You are on page 1of 10

Principal Component Analysis on a LES

of a Squared Ribbed Channel

Soledad Le Clainche Martı́nez1 , Carlo Benocci2 , and Alessandro Parente3

School of Aeronautics, Universidad Politécnica de Madrid, Spain
Department of Environmental and Applied Fluid Dynamics,
von Karman Institute for Fluid Dynamics, Belgium
Service d’Aéro-Thermo-Mécanique, Université Libre de Bruxelles, Belgium

Abstract. The present paper reports on the application of Principal

Component Analysis (PCA) on the flow and thermal fields generated by
the large-eddy simulation (LES) of a square ribbed duct heated by a con-
stant heat flux applied over the bottom surface of the duct. PCA allows
to understand the complexity of the resulting turbulent heat transfer
process, identifying the flow and thermal quantities which are most rele-
vant to the process. Different algorithms have been employed to perform
this analysis, showing high correlation between turbulent coherent struc-
tures, identified by Q − criterion, and the heat transfer quantified by the
non-dimensional magnitude Enhancement Factor (EF ), both identified
as Principal Variables (PV) of the process.

Keywords: Conjugate heat transfer, Large Eddy Simulation, Principal

Component Analysis, Flow topology.

1 Introduction
Most physical and engineering processes depend on the interaction of a high
number of different physical parameters. A such example is represented by the in-
ternal cooling of turbines blades. The turbine efficiency of a jet engine can be im-
proved increasing the turbine inlet temperature. However, temperatures higher
than the melting point temperature of the blade material could be reached, as
they would cause catastrophic failure. Thus, an efficient blade cooling must be
introduced to avoid overheating and internal ducts, transporting a cooling fluid,
are typically employed within blades. The efficiency of the cooling process can
be improved by means of turbulence generators, which increase the turbulent
mixing and turbulent heat transfer. A typical generator is a set of ribs installed
on the walls of the cooling ducts. The obvious drawback associated with the use
of ribs is the increase of pressure drop; therefore, the advantages and shortcom-
ings of this approach must be accurately assessed. A deeper discussion about
the topic can be found in [1], which provides an overview of heat transfer in gas

Á. Herrero et al. (eds.), International Joint Conference SOCO’13-CISIS’13-ICEUTE’13, 259

Advances in Intelligent Systems and Computing 239,
DOI: 10.1007/978-3-319-01854-6_27,  c Springer International Publishing Switzerland 2014
260 S. Le Clainche Martı́nez, C. Benocci, and A. Parente

The number of parameters that can potentially affect the heat transfer pro-
cess is very high and it is therefore very complex to identify the most relevant
ones. In the present work, Principal Component Analysis (PCA) is applied to
the matrix containing all the parameters potentially influencing the process, to
identify the Principal Variables (PV) and quantify their effect on the system
quantify of interest, i.e. the heat removal rate. PCA is a well-known statistical
technique that transforms a set of correlated variables into a smaller number
of uncorrelated variables, called Principal Components (PCs) [3,9]. The PC are
determined to successively maximize the amount of variance contained in the
original data sample. The approach performs a rotation within the original data
space to reveal the internal structure of the sample and to maximize the amount
of information accounted for by a smaller number of parameters. The PCs are,
by definition, linear combinations of all variables available. However, in some
applications, it can be useful working in terms of the original variables, to help
the physical interpretation of the results. Therefore, the PCA analysis of the
thermo-fluid dynamic state is completed with the determination of the Principal
Variables [3].
The present paper presents the application of PCA to the numerical simula-
tions carried out in [8], with the main objective of identifying the parameters
which control the heat transfer process in proximity of the rib and clarifying
the effect of the rib on heat transfer enhancement. Additionally, the study aims
to investigate the relevance of turbulent coherent structure on the heat trans-
fer process, as recent studies [6] pointed out how heat transfer enhancement
always occurs in regions of the flow characterized by the existence of presence
of organized turbulent flow structures.

2 Physical Principles and Numerical Simulations

The investigated case is presented in Fig. 1. It represents conjugate heat transfer
in a squared cross-section ribbed duct with five consecutive ribs [8]. The ribs
are characterized by a squared cross section positioned perpendicularly to the
channel axis. The rib hydraulic diameter is D = 0.075 m, the rib height is
h/D = 0.3, the pitch length, p/h = 10 and the thickness of the coupled solid
domain is s/h = 1.1. The Reynolds number based on the hydraulic diameter
and bulk velocity Ub is 4 · 104 and the Prandtl number is equal to 0.7. The
simulations were solved using the Large Eddy Simulations (LES) module of the
commercial software FLUENT 6.3. Numerical results have been validated with
some experimental measurements carried out at von Karman Institute employing
Particle Image Velocimetry (PIV) and Infrared thermography (ICR). In all cases,
a satisfactory agreement between simulations and experimental data was found.
Fig. 2 presents an overview of the high complexity fluid dynamic problem,
showing how the flow is dominated by regions of separated flow. Fig. 2a shows
the the vortex structures around the 4th rib, on the symmetry plane (up) and the
thermal field in the fluid and solid part of the channel on the same plane (down).
Principal Component Analysis on a LES of a Squared Ribbed Channel 261

Temperature decreases in presence of vortex structures. Fig. 2b shows the heat

transfer flux (up), defined as the non-dimensional magnitude Enhancement Fac-
tor (EF) and coherent structures (down) defined with the Q − criterion. The
presence of coherent structures enhances the heat transfer flux on the bottom

Fig. 1. Computational domain for the ribbed channel geometry

(a) (b)

Fig. 2. Left: (up) Vortex structures and (down) thermal field on the mid plane around
the 4thrib. Right: (up) Enhancenment Factor and (down) Coherent Structures iso-
surfaces (Q/(Ub2 /h2 ) = 24) on the 4th rib. Legend: blue colors: negative value, green
colors: zero value, red colors: positive value.

3 Theory
3.1 Coherent Structures
Coherent structures (CS) are large and strongly organized turbulent structures
associated with the process of production of turbulent kinetic energy. Within
the present context, the role of CS with respect to the heat transfer mechanism
is of interest. A common criterion for the detection of coherent structures in a
turbulent flow is the so-called Q-criterion [2], Q = 12 (Ωij Ωij − Sij Sij ), where
Sij is the strain tensor and Ωij is the rotation tensor. When rotational effects
are dominant over strain effects (Q > 0), a coherent structure identified by Q
exists. Fig. 2b down shows the iso-surface of Q/(Ub2 /h2 ) = 24, where Ub is the
bulk velocity and h is the height of the rib.
262 S. Le Clainche Martı́nez, C. Benocci, and A. Parente

3.2 Enhancement Factor

The effect of the rib on heat transfer is assessed by means of the so-called
Enhancement Factor (EF ), which represents the ratio of the Nusselt num-
ber along the actual (ribbed) duct, N u, and that on a smooth surface, N u0 ,
EF = N u/N u0 = h · Dh /kair · N u0 , where h is the heat transfer coefficient, Dh
is the hydraulic diameter, kair is the air thermal conductivity, Re and P r are
the Reynolds and Prandtl number, respectively. Fig. 2b, shows in the upper part
the EF distribution on the bottom solid surface of the body.

3.3 Principal Component Analysis

PCA is a well-known statistical technique that transforms a set of possibly corre-

lated variables into a smaller number of uncorrelated variables, called Principal
Components (PCs) [3,9]. For a multivariate data set, PCA is usually employed
for detecting the directions that carry most of the data variability, thus pro-
viding an optimal low-dimensional projection of the system. For a data set, X,
consisting of n observations of p variables, the sample covariance matrix, S, can
be defined as S = 1/(n − 1)XT X. Recalling the eigenvector decomposition of
a symmetric, non singular matrix, S can be decomposed as S = ALAT , where
A is the (p x p) matrix whose columns are the eigenvectors of S, and L is
a (p x p) diagonal matrix containing the eigenvalues of S in descending order,
l1 > l2 > . . . > lp . Once the decomposition of the covariance matrix is performed,
the Principal Components (PC), Z, are defined by the projection of the original
data onto the eigenvectors, A, of S, Z = XA. Then, the original variables can
be stated as a function of the PC as X = ZAT . The main objective of PCA is to
replace the p elements of X with a much smaller number, q, of PC, preserving
at the same time the amount of information originally contained in the data. If
a subset of size q << p is used, the truncated subset of PC is Zq = XAq . This
relation can be inverted to obtain an approximation of the original state space,
X q = Zq ATq . A representation of the PCA reduction process is shown in Fig. 3.

Fig. 3. PCA reduction process

3.4 Principal Component Rotation and Principal Variables

The physical interpretation of Principal Components is generally not straight-

forward as they are in principle linear combination of all the original variables.
Principal Component Analysis on a LES of a Squared Ribbed Channel 263

Rotation represents an attempt to overcome such difficulty: with rotation, the

weights can be redefined to attain a simple structure for Aq , so that weights on
a PC are either close to unity or close to zero and, thus, variables have large
weights on only few or (ideally) one PC. The most common orthogonal rotation
is based on the maximization of the VARIMAX criterion [4]. When rotation
is applied, the total variance within the rotated q-dimensional subspace is re-
distributed amongst the rotated components more evenly than before rotation;
however it remains the maximum that can be achieved.
Principal variable (PV) represents a further attempt to help the physical
understanding of Principal Components. Differently from PC rotation, PV al-
gorithms try to link the PC back to a subset of the original variables, which
satisfy one or more optimal properties of PCA. A number of methods exist for
selecting a subset of m original variables. Within the present investigation, the
M2 method by Krzanowski [5] and the approaches proposed by McCabe [7] were
selected. In the M2 method, PCA is performed on the original data matrix and
the scores Zq are then then evaluated. Assuming that q is the true data dimen-
sionality, the approximation, Z q , of Zq obtained by keeping only q variables of
the original data set is evaluated, to find the subset of variables yealding the
least difference between Zq and Z  q . The approaches by McCabe [7], MC1-MC3,
originates from the observation that the principal components satisfy a certain
number of optimality criteria, i.e. maximal variance and minimum reconstruc-
tion error. Therefore, a subset of the original variables optimizing one of these
criteria defines a set of principal variables.

4 Results

PCA analysis has been performed in order to determine the most relevant quan-
tities for the heat transfer process. The variables selected for the analysis are the
mean stream-wise (U ), chord-wise (V ) and span-wise (W ) velocities and their
variances, the mean temperature (T ) and its variance, the turbulent kinetic en-
ergy (k), the mean vorticity (Ω) and mean strain (S) tensors, the enhancement
factor (EF ), the mean heat flux (q) and the mean velocity and temperature
correlations U V , U W , V W , U T , V T , W T . The matrix to be processed with
PCA is built collecting data on different planes in the computational domain.
In particular, the heat transfer related quantities (heat flux, EF , T ) are taken
at the first grid point above the ribbed wall and around the rib itself. On the
other hand, the flow-related quantities (velocities, strain, vorticity, Q) are more
meaningful at a certain distance from the wall. Fig. 4a clearly indicates that the
coherent structures are present at a certain distance from the wall, they interact
with the bottom surface, leading to a reduction of temperature, and then they
mix back with the bulk flow. Fig. 4b shows the different sampling planes em-
ployed for the PCA analysis. Since the computational grid is unstructured, the
data extracted for the PCA analysis were interpolated over a structured grid.
Two different data set have been analyzed. The resulting data set for PCA
and PV analysis, indicated as CHT1, consists of 20 state variables and about
264 S. Le Clainche Martı́nez, C. Benocci, and A. Parente

(a) (b)

Fig. 4. Iso-surfaces at instantaneous Q/(Ub2 /h2 ) = 24 colored by temperature (a) and

sampling planes for the application of PCA and PV (b)

6000 observations. The second data set, indicated as CHT2, is then built to focus
the statistical analysis on the observations corresponding to positive values of
Q, to remove the effect of strain dominated regions on the results and point
out the role of coherent structures on the heat transfer process. The number of
observations for the reduced CHT2 data set is about 2000.

4.1 PCA Analysis

Fig. 5 shows the magnitude of the eigenvalues associated with the PCA reduction
(left), and the contribution of the largest eigenvalues to the amount of variance
explained by the new basis vectors for the CHT1 (a) and CHT2 (b) data sets.
The first five eingenvalues (representing 25% of the total number of variables)
account for more than 75% and 80% of the total variance in the original CHT1
and CHT2 data sets, respectively. On the other hand, the last six smallest eigen-
values contain no useful information and only explain linear dependencies among
the original variables. Therefore, a non-negligible size reduction can be accom-
plished with PCA, through the identification of the most active directions in the
data. Table 1 shows a selection of the correlation matrix, built for the extrac-
tion of the PC and PV, showing the correlations between k, EF and Q and a
subset of variables, for the CHT1 and CHT2 data sets. For CHT1, it is observed
that Q is mostly correlated to T  , V , and the correlations U V and V T (beside
the obvious correlation with strain and vorticity). This is in agreement with the
qualitative observation (Fig. 4a) that the vortex structures present in the flow
are driven by the temperature gradient determined by the heat exchange pro-
cess. Their movement is mostly perpendicular to the rib plane (thus explaining
the V component of velocity), allowing the structures to reach the ribbed surface
and to mix back into the flow bulk. On the other hand, EF is negatively corre-
lated with T (EF is larger when the temperature decreases) and it shows large
positive correlations with T  and V T , confirming the interaction between coher-
ent structures and the heat transfer process. As far as turbulent kinetic energy
is concerned, k shows a positive correlation with the variables representing the
heat transfer process, indicating that an increase in turbulence levels could be
effectively used to reduce rib temperatures (if compatible with pressure drops).
The same analysis for the CHT2 data set confirms the correlations pointed out
for CHT1 and supports the phenomenological explanation provided above, with
Principal Component Analysis on a LES of a Squared Ribbed Channel 265

(a) (b)

Fig. 5. Scree-graph showing the main contribution to the original data variance for the
CHT1 (a) and CHT2 (b) data sets

Table 1. Selection of the correlation matrix showing the correlation between k, EF

and Q and the variables defining the CHT1 and CHT2 data sets

k EF U V W T U V  W  UV
k 1.00 0.26 -0.33 -0.26 -0.02 -0.28 0.93 0.73 0.80 -0.54
CHT1 EF 0.26 1.00 -0.20 0.08 0.02 -0.74 0.34 -0.18 0.35 0.17
Q 0.07 0.21 0.13 0.58 -0.01 -0.19 0.03 -0.14 0.26 0.63
k 1 -0.08 0.15 -0.39 0.03 0.21 0.87 0.59 0.64 -0.48
CHT2 EF -0.08 1 -0.09 0.13 0.07 -0.78 0.06 -0.59 0.29 0.47
Q -0.14 0.35 0.32 0.80 0.00 -0.30 -0.18 -0.35 0.27 0.83
UW V W T  UT V T W T Ω S Q q
k -0.10 0.06 0.19 0.20 0.27 -0.08 0.06 0.03 0.03 0.25
CHT1 EF -0.05 0.08 0.72 0.35 0.73 -0.05 0.19 0.20 0.21 0.91
Q 0.01 -0.10 0.42 -0.27 0.45 -0.02 0.84 0.74 1.00 0.17
k 0.00 0.09 -0.19 0.02 -0.18 0.00 -0.23 -0.25 -0.14 -0.05
CHT2 EF -0.03 0.18 0.76 0.18 0.80 -0.01 0.32 0.30 0.35 0.91
Q 0.05 -0.04 0.62 -0.50 0.69 0.00 0.94 0.88 1 0.31

an exception concerning turbulent kinetic energy. In fact, the values listed in

Table 1 for CHT2 indicate almost zero correlations between k and the heat
transfer related variables, EF and q. Such result can be explained considering
that the CHT2 data set is built by assembling all those observations correspond-
ing to positive values of Q. Therefore, the values listed in Table 1 only indicate
that in the range of k compatible with the existence of CS, the heat flux span
an almost uniform range of values, leading to correlations close to zero.
Fig. 6 shows histograms of the structure of the first four rotated eigenvectors
of S, for the CHT1 and CHT2 data sets, respectively. The analysis of Fig. 6
clearly shows that the main eigenvector structure of the covariance matrix does
not significantly change switching from the CHT1 to the CHT2 data sets. It can
be observed that the first component is characterized by very large weights on V
and U V , and on Ω, S and Q, for both data sets. Such component can be then in-
terpreted as an indicator of organized coherent structures; moreover, it confirms
the relation between the latter and the vertical flow motion from the bulk to the
266 S. Le Clainche Martı́nez, C. Benocci, and A. Parente

wall and vice versa (V , U V ). Only a few differences can be emphasized when
comparing the two data sets. The CHT2 data set shows, in fact, non-negligible
weights for the correlations V W and W T , suggesting a non-negligible influence
of three-dimensional effects on the heat transfer process. The second rotated PC
is dominated by heat transfer related variables, EF and q (showing a correlation
of about 91%, Table 1), T and T , and the correlation V T , confirming the role
of the vertical velocity component on the heat transfer mechanism. Likewise the
first component, a major relevance of correlations related to three-dimensional
effects is observed when analyzing the CHT2 data set. As far as the third and

Fig. 6. Scree graph showing the main contribution to the original data variance

fourth PCs are concerned, these can be regarded as representative of the cor-
relation (U T ) between mean flow (U ) and the heat transfer from the wall to
the bulk (T ), and of turbulent kinetic energy (k). It is interesting to have a
closer look at the third component and particularly at the relative weights of
the original variables. In fact, the sign of the coefficients within a PC specifies its
orientation in the original multidimensional space, thus clarifying the relation
between the original variables of the sample. Given this remark, it becomes clear
how the variance on this component is directed towards concordant (positive or
negative) values EF , U T , but discordant values of EF and Q, or EF and T .
This is in agreement with the observation that the coherent structures in the
bulk flow (large values of Q) dissipate at the wall (small values of Q), leading to
an increase of EF and heat exchange (U T ) and a local decrease of temperature.

4.2 Principal Component Rotation and Principal variables

Fig. 7 shows the distribution of the first four unrotated PCs over the fourth rib
(a-d) and the corresponding distribution of the state variables showing highest
Principal Component Analysis on a LES of a Squared Ribbed Channel 267

Fig. 7. Distribution of the principal component modes (a-d) and of the variables show-
ing the largest weights on the rotated components (Fig. 6) over the fourth rib. From
up to down: vorticity, EF , U T , k.

contribution with the PCs (e-h). It can be clearly observed that the selected
variables very well mimic the PC behaviour, indicating that VARIMAX rotation
is an effective method to disclose the physical meaning of the PC.
The analysis presented in the previous sections is completed with PV analysis,
carried out on the CHT2 data sets, applying PV methods M2 and MC2-MC3.
All the employed methods indicate the following variables as PVs: k, T , W ,
Q and EF . Therefore, PV analysis confirms the results provided by PCA and,
in particular, by the rotation of the PCs using the VARIMAX criterion, being
the selected variables coincident or highly correlated to the ones dominating the
rotated principal components.

5 Conclusions

PCA and PV analysis have been applied to the flow and thermal fields of a ribbed
square duct, heated through the ribbed wall. The analysis has shown that most
of the information contained in the original thermo-fluid-dynamic field, defined
by twenty state variables, can be represented by the first five eigenvectors of the
sample covariance matrix, defining the problem principal components. The ap-
plication of the VARIMAX rotation criterion to the set of principal components
allowed isolating the different groups of physical parameters controlling the over-
all process and to identify the most relevant variables within the initial set of
twenty state variables. Importantly, the role of the rotation-dominated regions
268 S. Le Clainche Martı́nez, C. Benocci, and A. Parente

(coherent structures) in the heat transfer mechanism from the heated wall to
the fluid bulk has been highlighted. The PCA analysis has been completed with
the application of PV algorithms for the extraction of the most relevant physical
quantities driving the process. The outcome of PV analysis fully supported PCA
findings, demonstrating the suitability of PCA and PV for the investigation of
complex physical processes, characterized by a very large number of potentially
relevant physical parameters.

Acknowledgements. The authors acknowledge the financial support of the

Air Force Office of Scientific Research (AFOSR), Contract Number: FA8655-
08-1-3048, supervised by Dr. S. Surampudi and Dr. G. Abate of the European
Office of Aerospace Research and Development.

1. Han, J.C., Dutta, S., Ekkad, S.V.: Gas turbine heat transfer and cooling technology.
Taylor & Frances (2000)
2. Hunt, J.C.R., Wray, A.A., Moin, P.: Eddies, stream, and convergence zones in
turbulent flows. Center for Turbulence Research Report, p. 193 (1988)
3. Jolliffe, I.T.: Principal component analysis. Springer, New York (1986)
4. Kaiser, H.F.: The varimax criterion for analytic rotation in factor analysis. Psy-
chometrika 23, 187–200 (1958)
5. Krzanowski, W.: Selection of variables to preserve multivariate structure, using prin-
cipal components. Applied Statistics 36, 22–33 (1987)
6. Lohász, M.M., Rambaud, P., Benocci, C.: Flow features in a fully developed ribbed
duct flow as a result of les. Journal of Flow, Turbulence and Combustion 77, 59–76
7. McCabe, G.P.: Principal variables. Technometrics 26, 137–144 (1984)
8. Nakhle, D., Rambaud, P., Benocci, C., Arts, T.: Numerical investigation of flow
and heat transfer in ribbed square duct applying les. In: 10th International Sympo-
sium on Experimental and Computational Aerothermodynamics of Internal Flows
- Proceeding of ISAIF10, Bruxelles, Belgium, July 4-7 (2011)
9. Parente, A.: Experimental and numerical investigation of advanced systems for
hydrogen-based fuel combustion. PhD Thesis, University of Pisa (2008)