time in Zone 1 as
was 0.01 m. The time step used in the calculations was 0.05 s. It particles exit through the connecting door into Zone 2. Numerical
took 2 months to complete the 32 cases, requiring on an average, results agreed very well with the experimental data as shown in
48 h for each case. At the end of each case the number of particles the plot.
remaining in the room and the spatial distribution of the particles
in the room was determined.
Computational Results To determine the main and interacting effects of the factors on
the particles, the results were analyzed via ANOVA and regres-
The data from the 16 duplicate cases from the fractional factorial
sion methods using the statistical toolbox of MATLAB 共Math-
design were analyzed to determine the effect of the factors on the
Works 2007兲.
temporal variation of the number of particles in the room and on
the spatial variability of the particles within the room. Table 4
lists the levels of the 16 duplicate cases, the normalized number Temporal Variation of the Aerosols Released in the
of particles remaining in the room, i.e., 共N / No兲 and the mean and Room
standard deviation 共SD兲, of the number of particles at different
locations of the room at the end of each case. By randomizing the To determine the significant factors, an ANOVA table was gener-
particle positions during release, the results of duplicate cases ated. The residuals were plotted against the predicted value to
showed some variability. assess the adequacy of the model 共Devore 2004兲. The points on
Analyzing the results; for five cases, half or more of the par- the residual plot 关Fig. 7共a兲兴 are evenly distributed on both sides of
ticles remain in the room. In some cases the particles remained the horizontal line at zero. Some of the data points are clustered
clustered at different locations in the room while for others they because most of the data does not vary over a wide range. The
were more evenly dispersed. No particular dimensionless group normalized number of particles in the room was log-transformed
or parameter appeared to be the main reason for the particle be- to aid presentation of the data 关Fig. 7共b兲兴. Table 5 summarizes the
havior. Figs. 5共a and b兲 show the velocity vector plots for two results of the analysis. All the main effects and the two-factor
cases 共3 and 10兲 averaged over 60 s. The plots are representative interactions are listed in the first column of the table. The con-
of the airflow patterns observed in all the 16 duplicate cases. A secutive columns show the sum of square errors, the degrees of
central circulation zone forms at the core of the room and smaller freedom, mean square error and the F-test statistic. The last col-
circulation zones near the corners. The size of the circulation umn shows the p values. This is the probability that the observed
zones and the velocity magnitude at the different locations varied effect is due to random error. Any p value below 0.05 was noted
with the geometry and boundary conditions of each case. The to be significant, i.e., the observed effect is greater than would be
inlet velocities are 0.06 and 0.33 m/s and the dimensions are expected by chance.
6 m共L兲 ⫻ 8 m共H兲 ⫻ 14 m共W兲 and 9 m共L兲 ⫻ 12 m共H兲 Looking at the p values the factors A共L / W兲, D共 / 兲,
⫻ 2.7 m共W兲 for Cases 3 and 10, respectively. The Reynolds’ E共Q f / L兲 and G共L / d p兲 were identified as not significant. How-
number based on the inlet dimensions are 11,052 and 30,822, ever, the p values for the two way interactions indicate that except
respectively. Fig. 5共c兲 shows the root mean square of ūi on the for the two way interactions of 关AD兴 and 关AG兴 all other interac-
vertical axis located at the midpoint of the central plane for Cases tions are significant. This analysis seems to lead to the conclusion
3 and 10. The velocity magnitude for both cases is highest at the that the groups L / d p and / was not significant. These dimen-
inlet region. The turbulent kinetic energy, k 共m2 / s2兲 profiles for sionless groups are further discussed below.
the same cases are shown in Fig. 5共d兲. Near the vicinity of the The magnitude of difference between the scales of the length
inlet the kinetic energy is the highest. Fig. 6共a兲 shows the location of the room in meters and the particle diameter in micrometers is
of the particles for Case 10 after 240 s. The particles have entered 106. As a result the change of L / d p may not significantly influence
the central circulation zone of the room and some are exiting. the number of particles remaining in the room and could be ig-
After 2,025 s, the particle tracking time, nearly 80% of the par- nored from future calculations. The other dimensionless group
ticles had exited from the room. The remaining particles were which was classified as not significant is / . Identifying this
mainly dispersed near the top of the room 关Fig. 6共b兲兴. group as not significant indicates that as more time passes, the
Fig. 6. 共a兲 Particle positions after 240 s for Case 10 showing that the Fig. 7. 共a兲 Plot of standardized residuals against predicted values
particles have moved into the central circulation zone of the room 共dependent variable: 关N / No兴兲 determined via ANOVA; 共b兲 plot of
共front view of the room兲; 共b兲 particle locations at the end of the standardized residuals against predicted values 共dependent
particle tracking time, 2,025 s for Case 10 共front view of the room兲 variable: log关共N + 1兲 / 共No + 1兲兴兲 determined via ANOVA
accordance with the above conclusion. However, a different trend room with some portion of the particles first exiting as they be-
is noticed for Case 11 共Fig. 10兲. The residence time is 875 s but came entrained into the main stream. Because of the presence of
none of the particles exit the room until after 1,000 s. The par- a case which is not in agreement to the conclusion reached from
ticles also exit in two steps with the second step happening after ANOVA, further investigation is required before screening out
3,200 s. Analysis of the results of this case showed that the par- / from future simulations.
ticles had become entrapped in the vortex in the center of the The multiple linear regression model was developed through
forward step-wise regression. Based on the fractional factorial
design developed, there are seven main effects and eight two way
interaction effects. Each interacting effect is also a summation of
the effect of other two way interaction effects. The same conclu-
sions as ANOVA were reached through this method as well, but
the main effect of G 共L / d p兲 was also identified as significant. Eq.
共11兲 is the regression model in terms of the coded variables. The
R2 for the regression model is 0.94. Based on the magnitude of
the coefficients, H 共 p / f 兲 appears to be relatively the most sig-
nificant group among the main effects followed by C共WH / A兲,
B共I / H兲, and F共L / H兲. G共L / d p兲 is the least important. The magni-
tude of the coefficients of the interacting effects is similar to the
coefficient of G. Among the interacting effects the most important
is 关AE+ BF+ DG+ CH兴 and the least is 关AF+ CD+ BE+ GH兴. Fig.
11 is a plot of the standardized residual against the predicted
values from the regression model. The standardized residual is
Fig. 8. Number of particles remaining in the room as a function of obtained by dividing the residual by the estimated SD of the
time 共Case 5兲 residual 共Devore 2004兲. The points do not follow any trend show-
Fig. 9. Number of particles remaining in the room as a function of Fig. 10. Number of particles remaining in the room as a function of
time 共Case 16兲 time 共Case 11兲
ing that the residuals are not dependent on the predicted values.
Two points 关coordinates 共⫺7.8, ⫺3.8兲 and 共⫺7.8, 3.2兲兴, however
appear to be outlying values falling beyond the range of ⫺2 to
+2. The other residuals are within two SDs of their expected
value of zero Fig. 12. 共a兲 Positions of 49,795 particles in the room at the end of
Case 5; 共b兲 histogram showing the frequency of the number of par-
log关共N + 1兲/共No + 1兲兴 = − 4.37 − 1.042关B兴 + 1.349关C兴 + 0.918关F兴 ticles in each domain 共0.1⫻ 0.1⫻ 0.1 m3兲 at the end of Case 5, the
− 0.460关G兴 − 2.274关H兴 + 0.350关AB + EF y-axis has been log-transformed for a clearer presentation
+ CG + DH兴 + 0.220关AC + DF + BG
+ EH兴 + 0.366关AE + BF + DG + CH兴 Forward step-wise regression was again applied to determine
the coefficients of the linear regression model and the same con-
− 0.133关AF + CD + BE + GH兴
clusions were reached regarding the significance of the factors.
− 0.307关AH + BD + CE + FG兴 共11兲 Eq. 共12兲 is the developed regression model
log关 + 1兴 = 2.018 − 0.155关A兴 − 0.206关B兴 + 0.479关C兴 − 0.290关D兴 Validation Studies Applying the Multiple Linear
Regression Models
− 1.662关E兴 − 0.403关G兴 − 1.10关H兴 + 0.088关AB + EF
Thirty two separate and independent CFD simulations were con-
+ CG + DH兴 + 0.217关AC + DF + BG + EH兴 ducted, with the magnitude of the input variables kept within the
+ 0.035关AD + CF + EG + BH兴 + 0.084关AE + BF range of the maximum and minimum levels of each dimension-
less group, to obtain a data set for validation of the linear models.
+ DG + CH兴 + 0.090关AF + CD + BE + GH兴 The data obtained from the simulations were compared to the
+ 0.019关AG + BC + DE + FH兴 − 0.187关AH + BD results predicted from the developed models. First the regression
model, Eq. 共11兲 was applied to calculate the lognormalized num-
+ CE + FG兴 共12兲 ber of particles remaining in the room for the thirty two indepen-
dent cases. The average and SD of 共simulated− predicted兲 results
The R2 for the regression model is 0.99. In Eq. 共12兲 based on the was 0.927 and 4.00 while the standard error, from Eq. 共13兲 共De-
coefficient; E共Q / L兲 is the most significant followed by vore 2004兲, was 4.99. The standard error was 1.23 when the
H共 p / f 兲, C共WH / A兲, G共L / d p兲, D共 / 兲, B共I / H兲, and then model was applied to predict the data from which the model was
A共L / W兲. Among the interacting effects 关AC+ DF+ BG+ EH兴 is developed, i.e., the fractional factorial data set
relatively significant. Fig. 15 is the residual plot for the regression
Standard error
model. The points also do not follow any trend and appear to be
distributed evenly about the horizontal line with most points
冑
being within two SDs of the expected value, zero. One significant
outlier was located at coordinates 共3, ⫺3.8兲.
=
兺 共simulated results − predicted results兲2
number of runs − 共number of parameters + 1兲
共13兲
Fig. 14. Plot of standardized residuals against predicted values 共de- Fig. 15. Plot of standardized residuals against predicted values 共de-
pendent variable: log关 + 1兴兲 determined via ANOVA pendent variable: log关 + 1兴兲 from the regression model, Eq. 共12兲
Simulated log[σ+1]
6 The Buckingham pi theorem applied here helped to identify eight
5 dimensionless groups which are important in describing the dis-
persion of aerosols in a ventilated space. The application of clas-
4
sical experimental design methods allowed the development of an
3
efficient design plan to explore the relationship and influence of
2 these groups. Fractional factorial design approach was used to
1 identify the parameter values for detailed numerical simulations.
0 A comprehensive and validated CFD model was applied to con-
0 1 2 3 4 5 6 7 8
(a) Predicted log[σ+1] duct the simulations as planned out.
From the simulations two dimensionless groups 共the ratio of
8 the length of the room to the particle diameter and the ratio of
7 particle tracking time to residence time兲 were identified that did
not significantly influence the temporal change of the particle
Simulated log[σ+1]
6
5
number in the room. However they influenced the spatial distri-
bution of the particles within the room at the end of the simula-
4
tion. The dimensionless groups cannot be treated independently
3 as well since all interactions proved to be significant. Hence, all
2 the identified dimensionless groups will be considered for future
1 analysis. The use of two-level design which is based on assigning
0
maximum and minimum values to each factor implies the as-
0 1 2 3 4 5 6 7 8 sumption of linearity in the factor effects.
(b) Predicted log[σ+1] Forward step-wise regression method was applied to develop
the linear regression models for predicting the normalized number
Fig. 16. 共a兲 Plot of predicted log关 + 1兴 关applying Eq. 共12兲兴 against of particles remaining in a room and for determining their spatial
the simulation results from the 32 CFD cases generated through fac- dispersion. Thirty two more independent scenarios were simu-
torial design used to develop the linear model; 共b兲 plot of predicted lated by choosing points within the design space and the simu-
log关 + 1兴 关applying Eq. 共12兲兴 against the simulation results from the lated data compared with the predicted data from the developed
32 CFD cases not used to develop the linear model 共validation data linear regression models. The results showed that the linear model
set兲 predicted the data used in the fractional factorial design very well
but was not as accurate at predicting an independent data set.
A step by step method has been described in this work using
CFD toward the development of a robust model for predicting the
Based on the standard error it is evident that the model behaves number of aerosols and their distribution pattern in a room. The
much better in predicting the data set it was developed from but method outlined here shows how to apply the detailed results of a
not as well for another data set. Similarly Eq. 共12兲 was also ap- time consuming but accurate CFD model in a systematic manner
plied to predict log关兴 for the same 32 independent simulations. for the development of more user friendly solutions. Dimensional
The average and SD of 共simulated− predicted兲 was ⫺1.94 and analysis provided a convenient way of reducing a multidimen-
1.504, respectively. The standard error for the independent data sional problem into controllable proportions. The work can be
set was 3.35 while for the fractional factorial design data set it expanded by considering more variables or incorporating more
was 0.127. In this case as well, the linear regression model, Eq. complexities or phenomena into the model.
共12兲, is predicting the data from which it was developed well but
not doing as well for a separate and independent data set.
Figs. 16共a and b兲 are plots of the predicted results applying Eq. Acknowledgments
共12兲 against the simulated results from the fractional factorial
design data set and the separate validation data set. This provides The work described in this paper was supported by the L.D. Betz
a visual assessment of model effectiveness in making predictions. endowment.
If the model behaved well most of the points would be close to
the 45° line. Fig. 16共a兲 shows that the linear model predicts the
data set from which it was developed very well. In Fig. 16共b兲 Notation
there is more scatter about the 45° line.
A multizone airflow simulation program, CONTAM 共Walton The following symbols are used in this paper:
and Dols 2005兲 was used to simulate nine cases. The CONTAM A 共m2兲 ⫽ area of the inlet and outlet;
predictions were compared to the CFD results and to the predic- Cc ⫽ Cunningham correction factor;
tions from the regression model, Eq. 共11兲. Information about the CD ⫽ drag coefficient;
specific location of the particles cannot be obtained from CON- d p 共m兲 ⫽ particle diameter;
TAM, hence comparison with Eq. 共12兲 was not possible. The sum FD 共N兲 ⫽ drag force;
of square of errors 共SSEs兲 between simulated results from CFD FG 共N兲 ⫽ gravitational force;
and CONTAM predictions was 298. While the SSE between the Hr 共m兲 ⫽ height of the room;
simulated results from CFD and predictions from Eq. 共11兲 was I 共m兲 ⫽ vertical distance between the inlet and outlet;
196 indicating that the linear model showed marked improve- L 共m兲 ⫽ length of the room;
ment. m 共kg兲 ⫽ particle mass;