Article

Cite This: Ind. Eng. Chem. Res. 2019, 58, 12080−12096 pubs.acs.org/IECR

Optimal Design of Double Pipe Heat Exchanger Structures

Alice Peccini,† Julia C. Lemos,‡ André L. H. Costa,*,‡ and Miguel J. Bagajewicz+
†
Federal University of Rio de Janeiro (UFRJ), Escola de Química CT, Bloco E, Ilha do Fundão, Rio de Janeiro, Rio de Janeiro CEP
21949-900, Brazil
‡
Rio de Janeiro State University (UERJ), Rua São Francisco Xavier, 524, Maracanã, Rio de Janeiro, Rio de Janeiro CEP 20550-900,
Brazil
+
School of Chemical, Biological and Materials Engineering, University of Oklahoma, Norman, Oklahoma 73019, United States
*
S Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: This paper investigates the design optimization of

double pipe heat exchanger using mathematical programming.
Downloaded via UNIV OF OKLAHOMA on October 14, 2019 at 01:02:32 (UTC).

The heat exchanger area is minimized and the thermo-fluid

dynamic conditions are considered for the use of the right
transport correlations, together with design specifications, such
as, maximum pressure drops and minimum excess area. The
modular nature of this kind of heat exchanger and the allocation
of the streams (inside the inner tube or in the annulus) are also
contemplated. Two mixed-integer nonlinear programming
(MINLP) approaches are proposed. One approach relates the
binary variables to the nonlinear constraints directly. In the
second, the resulting nonlinearities involving binary variables are formally linearized, without loss of rigor (e.g., no use of
truncated Taylor series). The proposed methodology can get better solutions than traditional trial and error procedures. The
flexibility of the model is illustrated, together with a comparison between the performances of both MINLP formulations.
Additionally, computational time and local optimality issues are discussed.

1. INTRODUCTION determined through a sensitivity analysis study. In turn, Syed

Although shell-and-tube heat exchangers are the most common et al.6 investigated the optimal configuration of the annulus with
heat transfer equipment in chemical process plants,1 there are trapezoidal fins through the selection of the number of fins, fin
many cases for which other heat exchanger types become more height, fin thickness, and the radius ratio of the inner and outer
suitable. For instance, double pipe heat exchangers can be an tubes, using genetic algorithms and a trust region method. A
economically more advantageous option when smaller services similar investigation was also proposed by the same authors
are in place2 (e.g., heat transfer area lower than 50 m2). If the using parabolic fins7 and using a generalized optimization of the
stream contains solids in suspension, double pipe heat fin shape.8 Later, Han et al.9 investigated the design of the outer
exchangers may also be a better alternative, because they can surface corrugation of the inner tube using multiobjective
be built with an inner tube with larger diameter to avoid optimization for the determination of optimal values of the
plugging. Smaller diameters of the outer tube in double pipe heat pitch, height, and radius of the corrugation. Their optimization
exchangers are advantageous for high-pressure services, because was based on a response surface methodology (RSM) using
it implies a smaller wall thickness. In addition, double pipe heat simulation data generated through computational fluid dynam-
exchangers are easily cleaned, and the longitudinal flow avoids ics (CFD). Finally, Dastmalchi et al.10 employed a particle
the existence of stagnation regions, which in shell and tube swarm algorithm (PSO) associated with a CFD model for the
exchangers are prone to fouling.3 Double pipe heat exchangers optimization of the inner microfinned tube surface, contemplat-
have also the benefit of flexibility due to its modular structure, ing the selection of the number of microfins, the microfin height,
which allows an easier adaptation to process modifications. and the microfin helix angle.
Several papers addressed the design of double pipe heat The aforementioned studies did not account for the
exchangers; however, a large number of them are focused on operational and capital costs in their objective functions and
heat transfer enhancement devices, rather than on the general were focused on enhancing the heat transfer variables such as
optimization of the equipment design (e.g., diameter and length Nusselt number and heat transfer coefficient, sometimes the
of the tubes were usually considered fixed parameters).4 For
example, Sahiti et al.5 investigated the optimization of pin fins in Received: March 21, 2019
the heat exchanger annulus aiming at the minimization of Revised: May 16, 2019
entropy generation. Their heat exchanger model was based on Accepted: May 22, 2019
experimental data and the optimal set of variables was Published: May 22, 2019

© 2019 American Chemical Society 12080 DOI: 10.1021/acs.iecr.9b01536

Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

minimization of entropy generation, but never discussing a expressed in a form of discrete choices, as in industrial practice, is
complete optimal design. used as a convenient tool. A modified MINLP model, for which
Reducing costs is, after all, the main motivation for basic mathematical transformations are applied to exclude non-
design optimization in practice. The literature involving the linearities involving binary variables is also presented, so that
solution of the least-cost design problem of double pipe heat methods requiring linearity in binary variables (like Outer
exchangers is scarce and the available papers were based on Approximation14) can solve the model.
limited search spaces of the geometric variables. First, This paper is organized as follows: Initially the model is
Söylemez11 investigated the optimization of double pipe heat presented, followed by the first corresponding MINLP
exchangers for waste heat recovery, aiming at identifying the formulation and the modified MINLP problem formulation.
inner tube diameter, with a fixed ratio of the diameters of the To finish, numerical results are presented.
outer and inner tubes, minimizing capital and operating costs.
Later, Swamee et al.12 presented the optimization of double pipe 2. HEAT EXCHANGER ARCHITECTURE
heat exchangers focusing on operational costs, considering the The basic structure of a double pipe heat exchanger consists of
heat load, the pumping power, and the utility consumption. The two concentric tubes (Figure 1) and is usually commercialized in
decision variables were the inner and outer tube diameters and
the utility flow rate, but the tube length was assumed known. In
addition, both studies considered the design variables as
continuous variables, assuming implicitly that they may lead to
suboptimal solutions when the nearest commercially available
pipe is selected.
Finally, academic books2,13 usually use a larger set of design
variables (geometric dimensions, fluid allocation, and arrange-
ments with multiple units), and they use a traditional trial and Figure 1. Double pipe heat exchanger.
verification procedure; in such procedure, design variables are
selected first, followed by the calculation of the number of
a hairpin structure (Figure 2). If the thermal service demands a
required hairpins for that configuration. Then, if the resulting
higher heat transfer area, several hairpins can be interconnected.
heat exchanger is not feasible (e.g., the available pressure drop
For example, Figure 3 illustrates two hairpins connected in
for a given stream is exceeded and/or the stream velocities are series.
not within given bounds) a design change is proposed and the
calculations remade. This work points out that this approach is
dependent on the designer experience and does not guarantee
any optimality. The designer choices for new trials are various:
one can change length, pipe diameters, hairpin arrangements,
and other features aiming at reducing pressure drop and/or
increasing heat transfer coefficient, and the experts develop the
right intuitive choices to end up with a feasible exchanger, which
is the goal. It is for this reason that it is rare to see a set of rules as Figure 2. Hairpin structure.
what adjustment in geometry leads to the right answer. In
addition, it is unusual to see a trial−verification−improvement
procedure in which changes are proposed to seek improvements
over the feasible solutions found. In other words, in trial and
verification, optimality is definitely not the goal, and even in
trial−verification−improvement, optimality is not guaranteed.
To ameliorate the aforementioned deficiencies in the quest
for the optimal design, this paper presents the economic
optimization of the design of double pipe heat exchangers using
a mathematical programming method that at the very least
guarantees local optimality. The models are based on the hairpin
as the elementary component, arranged in series to form what is
here defined as a unit. These units are then considered as the
Figure 3. Two hairpins connected in series.
basic component of arrangements in series, parallel, and series-
parallel, and thus the modular structure of double pipe heat
exchangers is explored.
In addition, the set of design variables considered in this study Different interconnection patterns among the heat exchanger
is broader than the one used in previous literature papers, hairpins provide flexible design alternatives better suited to
considering the allocation of the streams (inside the inner tube attain the heat load and maximum pressure drop specifications
or in the annulus), the inner and outer tube diameters, the tube of the service. The mathematical formulation proposed explores
length, the number of parallel branches, the number of units in a general structure that encloses different arrangements. For
series and in parallel in each branch, and the number of hairpins that, the adopted nomenclature for the elements involved is as
per unit, which determine the arrangement of the existent follows:
hairpins. The optimization problem is formulated as a mixed- • hairpin: basic structure shown in Figure 2. The flow
integer nonlinear programming (MINLP) model. The fact that arrangement for a hairpin in this paper is always
several geometric parameters (diameters, lengths, etc.) can be countercurrent
12081 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Figure 4. Arrangements. (i) Type I, streams aligned in series; (ii) Type II, tube-side stream aligned in series and annulus-side stream in parallel; (iii)
Type III, annulus-side stream in series and tube-side stream in parallel. (a) Generic representation; (b) specific examples.

• unit: multiple hairpins connected in series (Figure 3 • diameter and length of the tubes of the heat exchanger
shows an example of a unit of two hairpins) hairpins (all hairpins are equal)
• branch: A structure composed of units that can be • layout (number of hairpins per unit, number of units in
arranged in three different ways, as illustrated in Figure 4: parallel for each flow side per branch and the number of
(i) tube-side and annulus-side streams aligned in series, branches)
(ii) tube-side stream aligned in series and annulus-side
stream aligned in parallel, and (iii) tube-side stream The traditional heat exchanger design equations (LMTD
aligned in parallel and annulus-side stream in series method and Darcy−Weisbach equation) are employed,
minimizing the total heat transfer area, constrained by maximum
• general structure: The association of one of the structures allowed pressure drops. Finally, it is only considered streams
shown in Figure 4 in a set of parallel branches, as without phase change and constant physical properties
illustrated by Figure 5 for the tube-side stream aligned in (representing average values), but considering any flow regime
parallel and annulus-side stream in series. (laminar, transitional, and turbulent). The LMTD method uses
The design problem involves the determination of the a logarithmic mean temperature difference given by inlet and
following: outlet temperatures of counterflow patterns. Although in Figure
4 each unit corresponds to a number of hairpins of counter-
• stream allocation of the hot and cold streams (tube-side or current flow, the overall flow pattern in Figure 4, (ii) and (iii)
annulus-side) does not correspond to a true countercurrent flow. To account
12082 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Figure 5. Double pipe heat exchanger general structure with multiple parallel branches: (a) generic representation; (b) three type III parallel branches
of four units.

for the deviation of the internal flow pattern from that of a sd max
countercurrent flow, the logarithmic mean temperature differ- ∑ ydsd = 1
ence is multiplied by a correction factor. This factor starts at 1 for sd = 1 (6)
pure countercurrent flow and reduces its value as the number of
units per branch increases. Because the factor is obtained sD max
analytically, it is a rigorous correction and therefore does not ∑ yDsD = 1
change the accuracy of the formulation. sD = 1 (7)

3. MODEL FORMULATION sLh max

The problem constraints are composed of the representation of ∑ yLhsLh = 1

the geometric variables, stream allocation equations, structural sLh = 1 (8)
constraints, thermal and hydraulic modeling, and pressure drop
and velocity bounds. The problem parameters that are kept where dte and dti are the outer and inner diameters of the inner
constant during the optimization, are identified with a symbol ∧ tube, Dte and Dti are the outer and inner diameters of the outer
on top. tube, and Lh is the tube length of each hairpin. The
3.1. Discrete Representation of Geometric Variables. corresponding binary variables which indicate the discrete
The diameter and length of the tubes employed in the options selected are ydsd for the inner tube diameter (discrete
construction of the heat exchanger are usually selected from a values: pdte sd̂ and pdti sd̂ ), yDsD for the outer tube diameter
set of discrete values according to the available standard options. (discrete values: pDte ̂ and pDti ̂ ), and yLhsLh for the tube
sD sD
This design feature imposes the following relations involving the
̂ ).
length (discrete values: pLh sLh
sets of binary variables that represent the available options:
sd max
The selection of the number of parallel branches present in
the heat exchanger design (NB) is represented by the binary
dte = ∑ pdte sd̂ ydsd
variables yBsB:
sd = 1 (1)
sBmax
sd max
NB = ∑ pNB sB̂ yBsB
dti = ∑ pdti sd̂ ydsd sB = 1 (9)
sd = 1 (2)
sBmax
sD max
Dte = ∑ pDte sD̂ yDsD ∑ yBsB = 1
(3) sB = 1 (10)
sD = 1

sD max where pNB sB̂ is the set of discrete values of the possible number
Dti = ∑ pDti sD̂ yDsD of parallel branches (1, 2, ..., sBmax).
sD = 1 (4) The number of units aligned in parallel in each branch for the
tube-side stream (NPt) and for the annulus-side stream (NPa)
sLh max
̂ yLh as well as the number of countercurrent hairpins per unit (Nh)
Lh = ∑ pLh sLh sLh are also represented using binary variables according to their
sLh = 1 (5) integer nature:
12083 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

sE max
NPt = ∑ pNE sÊ yPtsE
sE = 1 (11)
sE max
NPa = ∑ pNE sÊ yPasE
sE = 1 (12)
sNh max
Nh = ∑ ̂ yNh
pNh sNh sNh
sNh = 1 (13)
sE max
∑ yPtsE = 1
sE = 1 (14)
sE max
∑ yPasE = 1
sE = 1 (15)
Figure 6. General structure of three parallel branches (yBsB=3 = 1).
sNh max
∑ yNhsNh = 1
sNh = 1 (16) annulus; if yTh = 1, then the hot stream flows inside the inner
tube and the cold stream flows in the annulus.
where yPtsE and yPasE are the binary variables which represent One may associate known parameters of the hot and cold
the integer options of the number of units in parallel per branch streams, such as mass flow rates ( mh ̂ and mc )̂ , for instance, to the
for the tube-side and the annulus-side streams, and pNE sÊ is the corresponding variables for the tube-side and annulus-side
sequence of integer numbers representing the possible numbers flows:
of units interconnected along a branch (1, 2, ..., sEmax). For
mt = mcyT
̂ c + mh yT
̂ h (19)
branch structure type I (Figure 4i-a) the number of parallel units
selected for both flow-sides (NPt and NPa) are equal to unity ma = mcyT
̂ h + mh yT
̂ c (20)
(yPtsE=1 = yPasE=1 = 1). Similarly, yNhsNh is the binary variable
that represent the integer options of the number of hairpins per Thus, if the cold stream flows inside the inner tube (yTc = 1),
unit, and pNh sNh ̂ is the sequence of integer numbers then mt = mc ̂ and ma = mh .̂ Analogously, the following
equations relate physical properties and fouling factors of the
representing the number of hairpins possible per unit (1, 2, ..., hot and cold streams to the corresponding values of the tube-
sNhmax). side and annulus-side flows:
According to the structural options shown in Figure 4, an
additional constraint must be included to ensure that if the tube- ρt = ρĉ yTc + ρĥ yTh (21)
side has already more than one parallel passage, the annular side
can be only arranged in series and vice versa: ρa = ρĉ yTh + ρĥ yTc (22)
yPtsE = 1 + yPasE = 1 ≥ 1 (17) ̂ + Cp ̂ yTh
Cpt = CpyT
c c h (23)
To force that the outer tube inner diameter is larger than the
̂ + Cp ̂ yTc
Cpa = CpyT (24)
inner tube outer diameter one writes: c h h
sD max sd max μt = μĉ yTc + μ ̂h yTh (25)
∑ pDti sD̂ yDsD ≥ ∑ pdte sd̂ ydsd + ε
sD = 1 sd = 1 (18) μa = μĉ yTh + μ ̂h yTc (26)
where ε is the smallest diameter difference. Alternatively, one
can construct the model where each equation containing both kt = kĉ yTc + k ĥ yTh (27)
diameters is written only for combinations (sd,sD)* that are
ka = kĉ yTh + k ĥ yTc (28)
allowed. This latter option is favored when sets can be used in
programming environments such as GAMS. ̂ + Rf yT
̂
Rft = RfcyT c h h (29)
The utilization of the proposed set of binary variables to
describe the interconnection structure of double pipe heat ̂ + Rf yT
̂
exchanger elements is illustrated by the example depicted in Rfa = RfcyTh h c (30)
Figure 6. According to the proposed approach, the nonzero yTc + yTh = 1 (31)
binary variables that describe the structure presented in this
figure are yBsB=3 = 1 (three parallel branches), yPtsE=1 = 1 and where m is the mass flow rate, ρ is the density, Cp is the heat
yPasE=5 = 1 (five units per branch aligned in series for the tube- capacity, μ is the viscosity, k is the thermal conductivity, Rf is the
side flow and in parallel for the annulus-side flow) and yNhsNh=2 fouling factor, and the subscripts c and h indicate the cold and
= 1 (two hairpins per branch). hot streams, respectively.
3.2. Stream Allocation. The stream allocation is controlled 3.3. Structural Constraints. The flow area and length of the
by the binary variables yTc and yTh. If yTc = 1, then the cold hydraulic path of the streams considering the complete double
stream flows inside the inner tube and the hot stream flows in the pipe heat exchanger structure depend on the selection of the
12084 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

tube diameters and length, the number of hairpins and units in

series and in parallel in each branch, and the number of parallel
branches:
ij πdti 2 yz
At = jjj zzNB NPt
z
k 4 { (32)

ij πDti 2 πdte 2 yzz

Aa = jjj zNB NPa
k 4 4 z{
−
(33)

Lu = Nh Lh (34)

Lt = Lu NPa (35)
Figure 7. Inner tube friction factor profile in all regimes generated from
La = Lu NPt (36) eqs 41−43.

where At and Aa are the tube-side and annulus-side flow area, Lu

is the total length of one unit and Lt and La are the
corresponding flow path lengths for the streams flowing in the Since the laminar flow is more affected by the entry region, more
inner tube and in the annulus, respectively. than one equation is utilized, according to the proposal of
3.4. Inner Tube Side Thermal and Hydraulic Modeling. Incropera et al.16 For Prt > 5, the Hausen correlation is used:
The Prandtl number is given by 0.0668(2dti /Lh)Ret Prt
Nut Hau = 3.66 +
Cpt μt 1 + 0.04((2dti /Lh)Ret Prt )2/3
Prt =
kt (37)
for Ret ≤ 2300; Prt > 5 (45)
The flow velocity and the corresponding Reynolds number are For Prt ≤ 5, the Nusselt number is specified by the Sieder and
given by Tate (S&T) correlation unless its given value is lower than the
(mt /ρt ) theoretical (theo) Nusselt number for fully developed flow
vt = (3.66), in which case the latter is applied:
i Ret Prt dti yz
Nut S&T = 1.86jjjj
At (38)

zz
z
1/3

k Lh/2 {
dti vt ρt
Ret =
μt (39)
for Ret ≤ 2300, Prt ≤ 5, Nut S&T ≥ 3.66 (46)
Ignoring the minor head losses in the connections and bends,
the pressure drop of the flow in the inner tube is calculated by Nut theo = 3.66 for Ret ≤ 2300, Prt ≤ 5, Nut S&T < 3.66
the Darcy−Weisbach equation (omitting the viscosity correc- (47)
tion factor):15
Without loss of generality, the viscosity correction factor in the
Lt vt 2 Sieder and Tate correlation in eq 46 has also been omitted (i.e.,
ΔPt = ρt ft the ratio between bulk and wall viscosities were considered equal
dti 2 (40)
to 1).
where f t is the Darcy friction factor. To formulate a set of constraints able to represent the friction
The friction factor depends on the flow regime as follows:15 factor and Nusselt number evaluation for all possible conditions
64 described in eqs 41 to 47, binary variables are associated with the
ft lam = for Ret ≤ 1311 possible ranges of Ret, Prt, and Nut, as indicated in Figure 8. In
Ret (41)
this figure, URe , ̂ UPr , ̂ and UNu are
̂ maximum values of Re, Pr,
ft tran = 0.0488 for 1311 < Ret ≤ 3380 (42) and Nu, respectively. In each interval, the corresponding binary
variable is equal to 1 and the others are 0.
1.056 The following equations relate the binary variables and their
ft turb = 0.014 + for Ret > 3380 corresponding ranges:
Ret 0.42 (43)
̂
Ret ≤ 1311yRet1 + 2300yRet 2 + 3380yRet3 + URe yRet
The values of the intervals of the different flow regimes in eqs 41 4
to 43 guarantee a continuous profile of the friction factor, as (48)
shown in Figure 7. Ret ≥ 1311yRet 2 + 2300yRet3 + 3380yRet4 + ε (49)
Regarding the Nusselt number; laminar, transitional, and
turbulent flows are considered, with the threshold of Ret = 2300 ̂
Prt ≤ 5yPrt1 + UPr yPrt 2 (50)
between the laminar and the transitional/turbulent flow. For the
transitional and turbulent flow, the Gnielinski correlation is Prt ≥ 5yPrt 2 + ε (51)
used:16
Nut S&T ≤ 3.66yNut1 + UNu yNut
̂ 2 − ε (52)
Gni (ft /8)(Ret − 1000)Prt
Nut = for Ret > 2300 Nut S&T ≥ 3.66yNut 2 (53)
1 + 12.7(ft /8)1/2 (Prt 2/3 − 1)
(44) where ε is a small positive number.
12085 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

The annular region friction factor, analogously to the inner

tube, depends on the flow regime according to the following
equations:15
64
fa lam = for Rea ≤ 500
Rea (64)

32.656
fa tran = 0.02696 + for 500 < Rea ≤ 10000
Rea 0.93
(65)
0.178
fa turb = for Rea > 10000
Rea 0.1865 (66)
The threshold Reynolds number values in this case are 500 and
Figure 8. Possible ranges for Ret, Prt, Nut S&T
and its corresponding 10000 and these equations also form a continuous profile for
binary variables. calculation of the friction factor, as shown in Figure 9.

Since only one binary variable must be selected for each set of
intervals, it yields
sRet max
∑ yRetsRet = 1
sRet = 1 (54)
yPrt1 + yPrt 2 = 1 (55)
yNut1 + yNut 2 = 1 (56)
Therefore, the friction factor and the Nusselt number are
represented by
ft = ft lamyRet1 + ft tran(yRet 2 + yRet3) + ft turbyRet4 (57) Figure 9. Annular friction factor profile in all three regimes generated
from eqs 64−66.
Nut = Nut theo(yRet1 + yRet 2)yPrt1 yNut1
Regarding the Nusselt number, the same ranges and
+ Nut S&T(yRet1 + yRet 2)yPrt1 yNut 2 correlations used for the tube-side are applied, replacing the
+ Nut Hau(yRet1 + yRet 2)yPrt 2 inner tube diameter by the hydraulic diameter:

+ Nut Gni(yRet3 + yRet4) (58) Nua theo = 3.66 for Rea ≤ 2300, Pra ≤ 5, and NuaS&T
3.5. Annulus Side Thermal and Hydraulic Modeling. < 3.66 (67)

i Rea Pra dh zy
NuaS&T = 1.86jjjj zz
z
The Prandtl number is 1/3

k Lh/2 {
Cpa μa
Pra =
ka (59)
for Rea ≤ 2300, Pra ≤ 5 and NuaS&T ≥ 3.66 (68)
The flow velocity inside the annular region is given by
(ma /ρa) 0.0668(2dh/Lh)Rea Pra
va = Nua Hau = 3.66 +
Aa (60) 1 + 0.04((2dh /Lh)Rea Pra)2/3
The hydraulic diameter (four times the flow cross-sectional area for Rea ≤ 2300 and Pra > 5 (69)
divided by the wetted perimeter) can be simplified to16
(fa/8)(Rea − 1000)Pra
dh = Dti − dte (61) NuaGni =
1 + 12.7(fa/8)1/2 (Pra 2/3 − 1)
The Reynolds number is
for Rea > 2300 (70)
dh va ρa
Rea = Binary variables are then included to describe each interval of
μa (62)
Rea, Pra, and NuaS&T, as depicted in Figure 10. The same
Ignoring the head losses in the connections and bends, the approach used for the tube-side flow for the evaluation of the
pressure drop of the flow in the annulus is given by the Darcy− Nusselt number is implemented here as follows:
Weisbach equation using the hydraulic diameter (also omitting ̂
Rea ≤ 500yRea1 + 2300yRea 2 + 10000yRea3 + URe yRea
the viscosity correction factor):15 4
(71)
Lt va 2 Rea ≥ 500yRea 2 + 2300yRea3 + 10000yRea4 + ε
ΔPa = ρa fa (72)
dh 2 (63)
̂
Pra ≤ 5yPra1 + UPr yPra
where fa is the Darcy friction factor for the annular flow. 2 (73)

12086 DOI: 10.1021/acs.iecr.9b01536

Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

(Tih ̂ − Toc )̂ − (To h ̂− Tic)̂

ΔTlm =̂
̂ Toc )̂
ln
(Tih −
̂( ̂
(To h − Tic) ) (86)

where Tih ̂ and To h are

̂ the inlet and outlet temperatures of the
̂ ̂ the inlet and outlet temperatures
hot stream, and Tic and Toc are
of the cold stream.
As discussed above, the correction factor depends on the
structure of the arrangement of the double pipe heat exchanger
units. If the structure is based on an arrangement where both
streams are aligned in series, (type I, as shown in Figure 4i-a), the
correction factor is equal to unity. If the structure has one of the
streams in parallel, as in branch structures of type II (Figure 4ii-a
Figure 10. Possible ranges for Rea, Pra, NuaS&T and their and type III (Figure 4iii-a), its value depends on which stream is
corresponding binary variables. in parallel and on the number of existing units per branch.13 As
the number of units in parallel increases the correction factor
Pra ≥ 5yPra 2 + ε (74) decreases, accounting for the deviation from a countercurrent
flow pattern. The number of hairpins per unit and the number of
NuaS&T ≤ 3.66yNua1 + UNu yNua
̂ 2 − ε (75) parallel branches in the general structure do not implicate any
flow pattern change between streams and therefore do not have
NuaS&T ≥ 3.66yNua 2 (76)
any influence on the correction factor value. All the correction
factor options are calculated as parameters prior to the
sRea max optimization. For stream sST (cold, sST = c, or hot, sST = h)
∑ yReasRea = 1 aligned in series and the other stream in parallel, the correction
(77) factor is given by13
i 1 − pP ̂ yz
sRea = 1

lnjjj 1 − pP ̂ sST ̂ z
z
k sST pR sST {
yPra1 + yPra 2 = 1 (78) ̂ − pNE ̂ )
ij pNE sÊ y
(pR sST
zz
lnjjjj
sE

z
̂
pF sST =

pR ̂ z z
, sE
pNE ̂ (pR ̂ − 1)
k {
̂ − pNE ̂
pR sST
yNua1 + yNua 2 = 1 (79) sE sST sE
+
̂ (1 − pP ̂ pR ̂ )1/ pNEsÊ
pR sST sST
sST sST

fa = fa lam
yRea1 + fa tran
(yRea 2 + yRea3) + fa turb
yRea4 for sE ≠ 1 (87)
(80) ̂ and pP ̂ are specified as
l
where the factors pR sST
o
o
sST

o
o
o
o
Nua = Nua theo(yRea1 + yRea 2)yPra1 yNua1
o
(Toc ̂ − Tic)̂
o (Tih ̂ − To ĥ )
=m
for sST = c

o
o
+ NuaS&T(yRea1 + yRea 2)yPra1 yNua 2
o
o
o (Tih − To h )
̂
o
pR sST

o
o
+ Nua Hau(yRea1 + yRea 2)yPra 2
o
̂ ̂

n (Toc − Tic)
for sST = h
+ NuaGni(yRea3 + yRea4) (81) ̂ ̂ (88)

l
o
o
o
o
o
3.6. Heat Transfer Coefficients. The convective heat
o
o
(Tih ̂ − To h )̂
o (Tih ̂ − Tic)̂
transfer coefficient for the tube-side and annulus-side flows are for sST = c
=m
o
o
o
given by

o
o (Toc − Tic)
pP ̂
o
o
o
sST

o
Nut kt ̂ ̂

n (Tih − Tic)
ht = for sST = h
dti (82) ̂ ̂ (89)
Nua ka On the basis of these expressions, the constraint that represents
ha =
dh (83) the correction factor (F) evaluation, dependent on the structure
selected, becomes
The overall heat transfer coefficient is then determined by
sE max sE max
1 F=1+ ∑ ∑ {yTc[yPtsE(pFh ,̂ sE − 1)
U=
1 dte dte dte ln ( dtedti ) 1 sE = 2 sE ′= 2
+ Rft dti + + Rfa + (84)
ht dti 2ktubê ha
+ yPasE (pF ĉ , sE − 1)] + yTh[yPtsE(pF ĉ , sE − 1)
′ ′
where ktubê is the thermal conductivity of the inner tube.
+ yPasE (pFh ,̂ sE − 1)]} (90)
3.7. Heat Transfer Rate. The heat transfer rate is given by ′ ′
Q̂ = UA req ΔTlm F̂ In turn, aiming at guaranteeing a design margin, a minimum area
(85) excess (Â exc) is imposed.
ij  yz
A ≥ jjj1 + exc zzzA req
j 100 z{
where Areq is the required heat transfer area, ΔTlm is ̂ the

k
logarithmic mean temperature, and F is the correction factor.
The logarithmic mean temperature is defined as (91)

12087 DOI: 10.1021/acs.iecr.9b01536

Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

where the equipment heat transfer area is given by wi , j ,..., k = ypi yqj ... yzk (101)
A = π dte Lu NB NPt NPa (92) where this equality is guaranteed by the inclusion of these
Therefore, the heat transfer rate shown in eq 85 can be constraints:
reorganized as wi , j ,..., k ≤ ypi (102)
Aexĉ

UA ≥
Q̂ ( 100 )
+1 wi , j ,..., k ≤ yqj (103)
ΔTlm F̂ (93) ···
3.8. Pressure Drop and Velocity Bounds. The lower and wi , j ,..., k ≤ yzk (104)
upper bounds on the velocities of the tube-side flow and
annulus-side flow are given by wi , j ,..., k ≥ ypi + yqj + ... + yzk − (m − 1) (105)
̂
vt ≥ vtmin (94) where m is the number of binary variables in the product.
The reformulation of the product between a binary variable
̂
vt ≤ vtmax (95) and a function of continuous variables also involves the
introduction of an additional continuous variable and a set of
va ≥ vamin
̂ (96) inequality constraints.17
Let f(x1, x2, ...) be a function of continuous variables such that
va ≤ vamax
̂ (97) 0 ≤ f ≤ Φ̂ in the problem domain and y is a binary variable. The
While the pressure drop bounds are represented by product of this function and the binary variable can be
substituted by a continuous nonnegative variable w together
̂ yT + ΔP ̂ yT
ΔPt ≤ ΔPc disp with the inclusion of these constraints:
c h disp h (98)
w − Φ̂ y ≤ 0 (106)
̂ yT + ΔP ̂ yT
ΔPa ≤ ΔPc disp h h disp c (99)
f (x1 , x 2 , ...) − w − Φ̂(1 − y) ≤ 0 (107)
3.9. Objective Function. The objective function is given by
the minimization of the heat transfer area: f (x1 , x 2 , ...) − w ≥ 0 (108)
min A + ph ̂ Nh (100) 4.2. Resultant Formulation. The application of the
approach described above modifies the following set of
where ph is a penalty factor (a random small value) associated constraints: inner tube side thermal and hydraulic modeling,
with the number of hairpins directing the optimization, in case of annulus side thermal and hydraulic modeling, and heat transfer
equivalent solutions, to the one with the smaller number of rate. The detailed description of each equation modified is
elements. described in the Supporting Information.
The above model is here called the raw MINLP. It uses a
discrete representation of the geometric variables and several 5. HEURISTIC APPROACH
nonlinear relationships involving continuous variables. Addi-
tionally, some equations contain nonlinearities involving binary To later compare our MINLP approach to some heuristic
variables (eqs 57, 58, 80, 81, and 90), that are incompatible with approach, a proposal of a heuristic design procedure is presented
the use of some optimization algorithms. here. Unlike for other equipment (i.e., distillation trays) there is
To make the model more amenable to be solved by algorithms no step-by-step heuristic procedure widely available in the
that require linearity in binary variables, a reformulation to literature (i.e., the identification of a feasible solution depends
obtain another mixed-integer nonlinear model linear in the entirely on the intervention of the designer using a trial and
binaries is also presented, resulting in what is here called binary- verification method). Serth,13 for example applies a trial-and-
linear MINLP. viability-verification without outlining any rule. Thus, his
example is used here as a simple source to propose one such
4. MINLP REFORMULATION procedure. The proposed procedure is as follows:
This section presents the reformulation of some constraints to • Assume a starting structure, as simple as possible. A good
eliminate the nonlinearities involving binary variables (thus candidate is a one-unit branch structure as in Figure 4i-a.
rendering the binary-linear MINLP model). The linear • Pick hairpin dimensions from the discrete available
constraints or the constraints with continuous-only non- options (diameters and length).
linearities remain the same. • Calculate the heat transfer coefficient, the required area
4.1. Reformulation Techniques. In the original optimiza- (which together with a reasonable excess area, e.g.,
tion problem (raw MINLP model), there are two types of minimum of 10%, gives the necessary number of hairpins
nonlinear terms involving binaries: the product of two or more for the selected structure), and pressure drops.
binary variables and the product of a binary variable and a • If the results meet the design requirements (e.g., velocities
function of continuous variables. and pressure drops according to acceptable values) the
The product of binary variables can be substituted by a procedure stops.
continuous variable and a set of linear inequalities, as follows.17 • Otherwise, a structure change is made, a step that is highly
Let ypi, yqj, ..., and yzk be a set of binary variables. The product dependent on the designer, who, according to the unmet
of these variables can be substituted by a continuous requirement, can (a) change fluid allocation, (b) change
nonnegative variable wi,j,···,k: the hairpin dimensions (diameters and length), (c) select
12088 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

more or less branches, or (d) select another branch

structure (Figure 4).
Quite clearly, some changes, such as dimensions, are the
obvious first target: for example, if the area is smaller than the
area required, and the pressure drop is lower than the limit,
increasing length may be the first thing to consider. If pressure
drop is the limiting factor, the obvious dimension to modify is
the diameter, or the number of branches. The analysis of all the
starting options and the decisions to make when the service
requirements are not met, which would be in the direction of
Figure 11. Example 1, heat exchanger structure obtained by Serth.13
creating a rule-based automatic algorithm, is not pursued here.

6. RESULTS
Three aspects are explored in the analysis of the results: The results explored in sections 6.1 and 6.2 are generated by
our raw MINLP formulation, which is then compared with our
Table 1. Initial Estimates for Key Binary Variables linear-binary MINLP approach in section 6.3. All numerical
results were obtained using the GAMS software version 23.7.3
variable initial estimate and a branch-and-bound solver (SBB). Without good initial
inner tube diameter selection yd3 = 1 estimates, the SBB solver used does not converge for neither
stream allocation yTc = 1 MINLP models. The result of the root node relaxation of the
range of Reynolds identification yRet4 = yRea4 = 1 SBB solver is highly sensitive to the initial estimate. Thus, it was
range of Nusselt identification yNut2 = yNua2 = 1 necessary to identify a suitable initial set of values for some of the
problem variables to have convergence. GAMS default initial
Table 2. Initial Estimates for Key Continuous Variables value for all variables is zero, unless zero is not within the
bounded range (in which case the default is the bound closest to
variable initial estimate
zero) or the user determines otherwise. Six binary variables were
length of one unit Lhlo + Lh up
identified related to tube diameter, stream allocation, and range
Lh = 2
Reynolds numbera Rex lo + Rex up of Reynolds numbers that, in our examples, need initial values to
Rex = 2 promote the convergence. The values shown in Table 1 are the
Nusselt numbera Nux =
Nux lo + Nux up
ones found to work well for the set of proposed examples for this
2
Seider and Tate Nusselt numbera S&T + Nux S&T
Nux lo
set of variables. Furthermore, it was found that when using SBB,
S&T up
Nux = 2
the convergence also depends on the initial estimate for some
Hausen Nusselt number Hau
Nut = 3.66 continuous variables, as shown in Table 2. Finally, DICOPT
NuaHau = 3.66 does not provide results for any of the two MINLP approaches,
a
x being t (tube side) or a (annulus side). no matter what initial values are provided. This issue on
initialization is not investigated further, because it is not the aim
Table 3. Example 1: Stream Data of the current paper to seek a general systematic route to find
robust and automatic initialization procedures. Rather, all the
cold stream hot stream difficulties associated with the use of local solvers are simply
parameter unit (benzene) (aniline)
pointed out and these are difficulties known to many.
mass flow, m̂ sST kg/s 1.26 1.22 6.1. Comparison with Literature Results. Example 1 is a
̂
inlet temp, TisST °C 15.55 65.55 design problem proposed by Serth,13 which consists of
̂
outlet temp, TosST °C 48.85 37.75 determining the number of required hairpins and their
density, ρ̂sST kg/m3 879 1022 arrangement for the service shown in Table 3.
viscosity, μ̂sST Pa·s 5.5 × 10−4 2 × 10−3 The design methodology employed by Serth13 consists of a
heat capacity, Cp ̂ J/(kg °C) 1758.46 2177.14 trial-and-viability-verification procedure, similar to the one
sST
shown in section 5. Serth13 does not provide rules or hints of
thermal conductivity k̂sST W/(m °C) 0.159 0.173
how to obtain a different design and stops its procedure when
fouling resistance, Rf ̂ sST
m2 °C/W 1.76 × 10−4 1.76 × 10−4 the first viable equipment is found.
available pressure drop, The solution obtained by Serth13 is illustrated in Figure 11
ΔP ̂ kPa 138 138
sSTdisp and consists in a single branch of type II structure (Figure 4ii-a),
with an inner pipe stream connected in series and annulus side
i. comparison of the mathematical programming solution stream aligned in parallel in two units, each containing five
with a result from the literature hairpins using pipes with a thermal conductivity of 16.27 W/(m
°C).
ii. comparison of the mathematical programming solution The thermo-fluid dynamic relations used by Serth13 are
with a solution generated by the procedure explained in
different from the model employed in this paper, particularly the
section 5
friction factor and the Nusselt number correlations. For
iii. comparison between the proposed MINLP approaches comparison, Serth’s correlations are used in the optimization
Complementarily, the Supporting Information material applied to this example. Details of the resultant formulation
contains additional examples that explore the flexibility of the using Serth′s correlations employed in this example are shown in
proposed approach in relation to different flow regimes and the Supporting Information material. In all other examples, the
structural alternatives. correlations presented in section 3 are employed.
12089 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Table 4. Structural Discrete Options (Pipe Schedule 40)

parametera units discrete options
pNB sB̂ 1 to 20
pNE sÊ 1 to 20
pLĥ sLh ft 5 10 15 20 25 32b
pLĥ sLh m 1.524 3.048 4.572 6.096 7.620 9.75b
1 3 1 1
inner tube NPS in /2 /4 1 1 /4 1 /2 2 2 1/2 3 3 1/2
inner tube OD ( pdte sd̂ ) m 0.021 0.027 0.033 0.042 0.048 0.060 0.073 0.089 0.102
1 1 1 1 1
outer tube NPS in 1 /4 1 /2 2 2 /2 3 3 /2 4 4 /2 5
outer tube OD ( pDte sD̂ ) m 0.042 0.048 0.060 0.073 0.089 0.102 0.114 0.127 0.168
a
NPS = nominal pipe size, OD = outer diameter. bOption used only in section 6.1.

Table 5. Example 1: Results

raw MINLP using Serth correlations
a
variable units Serth 13
fixed geometry from Serth13 optimal solution
tube side stream aniline aniline benzene
total heat exchanger area, A m2 12.91 12.92 9.96
required heat exchanger area m2 11.68 11.55 9.02
inner tube NPS−OD in−m 11/4−0.042 11/4−0.042 3
/4−0.027
outer tube NPS−OD in−m 2−0.060 2−0.060 11/4−0.042
hairpin length, Lh ft−m 32−9.75 32−9.75 10−3.048
number of hairpins per units, Nh 5 5 13
number of branches, NB 1 1 3
tube side units in parallel, NPt 1 1 1
annular side units in parallel, NPa 2 2 1
tube side velocity, vt m/s 1.24 1.24 1.389
annulus side velocity, va m/s 0.93 0.93 0.979
tube side film coefficient, ht W/(m2 °C) 999 996 1727
annulus side film coefficient, ha W/(m2 °C) 1448 1446 977
overall heat transfer coefficient, U W/(m2 °C) 391.8 395.9 423.8
correction factor, F 0.836 0.836 1
tube side pressure drop, dPt kPa 82.0 78.9 49.0
annulus side pressure drop, dPa kPa 70.3 71.1 123.5
a
NPS = nominal pipe size, OD = outer diameter.

Table 6. Example 2: Stream Data

parameter unit cold stream hot stream
mass flow, m̂ sST kg/s 2.52 2.11
̂
inlet temperature, TisST °C 20 60
̂
outlet temperature, TosST °C 30 50
density, ρ̂sST kg/m3 850 1000
viscosity, μ̂ sST Pa·s 5.5 × 10−4 2 × 10−4
heat capacity, Cp ̂ sST
J/(kg °C) 1760 2100
thermal conductivity k̂sST W/(m °C) 0.160 0.175
−4
fouling resistance, Rf ̂ sST
m °C/W
2
2 × 10 2 × 10−4

Figure 12. Example 1: Optimal solution obtained by our raw MINLP

(with Serth correlations).

Additionally, it is also important to mention that the heat

exchanger modeling employed in the optimization does not
consider the pressure drops in the nozzles and return bends, and
therefore these elements were excluded from the analysis. The
results of pressure drop reported here from Serth13 also excluded
the head loss in these elements. An analysis of the trial-and-
verification procedure adopted by Serth13 indicates that the Figure 13. Example 2: Trial and verification procedure solution.
exclusion of these minor losses during the sequence of design
12090 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

structure (only the structural binary variables were fixed),

therefore validating the model implemented. The small
difference in the pressure drops were expected since Serth13
accounts for the viscosity correction factor which is here omitted
in eqs 40 and 63. Our optimization approach obtained an
optimal result (illustrated in Figure 12) that shows a reduction of
23% in area, which is attained due to a combination of higher
LMTD correction factor F resulting from the counter-current
arrangement of the set of units of the proposed heat exchanger
and a higher heat transfer coefficient resulting from a substantial
improvement of the annulus heat transfer film coefficient, in turn
enhanced by a higher velocity.
6.2. Comparison with Heuristic Approaches. The
advantages of an MINLP approach over a heuristic trial-and-
verification procedure are illustrated in example 1, originally
solved by Serth.13 For example 2, the stream data are depicted in
Table 6.
Figure 14. Example 2: Optimal solution obtained by our raw MINLP The discrete options applied to this design problem are shown
approach. in Table 4, where a thermal conductivity of 55 W/(m °C) is used
for the material of the pipes, flow velocity in both sides is
steps would not modify the result obtained, which enables a between 1 and 3 m/s, available pressure drop for both streams is
direct comparison of the results. equal to 50 kPa, and there is a minimum excess area of 20%.
The optimization search space employed in the proposed Heuristics shown in section 5 are used to obtain a viable
analysis is depicted in Table 4, and includes the original design solution, which is illustrated in Figure 13. The Supporting
proposed by Serth.13 The velocity bounds applied in the Information material depicts the steps taken during the
optimization runs were 0.9 to 3 m/s and a minimum excess area procedure.
of 10% (same as used by Serth13). Also, for comparison However, when using the raw MINLP formulation, one gets
purposes, the initial estimate was the first heat exchanger an area reduction of 17.8%, with the structure illustrated in
attempted by Serth in his trial-and-verification procedure, which Figure 14. This area reduction is achieved by using higher flow
consisted of a one countercurrent unit branch (structure on velocities compatible with available pressure drops (Table 7).
Figure 4i-a) with six hairpins (of the same dimensions as Serth’s 6.3. Comparison of the Performance of the MINLP
final solution) and benzene in the inner pipe. Table 5 contains Models. The two MINLP approaches are compared in this
the reported values by Serth,14 the results obtained using the raw section. For examples 1 and 2, the optimization using the linear-
MINLP model modified to use Serth′s correlations using the binary MINLP formulation results in the same solutions as the
same geometry, and the corresponding optimal solution. The ones from the raw MINLP formulation, already reported in the
diameters for both inner and outer tubes are calculated previous subsections. However, this is not a rule for every
considering pipe schedule 40. possible scenario, as illustrated here with examples 3 and 4,
The formulation proposed in this paper was able to find where both MINLP approaches give different results. The
equivalent results from Serth13 for the same heat exchanger stream data for both examples are shown in Table 8. The

Table 7. Example 2: Design Results

variablea unit trial and verification raw MINLP
tube side stream hot stream cold stream
total heat exchanger area, A m2 2.24 1.84
required heat exchanger area m2 1.79 1.51
inner tube NPS−OD in−m 1−0.033 1
/2−0.021
outer tube NPS−OD in−m 2−0.060 11/4−0.042
hairpin length, Lh ft−m 5−1.52 15−4.572
number of hairpins per unit, Nh 7 1
number of branches, NB 2 2
tube side units in parallel, NPt 1 3
annular side units in parallel, NPa 1 1
tube side velocity, vt m/s 1.89 2.52
annulus side velocity, va m/s 1.15 1.74
tube side film coefficient, ht W/(m2°C) 6496 4292
annulus side film coefficient, ha W/(m2°C) 1995 6046
overall heat transfer coefficient, U W/(m2°C) 824 991.6
correction factor, F 1 0.989
tube side pressure drop, dPt kPa 14.1 19.0
annulus side pressure drop, dPa kPa 8.0 30.3

a
NPS = nominal pipe size, OD = outer diameter.

12091 DOI: 10.1021/acs.iecr.9b01536

Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Table 8. Examples 3 and 4: Stream Data

example 3 example 4
parameter unit hot stream cold stream hot stream cold stream
mass flow, m̂ sST kg/s 16.5 34.32 25.9 13.14
̂
inlet temperature, TisST °C 90 30.1 60 10
̂
outlet temperature, TosST °C 50 40 50 25
density, ρ̂sST kg/m3 786 995 780 1050
viscosity, μ̂ sST Pa·s 1.89 × 10−4 7.2 × 10−4 9.5 × 10−4 0.024
heat capacity, Cp ̂
sST
J/(kg °C) 2177 4187 1900 2500
thermal conductivity k̂sST W/(m °C) 0.12 0.59 0.18 0.264
fouling resistance, Rf ̂
sST
m2 °C/W 2 × 10−4 4 × 10−4 3 × 10−4 3 × 10−4
̂
available pressure drop, ΔPsSTdisp kPa 100 100 150 150

Table 9. Example 3: Optimization Results

variable unit raw MINLP linear-binary MINLP
tube side stream hot stream hot stream
total heat exchanger area, A m2 88.73 91.15
required heat exchanger area m2 73.94 75.19
inner tube NPS−OD in−m 1 1/2−0.048 1−0.033
outer tube NPS−OD in−m 2 1/2−0.073 2−0.060
hairpin length, Lh ft−m 20−6.096 25−7.620
number of hairpins per unit, Nh 6 6
number of branches, NB 8 19
tube side units in parallel, NPt 1 1
annular side units in parallel, NPa 2 1
tube side velocity, vt m/s 2.00 1.98
annulus side velocity, va m/s 1.71 1.41
tube side film coefficient, ht W/(m2 °C) 1397 1436
annulus side film coefficient, ha W/(m2 °C) 9129 7449
overall heat transfer coefficient, U W/(m2 °C) 601.3 579.1
correction factor, F 0.979 1
tube side pressure drop, dPt kPa 76.3 79.0
annulus side pressure drop, dPa kPa 93.7 59.1

Table 10. Example 4 − Optimization Results

variable unit raw MINLP linear-binary MINLP
tube side stream cold stream cold stream
total heat exchanger area, A m2 64.60 40.86
required heat exchanger area m2 55.92 37.01
inner tube NPS−OD in−m 1 1/4 − 0.042 3 1/2 − 0.102
outer tube NPS−OD in−m 2 1/2 − 0.073 4 1/2 − 0.127
hairpin length, Lh ft−m 25−7.620 10−3.048
number of hairpins per unit, Nh 4 7
number of branches, NB 8 1
tube side units in parallel, NPt 1 1
annular side units in parallel, NPa 2 6
tube side velocity, vt m/s 1.62 1.96
annulus side velocity, va m/s 1.23 2.54
tube side film coefficient, ht W/(m2 °C) 420 656
annulus side film coefficient, ha W/(m2 °C) 1613 3276
overall heat transfer coefficient, U W/(m2 °C) 237.2 360.6
correction factor, F 0.992 0.986
tube side pressure drop, dPt kPa 117.1 110.7
annulus side pressure drop, dPa kPa 24.3 110.8

hairpins are built with a material with a thermal conductivity of example 4. The results obtained for example 3 and example 4, for
55 W/(m°C). The additional design specifications include flow both MINLP approaches, are depicted in Tables 9 and 10, and
velocity in both sides between 1 and 3 m/s for both examples, Figures 15 and 16, respectively. In the case of Table 9, there is a
and a minimum excess area of 20% for example 3, and of 10% for large difference between the number of branches associated with
12092 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Figure 15. Example 3: (i) raw MINLP solutions; (ii) linear-binary. (a) General structure; (b) branch structure.

different structures, but there is a consistency between them. For The computational times for both examples are shown in
the annular-flow side for example, in the solution obtained with Table 11, indicating that the linear-binary MINLP approach
the raw MINLP formulation there are eight branches (NB = 8), requires more computational effort, which also occurred on the
and in each branch there are two units in parallel (NPa = 2), first two examples where the heat transfer areas obtained were
while in the linear-binary MINLP solution, there are 19 branches the same for both formulations. These values were obtained
(NB = 19) but there are no units in parallel in each branch (NPa using a computer with a processor Intel Core i7 3.41 GHz with
= 1); therefore, in both cases the total flow for the annular side 16.6 GB RAM memory.
would be divided by 19. Since both inner and outer tubes had To reformulate the problem to obtain the linear-binary
their OD’s decreased from the first solution to the second, the MINLP, the problem variables and equations increase. This
annular flow area did not change significantly, and therefore, the appears to be directly connected to the larger computational
velocities are quite similar in both cases (2 and 1.98 m/s). Now, effort. For the given discrete available options given in examples
if one looks to the tube flow side, there are 8 branches for the raw 3 and 4, the raw MINLP has the number of 179 variables and
MINLP solution and 19 branches for the linear-binary solution, 120 constraints, compared to 283 and 440 of the linear-binary
and in both cases there are no units in parallel (NPt = 1). approach.
Therefore, the mass flow per unit in the second case would be The optimum values of the two MINLP approaches show
half the first one and a great decrease in the flow velocity would that, for a given example, there may be more than one optimal
be expected. However, there is also a decrease in the available solution. For example 3, the raw MINLP approach found a
flow area per unit, since the inner tube OD goes from 1 1/2 in to 1 solution with a heat transfer area only 2.65% smaller than the
in., compensating a portion of the decrease in mass flow, so that one found with the linear-binary MINLP approach. On the
the velocity only decreases from 1.71 to 1.41 m/s. contrary, for example 4 the smaller area was achieved using the
12093 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

Figure 16. Example 4: (i) raw MINLP solutions; (ii) linear-binary. (a) General structure; (b) branch structure.

Table 11. Examples 3 and 4: Performance Comparison of the initial estimates obtained from the solution of the SBB. The
Two MINLP Approaches results are shown in Table 12.
raw MINLP linear-binary MINLP For all three scenarios the global optimal solution was
example area (m2) elapsed time (s) area (m2) elapsed time (s) achieved for both examples 3 and 4. There was no significant
3 88.73 54.31 91.15 127.68 computational effort gained between scenarios (i) and (ii). In
4 64.60 26.15 40.86 102.24 scenario (iii), in which the solutions obtained from sBB runs
were employed as initial estimates, there is a significant
reduction in computational effort for the BARON run isolated.
linear-binary MINLP approach, with a significant reduction of Since it is necessary to generate the initial estimates first by a sBB
36.7% when compared to the solution obtained with the raw
MINLP result. The results do not show any pattern regarding optimization run, the computational effort shown in Table 12
which formulation is able to achieve a better solution, but the corresponds to the sum of both elapsed times. Nevertheless, in
wide difference between solutions in example 4 illustrates the two situations (example 3 with linear-binary approach and
importance of the utilization of a global optimization scheme to example 4 solved with raw MINLP approach), the gains are
address the design problem.
approximately 28% and 89%, respectively, showing that the
Thus, both MINLP approaches were also solved with
BARON considering (i) no initial estimates, (ii) the same initial methodology proposed here can also be applied as a
estimates used to run the problem using the solver SBB, and (iii) preprocessing method for global optimization procedures.

Table 12. Elapsed Time Comparison of BARON Solver with Different Initial Estimates

Elapsed time (s)

(ii) same initial estimates as the ones used (iii) results from sBB solver used as initial
(i) GAMS default initial estimates for sBB solver estimates
example area (m2) raw MINLP linear-binary MINLP raw MINLP linear-binary MINLP raw MINLP linear-binary MINLP
3 88.73 279.04 407.82 279.90 407.77 269.91 294.14
4 40.86 796.15 400.26 718.52 390.02 84.39 386.51

12094 DOI: 10.1021/acs.iecr.9b01536

Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

7. CONCLUSIONS k̂sST = Stream sST thermal conductivity (W/(m °C))

This paper presented the solution of the design problem of ktubê = Tube material’s thermal conductivity (W/m °C)
double pipe heat exchangers using mathematical programming. m̂ sST = Stream sST mass flow (kg/s)
The design variables are selected among discrete values, due to pdte sd̂ = Available inside tube external diameters (m)
their physical nature or available commercial alternatives.
Therefore, the problem is formulated as an MINLP problem, pDte ̂ = Available outside tube external diameters (m)
sD
which is presented here in two versions. The first approach (raw pdti sd̂ = Available inside tube internal diameters (m)
MINLP) is based on the original nonlinear equations and the
pDti ̂ = Available outside tube internal diameters (m)
second approach (linear-binary MINLP) involves mathematical sD
transformations that eliminate the nonlinearities in the binary ̂
pF sST , sE
= F factor in case stream sST is in series, for sE ≠ 1
variables, rendering a linear model in those. A comparison with ̂
ph = Objective function penalty factor
the literature shows that our MINLP approach can obtain a
solution superior to that of trial and verification. The application pP ̂ = P parameter for the calculation of F factor if the
sST
of the original and modified formulations to a sample of design stream sST is in series
problems indicated that the latter approach demands more ̂ = R parameter for the calculation of F factor if the
pR sST
computational time. The analysis of the heat transfer area of the stream sST is in series
solutions indicated the presence of multiple locally optimal ̂ = Available hairpin lengths (m)
solutions. The large difference between the objective functions pLh sLh
in a pair of local optima, resultant of the nonconvexity of the pNB ̂ = Number of available branches
sB
MINLP, illustrates the importance of the investigation of
pNE sÊ = Number of available heat exchangers in series by
alternatives that can identify rigorously the global optimum.

■
branch
ASSOCIATED CONTENT pNh sNh ̂ = Number of available hairpins per unit
*
S Supporting Information ̂
Q = Heat-transfer rate (W)
The Supporting Information is available free of charge on the RfsST̂ = Stream sST fouling resistance (m2 °C/W)
ACS Publications website at DOI: 10.1021/acs.iecr.9b01536.
TisST ̂ = Stream sST inlet temperature (°C)
Equations resultant from the MINLP reformulation
TosST ̂ = Stream sST outlet temperature (°C)
(linear-binary MINLP); formulation using Serth’s
correlations employed in example 1; steps of the heuristic UNu =̂ Nusselt number upper limit
procedure applied to example 2; additional examples that UPr =̂ Prandtl number upper limit
explore the flexibility of the proposed model (PDF) URe =̂ Reynolds number upper limit

■ AUTHOR INFORMATION
Corresponding Author
vamax
vamin
̂ = Annulus-side maximum velocity (m/s)
̂ = Annulus-side minimum velocity (m/s)
̂ = Tube-side maximum velocity (m/s)
vtmax
*E-mail: [email protected]. ̂ = Tube-side minimum velocity (m/s)
vtmin
ORCID
André L. H. Costa: 0000-0001-9167-8754 Continuous variables
Miguel J. Bagajewicz: 0000-0003-2195-0833 ΔPa = Annulus-side pressure drop (Pa)
ΔPt = Tube-side pressure drop (Pa)
Notes
μa = Annulus-side stream viscosity (Pa·s)
The authors declare no competing financial interest.

■
μt = Tube-side stream viscosity (Pa·s)
ACKNOWLEDGMENTS ρa = Annulus-side stream density (kg/m3)
ρt = Tube-side stream density (kg/m3)
Julia C. Lemos thanks the Coordination for the Improvement of A = Heat transfer area (m2)
Higher Education Personnel (CAPES) for the postdoctoral Areq = Required heat transfer area (m2)
fellowship through the PNPD Program. André L. H. Costa Aa = Annulus-side flow area (m2)
thanks Rio de Janeiro State University (UERJ) for the financial At = Tube-side flow area (m2)
support through the Prociência Program and the National Cpa = Annulus-side stream heat capacity (J/(kg °C))
Council for Scientific and Technological Development (CNPq) Cpt = Tube-side stream heat capacity (J/(kg °C))
for the research productivity fellowship (Process 311225/2016- dh = Hydraulic diameter (m)
0).

■
dte = Inner tube external diameter (m)
Dte = Outside tube external diameter (m)
NOMENCLATURE dti = Inner tube internal diameter (m)
Parameters Dti = Outside tube internal diameter (m)
̂
ΔPsSTdisp = Stream sST available pressure drop (Pa) F = Correction factor
fa = Annulus-side friction factor
ΔTlm =̂ Logarithmic mean temperature (°C) falam = Annulus-side friction factor for laminar flow
μ̂ sST = Stream sST viscosity (Pa·s) fatran = Annulus-side friction factor for transitional flow
ρ̂sST = Stream sST density (kg/m3) faturb = Annulus-side friction factor for turbulent flow
Φ̂ = Nonlinear continuous functions upper limit f t = Tube-side friction factor
 exc = Minimum area excess (%) f tlam = Tube-side friction factor for laminar flow
CpsST̂ = Stream sST heat capacity (J/(kg °C)) f ttran = Tube-side friction factor for transitional flow
12095 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096
Industrial & Engineering Chemistry Research Article

f tturb = Tube-side friction factor for turbulent flow sL = Index of tube length
ha = Annulus-side convective heat transfer coefficient (W/ sNua = Index of ranges of annular region Seider and Tate
(m2 °C)) Nusselt number
ht = Tube-side convective heat transfer coefficient (W/(m2 sNut = Index of ranges of inner tube Seider and Tate Nusselt
°C)) number
ka = Annulus-side stream thermal conductivity (W/(m °C)) sPra = Index of ranges of annular region Prandtl number
kt = Tube-side stream thermal conductivity (W/(m °C)) sPrt = Index of ranges of inner tube Prandtl number
La = Annulus-side hydraulic length (m) sRea = Index of ranges of annular region Reynolds number
Lh = Length of a hairpin (m) sRet = Index of ranges of inner tube Reynolds number
Lt = Tube-side hydraulic length (m) sST = Index of streams
Lu = Length of a unit (m)
ma = Annular region mass flow (kg/s)
mt = Inside tube mass flow (kg/s)
■ REFERENCES
(1) Towler, G.; Sinnott, R. Chemical Engineering Design Principles,
NB = Number of parallel branches Practice and Economics of Plant and Process Design; Butterworth-
Nh = Number of hairpins per unit Heinemann: Burlington, 2008.
NPa = Annulus-side units in parallel per branch (2) Kakaç, S.; Liu, H. Heat Exchangers Selection, Rating and Thermal
NPt = Tube-side units in parallel per branch Design; CRC Press: Boca Raton, 2002.
(3) Guy, A. R. Operational advantages. Heat Exchanger Design
Nua = Annulus-side Nusselt number
Handbook; Hewitt, G. F., Ed.; Begell House: New York, 2008.
NuaGni = Annulus-side Gnielinski Nusselt number (4) Omidi, M.; Farhadi, M.; Jafari, M. A comprehensive review on
NuaHau = Annulus-side Hausen Nusselt number double pipe heat exchangers. Appl. Therm. Eng. 2017, 110, 1075.
NuaS&T = Annulus-side Seider and Tate Nusselt number (5) Sahiti, N.; Krasniqi, F.; Fejzullahu, X.; Bunjaku, J.; Muriqi, A.
Nuatheo = Annulus-side theoretical Nusselt number Entropy generation minimization of a double-pipe pin fin heat
Nut = Tube-side Nusselt number exchanger. Appl. Therm. Eng. 2008, 28, 2337.
NutGni = Tube-side Gnielinski Nusselt number (6) Syed, K. S.; Iqbal, Z.; Ishaq, M. Optimal configuration of finned
NutHau = Tube-side Hausen Nusselt number annulus in a double pipe with fully developed laminar flow. Appl. Therm.
NutS&T = Tube-side Seider and Tate Nusselt number Eng. 2011, 31, 1435.
Nuttheo = Tube-side theoretical Nusselt number (7) Iqbal, Z.; Syed, K. S.; Ishaq, M. Optimal convective heat transfer in
double pipe with parabolic fins. Int. J. Heat Mass Transfer 2011, 54,
Pra = Annulus-side stream Prandtl number
5415.
Prt = Tube-side stream Prandtl number (8) Iqbal, Z.; Syed, K. S.; Ishaq, M. Optimal fin shape in finned double
Rea = Annulus-side stream Reynolds number pipe with fully developed laminar flow. Appl. Therm. Eng. 2008, 51,
Ret = Tube-side stream Reynolds number 1202.
Rfa = Annulus-side stream fouling resistance (m2 °C/W) (9) Han, H.; Bing-Xi, L.; Wu, H.; Shao, W. Multi-objective shape
Rft = Tube-side stream fouling resistance (m2 °C/W) optimization of double pipe heat exchanger with inner corrugated tube
U = Overall heat transfer coefficient (W/(m2 °C)) using RSM method. Int. J. Therm. Sci. 2015, 90, 173.
va = Annulus-side velocity (m/s) (10) Dastmalchi, M.; Sheikhzadeh, G. A.; Arefmanesh, A.
vt = Tube-side velocity (m/s) Optimization of micro-finned tubes in double pipe heat exchangers
using particle swarm algorithm. Appl. Therm. Eng. 2017, 119, 1.
Binary variables (11) Söylemez, M. S. Thermoeconomical Optimization of Double-
yBsB = Number of branches selection Pipe Heat Exchanger for Waste heat Recovery. J. Thermophys. Heat
ydsd = Inner tube diameter selection Transfer 2004, 18, 559.
yDsD = Outer tube diameter selection (12) Swamee, P. K.; Aggarwal, N.; Aggarwal, V. Optimum design of
double pipe heat exchanger. Int. J. Heat Mass Transfer 2008, 51, 2260.
yLhsLh = Hairpin tube length selection
(13) Serth, R. W. Process Heat Transfer Principles and Applications;
yNhsNh = Number of hairpins per unit selection Academic Press: Burlington, 2007.
yNuasNua = Range of annular region Seider and Tate Nusselt (14) Duran, M. A.; Grossmann, I. E. An outer-approximation
number selection algorithm for a class of mixed-integer nonlinear programs. Mathematical
yNutsNut = Range of inner tube Seider and Tate Nusselt Programming 1986, 36, 307.
number selection (15) Saunders, E. A. D. Heat Exchangers Selection, Design and
yPasE = Annulus-side number of units in parallel per branch Construction; John Wiley & Sons: New York, 1988.
selection (16) Incropera, F. P.; Dewitt, D. P. Fundamentals of Heat and Mass
yPrasPra = Range of annular region Prandtl number selection Transfer; John Wiley & Sons: New York, 2007.
yPrtsPrt = Range of inner tube Prandtl number selection (17) Williams, H. P. Model Building in Mathematical Programming;
yPtsE = Tube-side number of units in parallel per branch Wiley: New York, 2013.
selection
yReasRea = Range of annular region Reynolds number
selection
yRetsRet = Range of inner tube Reynolds number selection
yTsST = Stream sST allocation (1 = tube-side; 0 = annulus-
side)
Subscripts
sB = Index of number of branches
sd = Index of inner tube diameters
sD = Index of outer tube diameters
sE = Index of number of units per branch
12096 DOI: 10.1021/acs.iecr.9b01536
Ind. Eng. Chem. Res. 2019, 58, 12080−12096