SENSITIVITY ANALYSIS OF HEAT EXCHANGER DESIGN TO

UNCERTAINTIES OF CORRELATIONS

Julien Lambert, Louis Gosselin*

Department of Mechanical Engineering, Université Laval, Québec City, Québec, Canada

Abstract
Heat exchanger design procedures rely on a series of correlations to estimate heat transfer
coefficients, pressure drops, costs, etc. Each correlation is characterized by uncertainties,
i.e. it returns an approximate value of the variable that it aims at predicting. This paper
studies how sensitive the total cost evaluation is to these uncertainties for shell-and-tube
heat exchangers. A variance decomposition approach is used to calculate the total effect of
each uncertain variable. Three test cases are studied for which the most influential
uncertainties were found to be those on the correlations for evaluating the purchase cost
and the shell side heat transfer coefficient. Probability distributions of the total cost are
presented and exhibit a large variance of the total cost. The impact of the cost of energy
was also investigated, which revealed that uncertainty on energy cost had a smaller total
effect than that of other variables.

Keywords: heat exchanger design; sensitivity analysis; correlations; cost estimation; shell-
and-tube

Nomenclature
A surface area, m2
Cp heat capacity J kg-1 K
Ct Number of tube correction factor related to the tube layout pattern
Dctl diameter of a circle that pass through the center of the outermost tubes, m
di tube inside diameter, m
E(…) expected value

*
Corresponding Author: Louis.Gosselin@gmc.ulaval.ca; Tel.:+1-418-656-7829; Fax: +1-418-659-5343.

1
Er error on correlation
ec cost of energy, $/kWh
f friction factor
h heat transfer coefficient, W m–2 K–1
IC initial cost, $
j Colburn factor
Jb bundle bypass correction factor for the shell-side heat transfer coefficient
Jc baffle cut correction factor for the shell-side heat transfer coefficient
Jl baffle leakage correction factor for the shell-side heat transfer coefficient
Jr laminar flow correction factor for the shell-side heat transfer coefficient
Js unequal baffle spacing correction factor for the shell-side heat transfer coefficient
k thermal conductivity, W m-1 K
L tube length, m
ṁ mass flow rate, kg/s
Nb number of baffles
Nt number of tubes
OC operating cost, $
op number of operating hours, h
Pr Prandtl number
pt tube pitch, m
Re Reynolds number
s number of tube pass
S sensitivity index
v fluid velocity, m/s
V variance
w pumping energy, J

X average value of a random variable

Greek symbols
ΔP Pressure drop, Pa
δM correction factor related to materials for the purchase cost

2
δP correction factor related to operating pressures for the purchase cost
δT correction factor related to operating temperatures for the purchase cost
 relative error (%)
ζb bypass flow correction factor for the shell-side pressure drop
ζl baffle leakage correction factor for the shell-side pressure drop
ζs inlet and outlet baffle spacing correction factor for the shell-side pressure drop
μ viscosity, Pa s
ρ density, kg/m3
ϕ discrete interest rate factor
 standard variation
ѱ number of tube correction factor

Subscripts
b baffle
cc counter-flow
cw cross-flow
s shell-side
t tube-side
tw tube wall
w wall

3
1. Introduction
Heat exchangers are used in a wide variety of applications. They can come in various sizes
and designs. The problem of designing the best heat exchanger for a given task has been
addressed extensively in literature [1].

Over the years, different optimization techniques were employed to identify best
heat exchanger designs, which are most of the times defined as the cheapest, i.e. the ones
having the lowest cost for a certain application [2]–[5]. Multi-objective optimization
methods have also been used to design heat exchangers. For example, Wang et al. and Du
et al. used a multi-objective optimization strategy based on genetic algorithms [6], [7]. In
Ref. [8], a non-dominated sorting genetic algorithm optimization was used. Multi-objective
free search techniques [9], hybrid chaotic quantum behaved particle swarm optimization
[10] and firefly algorithm [11], among others, were also employed to identify best heat
exchanger designs.

These optimization and design methods rely on empirical models or correlations to

provide estimations of different parameters of interest. Correlations for estimating pressure
drop and heat transfer characteristics in heat exchangers are usually developed from
experimental measurements or numerical simulations, and correlations for purchase cost
estimation are based on economic analysis of products available on the market. Inherently,
correlations include some errors, i.e. they do not return the exact value but only an
approximation. For example, it is known that several heat transfer correlations have quite
a large error, in such a way that the convection coefficient obtained from them can be over
or under estimated by as a much as 25% [12]. The precision of correlations is dictated by
that of the experimental measurements on which they is based and by how close the data
points are to the regressions proposed by the correlations. For the design optimization of
heat exchangers, a few authors have thus proposed more precise ways to estimate the
required parameters, at the cost of heavier computations. For example, in Ref. [13], Kriging
response surface were used to overcome the dependence on empirical correlations.
Similarly, the correlations that were used for the optimization in [6] were built from CFD
simulations but this time with a Support Vector Machine (SVM) approach. More precise
cost estimation techniques for heat exchangers were developed in [14].

4
 Disregarding uncertainties in heat exchanger calculations would mean that only the
most probable value of each uncertain variable (e.g., heat transfer coefficient, pressure
drop, etc.) would be retained, not its probability distribution of possible values. With such
an approach, one only obtains the “expected value” of the calculated heat exchanger
performance. This method is by far the most widely used in literature on heat exchanger
performance assessment and design procedures, e.g., [1]- [11].

Since the actual real value of several parameters on which the performance
calculations rely might be different from the estimated ones (remembering that several
variables possess a certain uncertainty, i.e. they are only known with a certain level of
precision), the calculated performance is likely to be different from the actual one. Given
how widely used correlations are in literature and in practice, there is a need to develop a
better understanding of how the uncertainties of correlations propagate in the calculations
and affect the estimated performance of a heat exchanger. More specifically when
assessing the performance of a heat exchanger design, integrating the uncertainties
associated to the different variables involved in the calculations would allow to obtain a
probability distribution of the heat exchanger performance, not just its most probable value.
Studies on different systems have shown that taking into account uncertainties can change
the design process and the decision making [15]. As one obtains a range of possible
performance with a probability distribution, risk assessment naturally becomes an
obligation. For example, based on calculations involving uncertainties, a heat exchanger
design might have a low expected cost with a high level uncertainty (e.g., 10,000±5,000$),
and another design, a higher expected cost but with a lower level of certainty (e.g.,
12,000±1,000$). It is thus important to determine how the final performance is affected by
the different uncertainties (e.g., what is the resulting ±X$ that can be tagged to the cost
estimation in the example above) and to identify the most influential variables on the
precision of the heat exchanger performance. This information can also serve to identify
where the focus should be put in future experimental or numerical work to improve
correlations (i.e. increase their precision).

This paper aims a developing a better understanding and assessment of error propagation
in the evaluation of heat exchanger performance and proposes a methodology to do it. The

5
heat exchanger model is first presented in Section 2. The focus is on shell-and-tube heat
exchangers, but the ideas developed in this work could be exploited for other types of heat
exchanger as well. Section 3 recapitulates the different sources of uncertainties in the
correlations and provides numerical values for these errors. The sensitivity analysis
technique that we used and the three test cases that we studied are described in Sections 4
and 5, respectively. Finally, the main results are presented in Section 6, and Section 7
focusses on the impact of the cost of energy. This paper can be seen as a first step towards
a more rigorous heat exchanger design optimization that would take into account
uncertainties.

2. Heat exchanger model

2.1 Cost estimation
The initial cost of a heat exchanger is given by Eq.(1) and depends mainly on the total heat
transfer area A [16]:
0.68
 A
IC = 3.28  10  
4
 M  P T i ,n (1)
 80 
The purchase cost also depends on the material of construction and the operating pressure
and temperature which are expressed in Eq. (1) as correction factors δ with the subscript
P, M and T respectively. Moreover, in order to express the initial cost in terms of annuities,
it must be multiplied by an annualization factor ϕi,n considering an interest rate i and a
number of years n. Correlations such as Eq. (1), which are widely used in literature, are
typically found by comparing the real purchase cost of a series of pieces of equipment
available on the market.

The operational cost OC is mostly due to the pumping power required to circulate
the fluids in the shell-side and the tube-side of the heat exchanger:

OC =
( ws + wt )  op  ec
(2)
1000
In Eq. (2), the energy required for pumping the shell-side fluid and, and that for the tube-
side fluid are respectively ws and wt. The number of operating hours op was estimated to
5000 hours and the cost of energy ec to 0.10 $/kWh.

6
2.2 Heat transfer
The required heat transfer surface area depends on the shell-side and tube-side heat transfer
coefficients which are usually provided by correlations, which are derived from
experimental measurements. The heat exchanger model considered in this paper is derived
from that in Ref. [3], which could be consulted for more details. The complete model is
not explained here but the important correlations are presented in order to show where the
uncertainties are present in the model. The tube-side heat transfer coefficient depends on
fluid properties, mass flow rate and tubes features [1][17]:
0.14
k  
ht = 0.023 t Prt1/3 Ret0.8  t  (3)
di  tw 
The properties with the subscript tw are evaluated at the tube wall surface temperature. The
shell-side heat transfer coefficient also depends on fluid property and flow rate, but it must
be corrected by a series of factors J in order to take into account the effect of the baffles
and tubes configuration due to the complexity of the flow in the shell [1][18]:

 m& 
hs = J c J l J b J s J r  j C ps Pr −2/3  (4)
 A 
 0,cr 
With both correlations, one can determine the overall heat transfer coefficient that is used
to calculate the total required surface area for the heat duty considered.

2.3 Pressure drop

Pressure drop correlations are also based on experimental measurements. The tube-side
pressure drop required to calculate the pumping power is given by [1]:

 4 fL   v2
Pt = s  + 2.5  t t (5)
 di  2
where the friction factor f is determined from the Bhatti and Shah correlation [1]:

0.0054 + 2.3 10 ( Ret )

 −8
2100  Ret  4000
3/2

f = (6)
 0.00128 + 0.1143 ( Ret )
−0.311
 4000  Ret  107

The shell-side pressure drop is also given by correlations, but it is more complicated to
calculate. It involves a series of correction factors that consider the baffle leakages, bypass

7
flow and inlet and outlet baffle spacing which are respectively represented by ζl, ζb and ζs
in Eq.(7). It is also necessary to determine the number of baffles in the shell (Nb), the
number of tube rows in the cross flow of the window section (Nr,cw) and the number of tube
rows crossed during flow through cross flow section between baffle ends (Nr.cc) to calculate
the shell side pressure drop with Eq. (7), see Refs. [1] and [18]:

 N 
Ps = ( N b − 1) Pb ,id  b + Nb Pw,id   l + 2Pb ,id 1 + r,cw   b s (7)
 N 
 r, cc 

The required pumping power can be calculated once the pressure drops of Eqs.(6) and (7)
are established, since pumping power is equal to the pressure drop multiplied by the
volumetric flow rate, i.e. (m&P )  . The energy required for pumping is then simply
obtained by multiplying the pumping power by the number of seconds of operation. Then,
the operating cost is straightforwardly determined by Eq. (2).

2.4 Number of tubes

In the model used in this paper (Bell-Delaware method, see Ref. [1]), the number of tubes
contained in the shell is not a design variable. It is derived from a correlation based on the
value of the design variables. It is important to understand that the number of tubes must
be determined to calculate the required heat transfer surface area, which will affect the cost
function. In the context of this study, the correlation to calculate Nt thus plays a role
identical to the other correlations presented above and its uncertainty must be considered.
The number of tubes depends on the baffle dimension and the tube pitch, and can be
estimated with [1]:

 ( / 4 ) Dctl2
 2 (1 − c ) single tube pass
 Ct pt
Nt =  (8)
 ( / 4 ) Dctl 1 −
2

 C p2 ( n) multiple tube passes

 t t

where, Dctl is the diameter of the circle through the centers of the outermost tubes, Ct is
equal 0.866 for 30° and 60° tube layout and 1.00 for 45° and 90° tube layout, pt is the tube
pitch and ѱn is a correction factor that depends on the number of tube passes and on Dctl.

8
Correction factors are also needed in Eq.(8) in order to considerate the tube layout and the
number of tube pass.

3. Uncertainties related to heat exchanger model

The variables introduced in the previous section all have a certain degree of uncertainty
that is related to the accuracy of the correlations from which those variables are calculated.
In this document, the following six sources of uncertainty were taken into account:
(i) The accuracy of purchase cost estimation from relation such as Eq. (1) is typically
±30% [19];
(ii) The accuracy of the tubes side heat transfer coefficient, ht, as calculated from Eq.
(3), is from 19.8% low to 27.7% high according to [1][17];
(iii) For the shell-side heat transfer coefficient, hs, the accuracy of the correlation (4) is
from 50% low to 200% high, with a mean error of 15% low based on Refs. [1] and
[18];
(iv) The uncertainty of the tube-side pressure drop comes from the accuracy of the
correlation of the tube friction factor, Eq.(6). The accuracy related to this correlation
is of the order of 2% [1];
(v) For turbulent flow in the shell-side, Eq. (7) has an estimated accuracy of ±40% [20]
with a mean error of 5% low for Reynolds numbers greater than 1000 [1][18] (note that
laminar-turbulent transition occurs for shell-side Reynolds number of ~100);
(vi) Finally, the error on the number of tubes obtained from Eq. (8) is ±5% for any
number of pass if the shell diameter is greater or equal to 400mm [1][20].

It is worth to mention even though some correlations are less precise than others in this list,
it does not necessarily mean that they have a more important impact on the heat exchanger
performance. In fact, the purpose of sensitivity analysis techniques is exactly to identify in
a systematic and quantitative way the most influential variables in a problem.

It is important to mention that the accuracy related to those correlations is given relatively
to experimental data. The accuracies given in points (i) to (vi) must then be treated in order
to give the error on the predicted values obtained from the correlations. For example, if we
consider the element (ii), the tube-side heat transfer given by Eq.(3) can be 27.7%

9
overestimated or 19.8% underestimated compared to the real value that would be obtained
for a given heat exchanger. For the sake of illustration, let us consider the maximal
overestimation. The value of ht returned by the correlation would be equal to the real value
of ht multiplied by 1+27.7%. As a result, the real value of ht (i.e. the one that would be
measured experimentally) is the value returned by the correlation multiplied by (1+27.7%)–
1
. To facilitate the analysis, correlation error bounds thus were defined by the following
transformations:

Erlow = (1 + 27.7% ) − 1 = −21.7%

−1
(9)

Erhigh = (1 − 19.8% ) − 1 = 24.7%

−1
(10)

where Erlow is the low and Erhigh is the high error on the result of the tube-side heat transfer
coefficient given by Eq.(3). The same procedure was applied to the other variables. Note
that the high value of accuracy is used in the calculation of the low error and the low value
of accuracy for the high error. Table 1 gives the error bounds for all variables estimated
from correlations.

Table 1: Low, high and mean errors of estimated variables.

Estimated variable Low error High error Mean error

Purchase cost -23.1% 42.9%
Tube-side heat transfer -21.7 24.7%
Shell-side heat transfer -66.7% 100% 17.6% high
Tube-side pressure drop (friction factor) -1.96% 2.04%
Shell-side pressure drop -28.6% 66.7% 5.26% high
Number of tubes -9.09% 11.1%

4. Description of the sensitivity analysis method

To observe the effect of the uncertainties of the correlations on the heat exchanger model,
a variance-based method has been used in this study. This method can be seen as the
decomposition of the total variance of the model into different parts related to each

10
variable. Considering the facts that those variables possess an uncertainty described in the
preceding section and that no information was available on their probability distribution,
statistical distributions were generated in order to simulate the values that those variables
can take. To do so, three different methods of error simulation have been developed.

The first one consists in modeling the error as a normal distribution. The distribution was
then divided into five bands, all having the same probability of occurrence. Then the central
value of each band was multiplied by the error of each correlation. This procedure thus
generated five different values for each variable that were included in the range of possible
values given the uncertainties of the associated correlation. The standard deviation of the
five generated values for the error simulation of the variable i is:

 Xi
i = (11)
1.96
where γ is the relative error given in the preceding section and X i is the central value of

the five bands of equal probability.

The second method for error simulation was based on a Monte Carlo approach. Random
numbers corresponding to the different values that the studied variables can take have been
generated with a normal distribution in order to simulate the uncertainties given by the
correlations used in the model. More specifically, either 20 or 25 random values were
selected for each studied variables considered in this approach. Even though the size of the
samples might seem small, it will be shown below that it was sufficient to obtain a proper
ranking of the variables in terms of their influence in the sensitivity analysis.

The third method consists in dividing the error interval by the number of desired generated
values. In this method, the error is assumed to be a uniform distribution where all values
have the same probability and are separated by a gap corresponding to the error band
divided by the amount of desired values.

By computing all the different possible combinations of variables, it is possible to calculate

an average cost and its variance with the three methods described above.

11
To decompose the variance, it is now necessary to study each variable. To do so, the studied
variable can be fixed to a single value and then the mean value of the total cost has to be
calculated without considering all of the others values that the studied variable can take.
Then, an average cost must be calculated for each other possibility of the fixed variable.
The expected value of the cost C provided that one of the random variables Xi is fixed is

written E ( C | X i ) . After having calculated all the cost mean values for each fixed value

of the observed variables, it is necessary to calculate the variance of those average values
in order to have the first order contribution to the total variance. Eq. (12) shows the first
order variation Vi of the cost C for each value of the fixed variable Xi [21].

Vi = V ( E ( C | X i ) ) (12)

The right-hand side term represents the variance of the expected value of the cost when Xi
is fixed. Repeating the same procedure by fixing two different variables instead of only
one, it is possible to determine the second order variance Vij of the cost C for those two
variables Xi and Xj. Eq.(13) presents the second order variance Vij for two fixed variables
where the first order variances Vi and Vj must be subtracted to remove first order effects
[21]:

( )
Vij = V E ( C | X i , X j ) − Vi − V j (13)

For additive models, it is possible to show that the sum of the variances related to their
variable for each order is equal to the total variance. Eq.(14) shows this equality where
V(C) is the total variance calculated with each possible combinations of variables [21]:

V ( C ) =  Vi +  Vij + ... + V12...k (14)

i i j i

By now dividing each terms in the Eq.(15) by V(C), it is now possible to obtain sensitivity
indices S [21]:

1 =  Si +  Sij + ... + S12...k (15)

i i j i

where all S-terms are related to a part of the decomposed variance and give the importance
of a variable or a set of variable on the total variance of the model. The total effect of a

12
variable on a model is given by Eq.(16) which illustrates the sum of all the contribution for
each order of the variable i where i≠j≠k and l is the total number of variables [21]:

STi = Si +  Sij +  Sijk + ... + S12...l (16)

j j k j

It is also possible to express the total effect by calculating the variance of the mean values
of the cost by fixing all variables but the studied one as shown in Eq.(17) [21]:

V  E ( C | X : i ) 
STi = 1 − (17)
V (C )

where E ( C | X : i ) is the expected value of the cost when all variables are fixed expect Xi.

Note that for perfectly additive models, the sum of all STi is equal to one and greater than
one for nonadditive models. Moreover, if a difference exists between the total effect and
the first order effect of a variable such as 𝑆𝑇𝑖 − 𝑆𝑖 > 0, then it is said that Xi is involved in
interactions with other input factors [21].

5. Description of the three test cases

Three different test cases taken from literature are considered in this study. The
optimization of the heat exchanger design for these specific cases was also reported in
literature [3], which makes them convenient for the present work.
The first test case consists in crude oil passing in the tube-side that has to be cooled down
from 200 to 95°C by kerosene that enters the shell at a temperature of 35°C. The material
of construction of the shell and the tubes is carbon steel and Table 2 shows the fluids
properties and their mass flow rates. The reported minimized total cost related to this case
without considering the uncertainties was found to be 10,578$.

13
 Table 2: Heat transfer duty for the test case 1.

Tube-side Shell-side
Fluid Crude oil Kerosene
Viscosity (N s m-2) 0.0036 0.0004
Heat capacity (J kg-1 K) 2050 2470
Thermal conductivity (W m-1 K) 0.132 0.133
Density (kg m-3) 830 730
Flow rate (kg s-1) 85.49 27.01
Temperature (°C) 35/- 200/95
Fouling resistance (m2 K W-1) 0.0005 0.0002
Material of construction Carbon steel Carbon steel

The second test case consider heavy gas oil entering the tubes at a temperature of 319°C
and leaves cooled at 269°C by crude oil entering the shell at 226°C. It was reported that
the optimized solution related to this application gives a minimized total cost of 10,589$
and the fluids properties related to this test case are shown in Table 3.

Table 3: Heat transfer duty for the test case 2.

Tube-side Shell-side
Fluid Heavy gas oil Crude oil
Viscosity (N s m-2) 0.00032 0.00049
Heat capacity (J kg-1 K) 3161.03 2679.55
Thermal conductivity (W m-1 K) 0.089 0.1
Density (kg m-3) 678 723
Flow rate (kg s-1) 29.36 102.12
Temperature (°C) 319/269 209/-
Fouling resistance (m2 K W-1) 0.0006 0.0006
Material of construction Carbon steel Carbon steel

The last test case that is treated in this paper is Naphtha in the shell-side cooled from 114
to 40°C by cooling water at 33°C. The tubes are made of stainless steel and the shell of
carbon steel. Table 4 shows the fluids properties related to this test case and the reported
minimized total cost of this solution is 4,455$.

14
 Table 4: Heat transfer duty for the test case 3.

Tube-side Shell-side
Fluid Cooling water Naphtha
Viscosity (N s m-2) 0.00071 0.00037
Heat capacity (J kg-1 K) 4186.8 2646.06
Thermal conductivity (W m-1 K) 0.63 0.11
Density (kg m-3) 1000 656
Flow rate (kg s-1) 30 2.7
Temperature (°C) 33/- 114/40
Fouling resistance (m2 K W-1) 0.0004 0.0002
Material of construction Stainless steel Carbon steel

6. Results of the sensitivity analysis

6.1 Sensitivity analysis for the three test cases
First, for the three test cases, a series of scenarios was generated with the three methods
described in Section 4 and the total cost was evaluated for each scenario. Then, the average
total cost and the standard deviation were calculated, for each test case and with each of
the three methods. Results are reported in Table 5. It is possible to observe that values
obtained from the Monte Carlo approach vary from one try to another. This behavior can
be explained by the fact that the Monte Carlo approach is a stochastic method.
Nevertheless, all methods give average costs that are coherent with the total costs which
were previously calculated in [3] without any error analysis. However, the average costs
presented in this paper are always slightly higher than the costs from [3]. This is explained
by the lack of symmetry of the uncertainties that where considered in this paper. Indeed,
high errors from Table 1 are greater than low errors which results in greater chances to
obtain a higher total cost than one calculated without considering correlations accuracy. It
can also be seen in Table 5 that the standard deviations are relatively important. This
emphasizes the importance in a design optimization procedure not only to minimize
“average” total cost, but also the fact that the cost estimation is imprecise and can represent
a risk of a significant cost overrun.

15
 Table 5: Average costs and standard deviations for the three test cases.
Test case,
Method Average cost Standard deviation
Ref. [3]
Normal distribution 10,943.73 $ 1,666.01 $
Monte Carlo 20 No 1 10,854.99 $ 1,928.95 $
Test case 1 Monte Carlo 20 No 2 10,718.74 $ 1,911.73 $
10,578.00$ Monte Carlo 25 No 1 11,756.24 $ 1,911.71 $
Monte Carlo 25 No 2 10,724.21 $ 1,884.78 $
Uniform distribution 20 11,773.30 $ 2,229.49 $
Normal distribution 10,947.07 $ 1,736.67 $
Monte Carlo 20 No 1 11,558.85 $ 2,092.09 $
Test case 2 Monte Carlo 20 No 2 11,583.63 $ 2,247.79 $
10,589.10$ Monte Carlo 25 No 1 11,478.96 $ 1,969.32 $
Monte Carlo 25 No 2 11,568.93 $ 1,829.06 $
Uniform distribution 20 11,874.80 $ 2,357.76 $
Normal distribution 4,500.08 $ 867.65 $
Monte Carlo 20 No 1 4,657.46 $ 1,047.06 $
Test case 3 Monte Carlo 20 No 2 4,304.70 $ 783.94 $
4,455.13$ Monte Carlo 25 No 1 4,674.27 $ 987.33 $
Monte Carlo 25 No 2 4,728.82 $ 1,057.38 $
Uniform distribution 20 4,988.73 $ 1,279.65 $

The total effect (calculated from Eq. (17)) of each of the “uncertain” variables is presented
in Figure 1. Note that a log-scale was chosen for this figure to illustrate the difference in
terms of importance of the uncertainty on each variable, since some of them appear to have
virtually no effect. It shows that the variance of the total cost for test cases 1 and 2 mainly
comes from the uncertainty related to the capital cost correlation. Uncertainty of the shell-
side heat transfer coefficient, which is calculated with Eq. (4), also represents a significant
contribution to the variance of the total cost while the uncertainty related to the tube-side
heat transfer coefficient plays less important a role for the two same test cases. This can be
explained by the fact that the uncertainty on the shell-side heat transfer coefficient is greater
than the uncertainty related to the tube-side heat transfer coefficient.

As for the third test case, the importance of the uncertainty on the capital cost is similar to
that from the shell-side heat transfer coefficient. This can be explained by the fact that the
shell-side heat transfer coefficient has a higher value for the test case 3 and plays an
important role in the heat transfer calculation for this case. Thus, a variation of the shell-
side heat transfer coefficient brings relatively high variations of the average cost.

16
Moreover, sensitivity related to uncertainties on the friction factor, shell-side pressure drop
and tubes number are not important compared to the other sources of uncertainty described
in this paragraph. In fact, the friction factor and the shell-side pressure drop are used to
calculate the cost for the required pumping energy which is small compared to the purchase
cost in the present cases. Considering the fact that the number of tubes is used in different
equations, its small incidence can be explained by the small uncertainty to which it is
subjected. Thus, in order to lower the total cost variance, the uncertainties related to the
capital cost and the shell-side heat transfer coefficient correlations are the ones to be
minimized.

It should be noted that all sampling methods reveal the same trends in terms of which
variables are the most influential, even if the sampling number of the Monte Carlo approach
can seem relatively small. In other words, a similar ranking of the variables regarding the
impact of their uncertainties is achieved with all samples, even if the exact values of the
sensitivity coefficient can be slightly different.

6.2 High order effects

In order to verify the presence of higher order effects, the first order sensitivity is calculated
and subtracted from the total sensitivity presented in Figure 1 for each variable. The result
of this operation is presented in Figure 2 and shows that higher order effects are greater for
uncertainties related to the capital cost and shell-side heat transfer coefficient. However,
those higher order effects remain negligible compared to first order sensitivity.

6.3 Probability distribution of total cost

The probability distribution of the possible total cost while taking into account the
uncertainties are shown in Figure 3 for the three test cases. Again, it is visible in this figure
that the uncertainties on the correlations can result in quite important variations to the
estimated total costs.

Considering these results, one sees that the distributions varied slightly between different
runs with the Monte Carlo approach, given the relatively small number of design points
simulated. Moreover, by comparing the distributions obtained from the three methods used

17
in this study, it is possible to see that the breadth of the cost distribution with the normal
distribution method are usually somewhat thinner than the ones obtained with the other
methods. However, this situation does not seem to influence significantly the average costs
while only reducing slightly the standard deviations (see Table 5). Moreover, it is also
possible to see that the uniform distribution method leads to curves that are smoother than
those obtained from the other methods. This can be explained for the normal distribution
method by the fact that it requires less data in its calculation which lead to greater steps
between possible average costs calculated. Regarding the Monte Carlo method, the fact
that the occurrence distributions of this method are less smooth than the ones from the
uniform distribution method is explained by the accumulation of values that are close to
each other during the random generation. This behavior can then lead to a nonuniform
distribution of the average cost calculated and steps in the occurrence distribution for the
Monte Carlo method of Figure 3 (a,b,c).

6.4 Variation of the sensitivity analysis results in the design space

The sensitivity analyses presented above were performed for given heat exchanger designs.
These designs were obtained by an optimization procedure in Ref. [3]. It was also verified
to what extent the results of the sensitivity analysis were affected by the heat exchanger
design for which the sensitivity analysis is performed. For the sake of illustration, in Figure
4, a “non-optimal” heat exchanger design was considered for test case 1 and a sensitivity
analysis was performed with the uniform distribution approach. This “non-optimal” design
corresponds to a set of the design variables different than the optimal ones and was chosen
randomly in the design space to investigate the impact of the location in the design space
on the sensitivity analysis. After the sensitivity analysis around that specific design, the
average cost was 244,363$ with a standard deviation of 47,013$ for the cost distribution.
Compared to the average cost and standard deviation reported in Table 5, this shows that
the point analysed here is clearly far from the optimum. The total effect coefficients are
reported in Fig. 4. For the sake of comparison, the results that were achieved previously
for the optimal design are also shown in Figure 4. It is visible that the results are quite
different between the two designs. For the non-optimal design, the uncertainties on the
shell-side pressure drop estimation caused the greater variance of the total cost. This

18
situation can be explained by the fact that the unoptimized design has a very large shell-
side pressure drop which leads to an operation cost that is greater by several orders of
magnitude than the initial cost. The initial cost correlation then loses its importance in favor
of the shell-side pressure drop correlation. This shows that depending on where a design is
located in the design space, the ranking of correlations in terms of influence can be
different.

Moreover, by performing a sensitivity analysis of selected non-optimized designs covering

the entire design space, it was possible to find the extreme values of the sensitivity indices
for each correlation for the first test case. In other words, everywhere in the design space,
a sensitivity analysis was performed locally, and then, the maximal and minimal values of
the sensitivity index for each variable were identified. Table 6 shows the extrema of the
sensitivity indices related among all designs tested and their associated average cost and
standard deviation. Thus, we can see that all the correlations except the one used to
calculate the friction coefficient in the tubes can take significant relative importance in the
design space. As a result, the error on the friction coefficient correlation could be neglected
in future work in order to minimize calculation times, at least for test case 1.

Table 6 : Sensitivity index extrema, and corresponding average cost and standard
deviation, as a function of correlations
Correlation Index extrema Average cost Standard deviation
fmax 8.18×10-3 10,288,282.40 $ 1,505,720.95 $
fmin 7.71×10-7 392,878.90 $ 121,231.00 $
ΔPsmax 8.32×10-1 264,615.95 $ 67,568.43 $
ΔPsmin 1.48×10-8 600,378.58 $ 261,352.95 $
hsmax 9.77×10-1 224,590.95 $ 133,489.45 $
hsmin 1.37×10-1 264,615.95 $ 67,568.43 $
htmax 1.29×10-1 22,380.03 $ 5,083.74 $
htmin 5.09×10-5 3,546,873.33 $ 2,188,045.49 $
Cmax 6.25×10-1 13,734.40 $ 3,359.93 $
Cmin 4.08×10-8 25,090,635.68 $ 12,238,423.24 $
Ntmax 4.15×10-1 10,288,282.40 $ 1,505,720.95 $
Ntmin 8.60×10-5 26,146.27 $ 10,827.28 $

6.5 Impact of uncertainties on heat exchanger design

The previous sections have revealed that uncertainties can be quite important when
estimating heat exchanger total cost. When taking into account these uncertainties, it

19
follows that for a given set of design variables, one actually obtains a probability
distribution of possible total costs. From this distribution, it is possible to evaluate an
‘average cost’ (or more rigorously, an expected value), but also the standard deviation of
the distribution, which reveals the risk that a given heat exchanger design turns out more
or less expensive than expected. This section illustrates how uncertainties could be taken
into account in heat exchanger design procedure. The first “test case” was chosen for this
purpose. An optimization problem was formulated with two simultaneous objectives,
namely the minimization of expected total cost and the minimization of the standard
deviation of the cost distribution. A multi-objective genetic algorithm (NSGA-II) was used
to solve the problem. Preliminary results for that problem were reported in Ref. [22]. Figure
8 shows an example of a resulting Pareto front. Each point represents a non-dominated
solution. In other words, for a given total cost, there is no solution providing a lower
standard deviation than the one in Fig. 8, and vice versa. For the sake of comparison, the
average cost and standard deviation for the same test case and sampling procedure is shown
in Fig. 8 by a small cross and is slightly above the Pareto front. Two main conclusions can
be drawn. First, it should be noted that the total costs in Fig. 8 are higher than the one
obtained when no uncertainty is considered, i.e. when only the expected value of each
uncertain variable is used in assessing heat transfer performance. As was noted earlier, the
probability distribution is not symmetrical. The expected value of the total cost is thus
different from the total cost obtained when not considering uncertainties. Second, one sees
in Fig. 8 that when reducing the expected total cost, the standard deviation (i.e. the risk for
a design to be more expensive than estimated) actually increases, and vice versa. In other
words, the designs expected to be the cheapest can turn out to be “riskier” than others.
Larger surface area heat exchangers tend to are chosen in order to obtain small expected
costs, whereas smaller heat exchangers are preferred to minimize the standard deviation.
These two conclusions illustrate the importance of understanding how error propagate in
heat exchanger design procedures.

7. Sensitivity related to the energy cost

One of the important parameter that is highly uncertain is the value of the energy. This
value can fluctuate significantly from year to year. An attempt was made to establish the

20
impact of this on the sensitivity analysis. For the sake of illustration, test case 1 was
considered. For the design in Table 2, the sensitivity analysis was repeated for different
cost of energy. Figure 5 reports that when the energy cost increases, the sensitivity related
to the capital cost error decreases while the sensitivity related to other sources of
uncertainty gain in importance. This happens because a greater contribution to the total
cost comes from the energy needs and thus, decreases the relative importance of the capital
cost.

Moreover, Figure 6 shows the total cost probability distribution for three different energy
cost where it is possible to see that while the energy cost increases, the occurrence
distribution moves to the right on the cost axis. This behavior, is accompanied by an
augmentation of the standard deviation of the average cost and therefore of the economic
risk related to the heat exchanger. Table 7 shows the average cost, its standard deviation
and the maximal cost according to the energy cost for the test case 1.

Table 7 : Average cost and standard deviation for different energy costs
Energy cost Average cost Standard deviation
0.05 $/kWh 11,074.29 $ 2,230.90 $
0.10 $/kWh 11,796.77 $ 2,281.53 $
0.15 $/kWh 12,519.25 $ 2,337.05 $

In order to observe the influence of the energy cost, another sensitivity analysis was made
by adding an arbitrary uncertainty of ±50% to the energy cost and considering energy cost
as an uncertain parameter. Considering the fact that the uncertainty related to the friction
factor, the shell-side pressure drop and the number of tubes have a small impact on the total
cost, they were not considered in this second sensitivity analysis in order to decrease the
calculation time. The sensitivity indices are shown in Figure 7 for the three test cases
studied in this paper. It is possible to see that the uncertainty related to the energy cost only
has a weak importance compared to the other uncertainties for these cases. Note that the
uniform distribution method was used for the sensitivity analyses of this section. Based on
the results presented in the other sections, similar results would be expected with the other
sampling methods.

21
8. Conclusions
Several correlations are typically used when assessing the performance of a heat exchanger.
This paper presents a sensitivity analysis of how the estimation of the total cost of shell-
and-tube heat exchangers is affected by the uncertainties of different correlations.
Sensitivity indices and probability distributions of the total cost were calculated for three
test cases, based on the reported precision of the correlations. The main contributions of
this work can be summarized by:

• The probability distribution of total cost can be relatively wide, meaning that there
is a large uncertainty on the total cost estimation (standard deviation of total cost
over expected total cost can be as high as 25%).
• The uncertainty that has the greatest influence for previously optimized designs was
that relating the purchase cost to the surface area, emphasizing that more precise
cost estimation approaches such as that in [14] could be helpful in achieving better
and less “risky” designs. The second most influential uncertainty was that used to
estimate the shell heat transfer coefficient.
• The paper demonstrates that the most influential uncertainties are not always the
same in the design space, i.e. that the total effect of the different variables changes
depending on which design is evaluated.
• The total effect due to the uncertain cost of energy was shown to be relatively small
compared to that of other variables.
• Finally, an example of heat exchanger design optimization based on minimizing
expected total cost and standard deviation is presented, revealing a set of non-
dominated optimal solutions.

The paper shows the relevance of developing correlations or meta-models that are more
precise. In particular, due to their larger influence, the correlations for the purchase cost
and the correlation for the shell side heat transfer coefficient were the ones that deserve to
receive more attention. Future work, either experimental or numerical, could thus focus on
reducing the uncertainties associated with these widely used correlations.

22
Furthermore, the paper highlights the need to develop design procedures that take into
account uncertainties [23], for example by optimizing heat exchangers while minimizing
simultaneously the average and the variance of the total cost probability distribution.

Acknowledgements
This work was supported by the Natural Sciences and Engineering Research Council of
Canada (NSERC).

References

[1] R. K. Shah, Fundamentals of heat exchanger design. Hoboken, NJ: John Wiley &
Sons, 2003.
[2] I. H. Kim, X. Zhang, R. Christensen, et X. Sun, « Design study and cost assessment of
straight, zigzag, S-shape, and OSF PCHEs for a FLiNaK-SCO2 Secondary Heat
Exchanger in FHRs », Ann. Nucl. Energy, vol. 94, p. 129‑137, août 2016.
[3] P. Wildi-Tremblay et L. Gosselin, « Minimizing shell-and-tube heat exchanger cost
with genetic algorithms and considering maintenance », Int. J. Energy Res., vol. 31, no
9, p. 867‑85, juill. 2007.
[4] B. V. Babu et S. A. Munawar, « Differential evolution strategies for optimal design of
shell-and-tube heat exchangers », Chem. Eng. Sci., vol. 62, no 14, p. 3720‑3739, juill.
2007.
[5] R. Selbas, O. Kizilkan, et M. Reppich, « A new design approach for shell-and-tube
heat exchangers using genetic algorithms from economic point of view », Chem. Eng.
Process., vol. 45, no 4, p. 268‑275, avr. 2006.
[6] L. Wang, L. Deng, C. Ji, E. Liang, C. Wang, et D. Che, « Multi-objective optimization
of geometrical parameters of corrugated-undulated heat transfer surfaces », Appl.
Energy, vol. 174, p. 25‑36, juill. 2016.
[7] J. Du, M.-N. Yang, et S.-F. Yang, « Correlations and optimization of a heat exchanger
with offset fins by genetic algorithm combining orthogonal design », Appl. Therm.
Eng., vol. 107, p. 1091‑1103, août 2016.
[8] J. Y. Q. Wong, S. Sharma, et G. P. Rangaiah, « Design of shell-and-tube heat
exchangers for multiple objectives using elitist non-dominated sorting genetic
algorithm with termination criteria », Appl. Therm. Eng., vol. 93, p. 888‑899, janv.
2016.
[9] H. V. Hultmann Ayala, P. Keller, M. de F. Morais, V. C. Mariani, L. dos S. Coelho, et
R. V. Rao, « Design of heat exchangers using a novel multiobjective free search
differential evolution paradigm », Appl. Therm. Eng., vol. 94, p. 170‑177, févr. 2016.
[10] O. E. Turgut, « Hybrid Chaotic Quantum behaved Particle Swarm Optimization
algorithm for thermal design of plate fin heat exchangers », Appl. Math. Model., vol.
40, no 1, p. 50‑69, janv. 2016.
[11] D. K. Mohanty, « Application of firefly algorithm for design optimization of a shell
and tube heat exchanger from economic point of view », Int. J. Therm. Sci., vol. 102,
p. 228‑238, avr. 2016.

23
[12] « Wiley: Fundamentals of Heat and Mass Transfer, 7th Edition - Theodore L.
Bergman, Adrienne S. Lavine, Frank P. Incropera, et al ». [En ligne]. Disponible sur:
http://ca.wiley.com/WileyCDA/WileyTitle/productCd-EHEP001810.html. [Consulté
le: 15-mars-2017].
[13] J. Wen, H. Yang, G. Jian, X. Tong, K. Li, et S. Wang, « Energy and cost
optimization of shell and tube heat exchanger with helical baffles using Kriging
metamodel based on MOGA », Int. J. Heat Mass Transf., vol. 98, p. 29‑39, juill. 2016.
[14] A. C. Caputo, P. M. Pelagagge, et P. Salini, « Manufacturing cost model for heat
exchangers optimization », Appl. Therm. Eng., vol. 94, p. 513‑533, févr. 2016.
[15] W. C. Long, R. Luck, et P. J. Mago, « Uncertainty based operating strategy
selection in combined heat and power systems », Appl. Therm. Eng., vol. 98, p.
1013‑1024, avr. 2016.
[16] R. Smith, Chemical process design and integration. Chichester, West Sussex,
England ; Hoboken, NJ: Wiley, 2005.
[17] S. Kakaç, R. K. Shah, et W. Aung, Éd., Handbook of single-phase convective heat
transfer. New York: Wiley, 1987.
[18] Advanced Study Institute on Heat Transfer Equipment, Heat transfer equipment
design. New York: Hemisphere Pub. Corp, 1988.
[19] G. P. Towler, Chemical engineering design: principles, practice, and economics of
plant and process design, 2nd ed. Boston, MA: Butterworth-Heinemann, 2013.
[20] G. F. Hewitt, Éd., HEDH: heat exchanger design handbook 2002, vol. 3, 4 vol.
New York: Begell House, 2002.
[21] A. Saltelli, Global sensitivity analysis the primer. Chichester, England; Hoboken,
NJ: John Wiley, 2008.
[22] Julien Lambert et Louis Gosselin, « Heat exchanger design optimization taking into
account uncertainties of different correlations », in 4th International Conference on
Engineering Optimization, Lisbon, Portugal, 2014.
[23] M. J. Mohideen, J. D. Perkins, et E. N. Pistikopoulos, « Optimal design of dynamic
systems under uncertainty », Aiche J., vol. 42, no 8, p. 2251‑2272, août 1996.

24
Figure captions
Figure 1 Total sensitivity of studied variables (STi on a log-scale) on the total cost,
for the three test cases and with the different error distribution approaches.
Figure 2 Presence of higher order effects (STi – Si), for the three test cases and with
the different error distribution approaches.
Figure 3 Total cost probability distribution with the three error distribution
approaches, for: a) test case 1, b) test case 2 and c) test case 3.
Figure 4 Total sensitivity (STi on a log scale) on total cost for two different designs
and for case 1.
Figure 5 Total sensitivity (STi) of studied variables for different energy cost, for
case 1.
Figure 6 Total cost probability distribution for different energy cost (test case 1).
Figure 7 Total sensitivity (STi) of total cost with respect to uncertain variables
including the energy cost, for case 1.
Figure 8 Example of simultaneous minimization of total cost standard deviation and
expected total cost for test case 1.

25
 test test test
case 3 case 2 case 1

26
Figure 1
 test test test
case 3 case 2 case 1

27
Figure 2
 (a)

(b)

28
 (c)

Figure 3

29
Figure 4

30
Figure 5

31
Figure 6

32
Figure 7

33
 From Table 5

Figure 8

34