International Journal of Heat and Mass Transfer 207 (2023) 124003

Contents lists available at ScienceDirect

International Journal of Heat and Mass Transfer

journal homepage: www.elsevier.com/locate/hmt

Comparison of Shape Optimization Methods for Heat Exchanger Fins

Using Computational Fluid Dynamics
Justin Weber a, E David Huckaby a, Douglas Straub a
US Department of Energy, National Energy Technology Laboratory, 3610 Collins Ferry Rd., Morgantown, 26505, WV, USA

a r t i c l e i n f o a b s t r a c t

Article history: Inverse design techniques are one way to leverage advances in 3D printing, artificial intelligence, and
Received 8 November 2022 computational resources to achieve increased performance of heat exchangers. Two optimization tech-
Revised 8 February 2023
niques (genetic algorithm and particle swarm) and three geometry representations (binary level set, com-
Accepted 19 February 2023
posite Bézier, and free form deformation) are used to increase the heat transfer and reduce the pressure
Available online 6 March 2023
drop of a heat exchanger fin. After running 210,810 OpenFOAM simulations, results indicate that a signif-
Keywords: icant performance increase of the fin can be realized in less than 48 hrs, allowing for such a process to
Shape optimization be integrated in traditional design processes. The best design increased the performance of the objective
Topology optimization function, compared to the baseline rectangular geometry by 75%. A custom distributed infrastructure was
Heat exchanger built, allowing for all methods to reach 95% of the final objective values in a little over 4 hrs, handling
Level set 1674 OpenFoam simulations per hour.
Composite bézier
Free form deformation Published by Elsevier Ltd.
Genetic algorithm
Particle swarm

1. Introduction optimization (TO) and results in more organic and performant ge-
ometries. The challenge with topology optimization is developing
With the advent of 3D printing, machine learning, and ever methods and processes that find solutions in reasonable time us-
increasing computing resources, inverse design techniques are an ing readily available resources. If TO methods can impact the de-
area of growing interest for heat transfer applications and are al- sign cycle, then these methods will become useful design tools for
ready widely available for other areas like solid mechanics. Instead the heat exchanger industry. A key question for current heat trans-
of using intuition to suggest designs, an algorithm is used to gen- fer TO methods is the time-to-solution using readily available re-
erate designs to meet specified performance goals and design re- sources.
quirements. This concept is already being used to design unique For heat transfer applications, most prior TO publications uti-
solutions to optics [1] and structures using Generative Adversar- lize a finite-element formulation to discretize the governing equa-
ial Networks [2]. However, when it comes to heat transfer applica- tions and the same mesh is typically used for all design iter-
tions, recent review articles [3–5] agree that significant improve- ations [3,5]. Finite-element methods (FEM) are more commonly
ments in the current state-of-the-art are needed to become a use- used for structural analysis, whereas finite volume methods (FVM)
ful design tool. are more commonly used for fluid flow and heat transfer analy-
Improved heat exchanger designs have been a topic of inter- sis. Finite volume methods guarantee element-wise conservation of
est for over 40 years. Although inverse design optimization is the fundamental physical quantities, but few studies have utilized
a relatively new term, multi-objective performance criteria have FVM for discretization in TO heat transfer studies [10–15]. In 2006,
been developed for heat exchanger applications [6]. Some previ- Gersborg-Hansen et al. [10] published one of the first heat transfer
ous parametric design studies for heat exchangers include inter- topology optimization papers that utilized a FVM and a fixed uni-
nally finned tubes [7], roughened surfaces [8], and helical rib- form mesh throughout the domain. In this paper, Gersborg-Hansen
roughened tubes [9]. Parametric designs pre-select important de- et al. [10], applied FVM discretization to a conduction problem and
sign variables (i.e., fin heights, angles, or repeating patterns), sig- reported issues with “jumps” in property and design parameters
nificantly restricting the degrees of freedom. near the gas-solid interface. FEM discretization methods can also
For inverse design methods, ideally there should be no pre- develop numerical instabilities, particularly near the interface [16].
selection, allowing the optimization algorithms complete control For convection-dominated problems, Bruns [16] has described nu-
of the design. This approach is typically referred to as topology merical issues for FEM approaches, and Gersborg-Hansen [10] has

https://doi.org/10.1016/j.ijheatmasstransfer.2023.124003
0017-9310/Published by Elsevier Ltd.
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

edge, only one publication [22] reports the combination of LSM

Acronyms with FVM discretization for heat transfer applications. Yoshimura
et al. [22] reports a specialized Cartesian-mesh computational fluid
BH Basin hopping dynamics (CFD) solver in combination with LSM to optimize some
CFD Computational fluid dynamics example channel flows using a fixed uniform mesh.
DA Dual annealing In another interesting study [21], the concept of body-fitted
DE Differential evolution meshing combined with LSM has been used to optimize 2D and 3D
FVM finite volume method heat exchanger concepts. This approach does not require any mod-
GA Genetic algorithm ification of the original equations describing the physical phenom-
LSM level-set method ena at the interface, but the mesh must be re-generated at each
PS Particle swarm design iteration. According to Feppon et al. [21], the computational
STL Standard Triangle Language time required to run 359 iterations of the optimization algorithm
TO Topology Optimization was about 8 days using 30 processing units. The initial mesh con-
sisted of about 3.8 million tetrahedra elements, but the number of
Variables
elements decreased with each design iteration. The last mesh in
c Child
the design iteration had 1.7 million tetrahedra elements.
c1 PS cognitive term weight, 1 unless otherwise stated.
For heat transfer TO applications, only a few prior studies have
c2 PS social term weight, 2 unless otherwise stated.
J reported gradient-free optimization methods. According to Fawaz
Cp specific heat of air, held constant at 1006 ( kgK )
et al. [5], approximately 95% of the publications prior to 2022, used
h Heat exchanger surface height (m) gradient based optimization algorithms. Gradient based optimiza-
m˙ mass flow rate of air (kg/s) tion techniques are attractive for their computational efficiency, es-
ngenes Number of genes (in GA terms) or parameters pecially adjoint-based optimization where the sensitivities of the
P Pressure (Pa) objective function can be computed for each of the design vari-
p Parent in the GA method ables. These sensitivities can then be used to direct the deforma-
pg Best global position. tion of the geometry. However, gradient based methods converge
pi Best particle position. to a local optimum instead of exploring the entire design space,
ranneal Mutation anneal rate (applies to GA), 0.99 unless making it difficult to select which algorithm is needed without
otherwise stated. prior knowledge of the response surface.
rmutate Mutation rate (applies to GA), 0.2 unless otherwise For non-gradient based optimization algorithms applied to heat
stated. exchangers, there are only three publications [22–24]. Yoshimura
rn Random number between 0 and 1. et al. [22] used a genetic algorithm (GA) assisted by a Krig-
T Temperature (K) ing surrogate model to optimize flow channels with different in-
V Gas velocity (m/s) let/outlet configurations. No solution times were reported. Shi-
vi Particle velocity. moyama et al. [24] used a commercial CFD code with FVM dis-
w PS constant inertial weight, 0.5 unless otherwise cretization to optimize a heat sink with natural convection. Instead
stated. of using gradient-based optimization, Shimoyama et al. [24] com-
xi Current particle position. bined a genetic algorithm (NSGA-II) and a Kriging surrogate model.
Sub and super scripts Although the mesh was regenerated with each design iteration, the
i the ith gene of a chromosome representing a design, possible geometry variations were severely restricted (i.e., neither
or the ith variable in the particle swarm algorithm the density method nor the LSM approach were used). According
in inlet value to Shimoyama et al. [24], 167 design iterations are required to find
k the current iteration an optimum and each evaluation requires 25–40 minutes to con-
out outlet value verge.
The present paper is closely aligned with Mekki et al [23,25].
The similarities include the size of the domain, the boundary con-
described numerical issues with FVM approaches for heat conduc- ditions and the baseline single rectangular fin dimensions [23]. An
tion problems. open-source CFD solver (OpenFOAM) based on a FVM discretiza-
Marck et al. [11,12] also used FVM discretization because it is tion is used in both the current work and prior work [23,25]. In-
widely accepted in the industrial field, but no issues with nu- stead of using a GA to generate new geometries on a voxel basis
merical instability were noted. The lack of numerical instabilities as described by Mekki et al. [23,25], the current paper uses three
with the FVM formulation is consistent with more recent publica- different approaches to represent the geometry, including a LSM,
tions [13–15]. Two common features for all prior FVM discretiza- Bézier curves, and free-form deformation (FFD). Similarly Keramati
tion efforts [10–15] include: 1) a density-based design representa- et al. used the same optimization test problem, but with a deep re-
tion, and 2) a fixed mesh for all design iterations. inforcement learning optimization technique and Bézier curves for
A density-based design representation means each element has defining the geometry [26]. Further discussion and comparison to
two possible geometry states (i.e., solid or fluid) and this approach Mekki et al. and Keramati et al. are provided in Section 6.
has been used in roughly 80% of the TO heat transfer studies Based on the heat transfer TO literature, free form deformation
prior to 2022 [5]. As an alternative, level-set methods (LSM) clearly (FFD) methods have not been reported previously. Instead of using
define the surface, because the element can have three possible a fixed mesh, the current paper generates a new mesh for each de-
states (i.e., solid, fluid, or an interface). LSM approaches have been sign iteration which is consistent with Shimoyama et al. [24] and
used to optimize heat exchanger components [17–21]. Santhanakr- Feppon et al. [21]. This paper will also use non-gradient based op-
ishnan et al. [19] have compared the conventional density-based timization algorithms. In addition to using the GA optimizer, the
method to LSM using an FEM formulation. A fixed mesh for all de- current paper also investigates the use of a particle swarm (PS)
sign iterations was used in this study which concluded the density optimizer.
method was more robust and faster, but LSM predicted a better The purpose of this paper is to benchmark the time required to
design in terms of the objective function. To the authors’ knowl- optimize a fundamental heat exchanger fin using a readily avail-

2
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

able CFD solver (OpenFOAM). The results will provide a range of

benchmark time-to-solution values using combinations of geom-
etry representations and gradient-free optimization strategies. Al-
though many of the methods described in this paper have not been
reported previously, these time-to-solution results are an impor-
tant first step toward the realization of TO as an important tool for
optimized heat transfer solutions. In order to achieve this purpose,
the following were necessary:

1. Develop an infrastructure to manage thousands of CFD simu-

lations in parallel, including geometry and mesh generation at
each design iteration.
2. Implement alternative geometry representation methods (LSM,
Bézier curves, and free-form deformation) without producing
errors in the automated mesh generation.
3. Implement minor modifications to conventional algorithms to
utilize parallel computational resources more efficiently. Fig. 1. 2D voxel array and level set representing a geometry.

4. Compare optimized solution to recent results such as Mekki

et al. [23] and Keramati et al. [26].

The body of this paper will be organized as follows.

Section 2 will describe three different geometry representations,
including a modified level set approach. Section 3 will describe the
computational infrastructure, as well as, the genetic algorithm and
particle swarm optimization methods used in this study, includ-
ing the modifications incorporated to more effectively leverage the
parallel computing resources. Section 4 will describe the results
from the test problem, as well as, the time required to reach 95% of
the optimum solution. Section 5 compares the results and perfor-
mance of the optimization campaigns, followed by Section 6 which
compares the results of this work to Mekki et al. [23] and Kera-
mati et al. [26]. The final section summarizes the work, concluding
that the genetic algorithm finds the ”best” solution, and the parti- Fig. 2. Composite Bézier curve example using four control points and four Bézier
cle swarm method with the free-form deformation is the fastest to curves.
converge to 95% of the final solution. All methods described in this
paper reached 95% of the final objective values in approximately 4
points, optimization algorithms should converge faster due to the
hrs.
fewer degrees of freedom. A quadratic Bézier curve can be defined
using 3 points, P0 , P1 , and P2 [27].
2. Geometry representations
B(t ) = (1 − t )[(1 − t )P0 + tP1 ] + t[(1 − t )P1 + tP2 ] (1)
2.1. Binary level sets
A composite Bézier curve can be constructed by combining a
One of the simplest approaches to performing topology opti- series of quadratic Bézier curves. To ensure a smooth transition
mization is to simply discretize the design space into a number between Bézier curves, mid points between the control points are
of voxels. Each voxel can either have a value of 1, meaning that used for the endpoints P0 and P2 , while the control point is used
that voxel is solid, or a value of 0, meaning that that voxel is gas. for center point, P1 , following Stärk’s construction, Fig. 2.
To generalize this approach specifically for CFD meshers that typi- The Bézier curves are then sampled at an arbitrary resolution,
cally use a CAD format as geometry input, a level set is calculated t. The sampled points are then copied into the 3rd dimension and
at a value of 0.5, Fig. 1. If the geometry is 2D, this level set can triangulated to create an STL file which can then be used by vari-
be extruded into the third dimension and triangulated to create ous meshing tools. This geometry representation requires continu-
a Standard Triangle Language (STL) file. This geometry representa- ous variables.
tion requires an optimization algorithm that can work with binary
variables. 2.3. Free form deformation
This geometry approach is similar to what Mekki et al.
used [23,25]. However, difficulties arise when meshing the solid- Free form deformation is a technique for creating new geome-
gas interface around some voxels, so in this work, smoother ge- tries from a transformation of an initial geometry. The undeformed
ometries are generated by interpolating the surface profile through geometry, that can be defined with a collection of points (geomet-
the midpoint of the ’edge’ voxels, Fig. 1. This allows for represen- ric points) in space, is enclosed with a surrounding cube, or hull,
tation of more aerodynamic shapes without stair-stepped geome- an object which can be defined with another set of points (control
tries. Additionally, since an STL is also created, this method works points). These control points are moved in space to define a gen-
with existing CFD workflows. eral transformation which is used to deform the original geometry.
Fig. 3 illustrates the construction of an airfoil shape (solid) from a
2.2. Composite Bézier circle (dashed) by moving the control points from the initial loca-
tions generated from the circle (open) to a different set of positions
To generate a smooth surface, specifically for aerodynamic fea- (closed). These control points can then be moved and the new po-
tures, composite quadratic Bézier curves can be used to gener- sitions of the geometry points are calculated using trivariate ten-
ate complex features with few control points. With fewer control sor product Bernstein polynomial [28]. This technique allows for

3
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Fig. 3. Free form deformation of a circle with a 3x3 lattice.

smooth deformations of any geometry and is commonly used in

shape optimization of aircraft wings.
In this work, the geometry is defined with an STL file that is
fit with a lattice of control points. The same number of control
points are used for both the free form deformation technique and Fig. 4. Structure of the software used to handle concurrent evaluations needed by
the composite Bézier curve method for representing the geometry. the optimization algorithms.

3. Optimization methods
Since the server and clients talk to each other using https, this
framework could also handle evaluations distributed over a local
The optimization approaches in this work are focused on meta-
network or internet. This allows for usage of distributed resources
heuristic black box algorithms, where underlying knowledge of the
instead of being limited to a single large cluster. The user could
evaluator is not known. This approach was taken for the following
test the algorithm with local resources and scale up evaluations
reasons:
by adding other remote or cloud resources. Further, the framework
1. Any method (i.e., CFD, test functions, etc.) can be used to pro- allows for on-the-fly scale-up by dynamically adding clients while
vide an objective value for the optimization method. the algorithm is running.
2. No gradient calculations or estimates are needed, removing the
need for significant code development for the method of ad- 3.2. Genetic algorithm
joints.
3. These methods allow for many concurrent evaluations of the A custom asynchronous genetic algorithm was developed and
objective function, effectively utilizing large computational re- implemented in python to handle both binary and continuous vari-
sources. ables. The algorithm follows the standard selection, crossover, and
4. These methods are global optimizations instead of local opti- mutation operations. However, new children are created on de-
mizations. mand, as opposed to generating the new population all at the
same time [31]. The consequence is that the population entirely
Since these optimization algorithms are not using more seri-
consists of elites. In other words, when a new evaluation is com-
ally efficient approaches that calculate and use sensitivity gradi-
pleted, it is added to the population and the worst performing
ents, such as the adjoint method, they need to handle hundreds
member in the population is removed. Additional details on the
of thousands of simultaneous CFD simulations. The methods have
selection, crossover, and mutation operations are described in the
been modified as described below to allow for asynchronous or
following paragraphs.
continuous evaluations so that a single slow evaluation (in this
Selection. For both continuous and binary variables, a tourna-
case, a CFD simulation) does not prevent other evaluations from
ment selection is used to select two parents for crossover. The
executing. This approach will effectively maximize the utilization
method works by selecting five individuals from the population at
of parallel computational resources, reducing the overall time to
random and returning the best scoring individual for the first par-
solution.
ent. This processes is repeated for selection of the second parent.
Crossover. Once two parents are selected, their genes are mixed
3.1. Infrastructure
using a crossover operator. For Binary variables where the variable
is either True or False, bit-wise and and or operators are typically
To further improve evaluation throughput on a traditional clus-
used to generate a single child which consists of bits where both
ter using a batch scheduling systems (Slurm in this case), a cus-
parents are True or where either parent is True, respectively. Addi-
tom server-client architecture was constructed, Fig. 4. The server,
tionally, random, random point, and random box operators can be
implemented with Flask, hosts the optimization algorithm, collects
used which simply swap genes between the two parents based
results, generates new samples, and displays progress through a
on randomly selected genes, a single randomly selected point in
dashboard using Dash, Fig. 5 [29,30]. The clients poll the server
the flattened array, or a randomly selected box in the 2D or 3D
for new samples, evaluate the sample, and return the results.
array, respectively. For Continuous variables, a heuristic crossover
The clients can handle an arbitrary number of simultaneous
method is used where the values of the variables are combined
evaluations. This allows for efficiently packing evaluations into
using a random weight [32].
batch scheduling systems limits. For example, if the minimum
batch size is 40 cores and we want to run a serial evaluation, c1i = p1i − βi ( p1i − p2i ) (2)
the client can manage 40 concurrent serial evaluations in that sin-
gle batch request. Additionally, this framework creates a personal
c2i = p2i + βi ( p1i − p2i ) (3)
queue for the optimization process and avoids the need for con-
stantly submitting requests to the batch queuing system, further Where p1i is the ith
gene of the first parent, p2i is the gene ith
reducing the time to solution. of the second parent, c1i is the ith gene of the first child, c2i is the

4
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Fig. 5. The dashboard, implemented with Dash, used to monitor optimization progress of an optimization run using the genetic algorithm and binary level set geometry
representation.

ith gene of the second child, and βi is a randomly chosen value generated. Each particle has a position, which is the current values
between 0 and 1 for the ith gene. Since the GA is asynchronous, of the parameters being optimized, and a velocity [33]. The new
only one child is needed. If the crossover operator generates two particle velocity, vki +1 , is composed of an inertia term, a cognitive
children, one is randomly selected. term, and a social term.
Mutation. Once a child has been generated, there is a proba-
bility, rmutate , that the child will be mutated. For Binary variables, vki +1 = wk vki + c1 r1 ( pki − xki ) + c2 r2 ( pkg − xki ) (5)
three mutation operators are available including random and two
Where wk is a constant inertial weight, vki is the current particle
more morphological operators edge and erosion/dilation. The ran-
velocity, c1 is the cognitive term, r1 and r2 are random numbers
dom operator simply picks rmutate ∗ ngenes genes to randomly flip
between 0 and 1, pki is the best individual particle position, c2 is
(False to True or True to False). The edge operator finds genes on the
edge of where opposite bits meet, randomly picking rmutate ∗ ngenes the cognitive term, and pkg is the best swarm position. For Contin-
genes from that edge list to flip. The erosion/dilation operator is uous variables, the new particle position can simply be calculated
similar to the edge operator except for instead of flipping the se- by the addition of the current position and updated velocity. Posi-
lected bits, those bits are used as the center of a randomly selected tions are not allowed to exceed the variable limits.
erosion or dilation of a user selected shape.
xki +1 = xki + vki +1 (6)
For mutation of Continuous variables, the variables are first
scaled to be from 0 to 1. Two different operators can then be used To achieve an asynchronous or parallel particle swarm algo-
to change rmutate ∗ ngenes randomly selected genes. The random op- rithm, the particles continuously move around the domain evaluat-
erator simply replaces the selected genes with a random value be- ing the objective function. The only value that needs to be updated
tween 0 and 1, while the normal operator adds a random value and synchronized across the particles is the global best position,
from the normal distribution, centered at 0 with a standard devi- which is updated every time a particle completes a new evalua-
ation equal to rmutate . In this work, the random mutation operator tion.
is used. For Binary variables, the continuous velocity variables need to
If rmutate stays constant, the GA can explore a large design space be cast to a binary value. Although this seems possible with the
but has trouble resolving a global minimum. As a result, the rmutate particle swarm algorithm [34], the authors have been unable to get
is allowed to decrease during the optimization process if the ob- a particle swarm algorithm to make significant progress with the
jective function does not improve over 500 iterations. After 500 binary level set geometry approach.
stalled iterations, the new mutation rate is then determined by:
rmutate,new = rmutate ∗ ranneal (4) 3.4. Verification
Where ranneal is value less than 1, typically 0.99.
Since these algorithms are slightly different than standard ap-
3.3. Particle swarm proaches, they are compared to the global optimization algorithms
implemented in SciPy including differential evolution (DE), dual
A custom asynchronous particle swarm algorithm was devel- annealing (DA), and Basin hopping (BH) [35] as well as the genetic
oped and implemented in python. This implementation is the most algorithm (GA) implemented in PyGAD [36] with all the defaults.
primitive form of the algorithm where a population of particles is The test function used is the two parameter, continuous variable,

5
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Table 1
Comparison of base simulation parameters. Note: As-
sumptions were made and values extracted from plots
to estimate values from Mekki et al. [23].

Mekki[23] This work

Total cells 10,000 11,933

Meshing - 78 s
Solving - 15 s
Q/h 178 168 W/m
dP 4.476 4.748 Pa
f ob jective 0.108 0.100 W/Pa1/3

Where m˙ is the mass flow rate of air, C p is the specific heat of air
at 300K, Tout is the mass averaged outlet temperature, Tin is 300K,
Pin is the inlet pressure, and Pout is the outlet pressure (101325 Pa).
This function, developed by Webb and Eckert [8], corrects the heat
transfer rates of different designs at fixed flow rate to fixed pump-
Fig. 6. Comparison of the number of evaluations of the Ackley function required ing power and has become a commonly used performance mea-
by the optimizations methods. The violin plot displays the kernel density estimated
sure in the heat transfer enhancement community [38,39].
distribution, 25% quartile (Q1), median, and 75% quartile (Q3).
The fin geometry is allowed to vary anywhere in a subset of
the computational domain so that the geometry does not touch
Ackley function, defined as [37]: any of the domain boundaries. Specifically, the geometry is limited
  1 n  to fit inside a box from 0.1 to 9.9 mm in the x direction and 0.1 to
fAckley (x0 · · · xn ) = −20 exp −0.2 i=1 x2i 2.9 mm in the y direction. This allows for at least one background
   n
(7)
mesh cell to be between the geometry and the edge of the simu-
− exp 1n ni=1 cos(2π xi ) + 20 + e
lation domain.
where − 32 ≤ xi ≤ 32
4.1. Computational model
Where there are many local minima and one global minimum
at xi = 0 with a value of 0. By using this test problem, the ro- Similar to Mekki et al., bouyantSimpleFoam from the OpenFOAM
bustness and time to solution of the optimization algorithms can CFD toolbox is used to predict the flow and heat transfer for the
be quickly tested and compared. Due to the stochastic nature of different geometries [23]. bouyantSimpleFoam is an incompressible
the optimization algorithms, each optimization method is run 100 flow solver with heat transfer using the Boussinesq approximation.
times and stopped when the objective value is less than 0.001. However, for this application the gravity is turned off; eliminat-
The particle swarm algorithm implemented in this work is the ing the feedback from the temperature field to the flow. Further,
fastest and most consistent converging algorithm, averaging 406 for turbulence, the k-ω SST model is used along with the asso-
function evaluations to find the minimum, Fig. 6. Next best is the ciated momentum and thermal wall-functions. Air is the working
dual annealing, followed by basin hopping, differential evolution, fluid. snappyHexMesh, a Cartesian dominant meshing tool, is used
the genetic algorithm used in this work, and finally PyGAD’s ge- to mesh the candidate geometries (STL files). The models start with
netic algorithm, averaging 662, 882, 925, 1923, and 2130 function a background mesh of 100x50x1 cells. During meshing, two lay-
evaluations, respectively. Based on this comparison, the optimiza- ers of cell refinement with 10 layers are added around the geome-
tion algorithms used in this work represent one of the fastest and try, allowing for fine details of the surfaces to be resolved without
one of the slowest converging algorithms. The GA implemented significantly increasing the overall cell count. Since the refinement
here is also more consistent than PyGAD’s, most likely due to the adds cells in the z direction, the first layer of cells in the x-y direc-
population consisting entirely of elites. Interestingly, it seems that tion are extruded into the z direction using the OpenFOAM utility,
PyGAD’s GA occasionally gets lucky and finds the minimum excep- extrudeMesh.
tionally fast. All models and optimization campaigns are run on NETL’s Joule
2.0 supercomputer. Each node contains two 20 core Intel Xeon
4. Test problem Gold 6148 CPUs clocked at 2.40GHz [40]. The optimization cam-
paigns each use four nodes, for a total of 160 cores. 20 simultane-
Mekki et al. proposed a simple test problem that optimizes a ous models are allowed to run on each node, totaling 80 concur-
single heat exchanger fin [23]. This problem is used to compare rent simulations.
the two optimization strategies and three geometry representation The base model compares well with Mekki et al., Table 1. 11,933
strategies described in this work. total cells are used. The objective function value for the base model
This problem consists of a 3 mm wide and 10 mm long chan- is coincidentally 0.1, with a length normalized heat transfer rate,
nel containing a rectangular fin, Fig. 7. The geometry represents a Q/h, of 168 W/m and a pressure drop, dP , of 4.748 Pa. The model
two-dimensional simplification of a “flow unit” of a phase-fin heat converges in 249 iterations, taking about 15s. However meshing
exchanger. The starting geometry consists of a 0.5 x 5 mm rectan- with snappyHexMesh takes 85% of the total evaluation time, or 78s.
gular fin, centered in the domain. Air at 300K is injected at 3 m/s
at the inlet. The fin’s walls are fixed at 450K and system is main- 4.2. Solution checking
tained at atmospheric pressure. The optimization goal is to maxi-
mize the objective function, which combines two parameters, the After running hundreds of thousands of CFD simulations blindly
pressure drop across the fin, dP , with the heat transfer from the through optimization algorithms, it became quickly apparent that
fin to the fluid, Q: the optimization algorithms are very good at finding and exploit-
Q m˙ C p (Tout − Tin ) ing issues with the mesh. Just because the residuals of the solution
fob jective = = (8) indicate that the model has converged, does not mean the solution
dP 1/3 (Pin − Pout )1/3
6
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Fig. 7. Description of the base simulation.

is valid. When the STL file generated with the composite Bézier
or free form deformation techniques is self-intersecting, then over-
lapping and malformed cells can occur. To screen out malformed
meshes, OpenFoam’s checkMesh utility is used. Any proposed de-
sign that results in a checkMesh failure is discarded.
If the mesh is accepted, then the simulation is run. Once the
simulation is finished, the contGlobal residual is checked to ensure
that residual has fallen below 10−6 . Finally, the objective function
is calculated. Since the pressure drop is in the denominator, any
designs with a dP <= 0 are rejected.
If a design is rejected for any reason during the evaluation of
the objective function, a NaN (not a number) objective value is re-
turned to the optimization algorithm. Luckily, the GA and PS al-
gorithms seem resilient to these missing values. For the GA algo-
rithm, since the population is continuously sorted, the failed de-
signs are simply discarded during the sorting of the population and
the algorithm can continue. For the PS algorithm, the rejected de-
signs do not change the global or particle best positions and the
next particle can continue as if the failed position never happened. Fig. 8. Objective value over time for all geometry and algorithm combinations. For
rendering purposes, only about 11% of the evaluations are displayed, decimated with
spatial consideration.
4.3. GA with composite Bézier

A combination of the genetic algorithm with the composite heat transfer through the fin itself needs to be imposed to opti-
Bézier geometry representation was used to optimize the base heat mize a real design.
exchanger fin design. 10 control points were used to define the The combination of the GA with the composite Bézier finds a
base geometry. This required a total of 20 continuous variables to solution that is 174% better than the baseline objective function.
be optimized using the GA. The GA has a population of 50 and The optimal design has about 70% the pressure drop of the original
a starting mutation rate, rmutate , of 0.2. After 500 stalled itera- design and 155% of the heat transfer to the fluid, Table 2.
tions without improvement in the objective value, the mutation
rate is reduced using ranneal = 0.99. The random mutation method 4.4. GA with FFD
was used. The initial population was created by mutating the base
design. The free form deformation geometry was used in combination
Over 48 hrs, which is the job limit on Joule 2.0, a total of 25,511 with the genetic algorithm to optimize the baseline design. A 5x2
successful objective evaluations were completed. Over the entire lattice was used as the control points to deform the original geom-
optimization campaign, about 531 evaluations were completed ev- etry. Consistent with the previous Bézier geometry, this requires a
ery hour. Most of the initial innovation occurs within the first 2.4 total of 20 continuous variables to be optimized using the GA. A
hrs, reaching an objective value within 95% of the final objective population of 50 is used with a starting mutation rate, rmutate , of
value of 0.175, Fig. 8. The GA seems to do a good job of proposing 0.2, and after 500 stalled iterations, the mutation rate is reduced
a wide variety of designs, resulting in a wide variety of objective using ranneal = 0.99. The random mutation method was used.
values with different pressure drops and heat transfer rates, Fig. 9. During this 48 hrs optimization campaign, 27,419 simulations
The final design is significantly thinner than the original de- were completed, averaging 571 simulations per hour. In about
sign and tries to maximize the length of the fin by reaching from 1.3 hrs, the best preforming design has already reached 95% of the
one corner of the domain to the opposite corner, Fig. 10. Un- final best objective value. The GA does a good job of generating a
fortunately, this is an unrealistically thin fin, with the geometry wide variety of designs with different heat transfer and pressure
approaching a zero thickness solution. Random perturbations by drops, Fig. 9.
the optimization algorithm can be larger than the thickness, re- The final optimal design is almost identical to the GA with com-
sulting in self-intersecting STLs. This is not surprising since there posite Bézier final design, except flipped, Fig. 10. Due to the sym-
is no constraint on the thickness. In reality, manufacturing toler- metry, this suggests that there are two global minima in the design
ances, thermo-mechanical material failure criteria, and conjugate space, as both designs have the same objective function value of

7
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Table 2
Comparison of base design to the optimized designs.

GA Bezier GA FFD GA lvlset PS Bézier PS FFD

f ob jective 0.175 0.175 0.152 0.165 0.166 W/Pa1/3

dP 3.3 3.2 3.0 3.4 2.8 Pa
Q/h 261.2 258.0 220.3 248.0 233.1 W/m
Evaluations 25,511 27,419 80,351 39,801 37,728 -
Rate 531 571 1674 829 786 1/hr
Time to 95% 2.4 1.3 4.2 0.7 0.2 hr
f ob jective / f ob jective,Base 174 175 152 165 165 %
dP/dPBase 70 68 64 71 59 %
Q/QBase 155 153 131 147 138 %

Overall, the GA with either the composite Bézier or free form

deformation geometry strategies resulted in almost the same de-
sign. Interestingly, the free form deformation approach achieved
the design slightly faster.

4.5. GA with level set

The genetic algorithm and the binary level set geometry repre-
sentation are combined to optimize the base geometry. Consistent
with the previous GA based optimization runs, a population of 50
is used with a starting mutation rate, rmutate , of 0.2, and after 500
stalled iterations, the mutation rate is reduced using ranneal = 0.99.
The design is limited to a single body and any holes that appeared
inside the body after mutation are filled.
The design space is discretized using 50 cells in the x direc-
tion and 50 cells in the y direction, resulting in 2500 boolean vari-
ables that the GA mutates between 0 and 1 to represent either gas
or solid. This geometry discretization is independant of the Open-
Fig. 9. Pareto plot of the two parameters for all geometry and algorithm combi- FOAM fluid mesh discretization.
nations. Contour lines are the values of the objective function, Eq. 8. For rendering A total of 80,351 OpenFOAM simulations were completed in
purposes, only about 11% of the evaluations are displayed, decimated with spatial con-
48 hrs using 4 clients each running 20 simulations for a total of
sideration.
80 concurrent simulations. The infrastructure successfully handled
1675 evaluations per hour of the objective function. The GA algo-
rithm successfully explores a significant design space, generating
0.175. Similar to the other Bézier geometry cases, the optimal de- many different designs of different sizes and shapes. 4.2 hrs into
sign is just thick enough for the mesher to intersect the geometry the optimization, the design already achieves an objective value
and generate the mesh. within 95% of the final optimal objective, Fig. 8.
The free form deformation combined with the GA worked well, Although this is a single objective optimization, the algorithm
resulting in a 175% better objective function. The optimal design seems to find and start to resolve a Pareto front as the pressure
has 68% the pressure drop of the original design and 153% of the drop and heat transfer are in competition with each other, Fig. 9.
heat transfer to the fluid, Table 2. The algorithm also explores the largest space, compared to the

Fig. 10. Gas temperatures of the base design compared to the optimal designs achieved using two optimization algorithms and three geometry representations.

8
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

other campaigns, finding designs that have pressure drops all the transfer increased to 153% of the baseline. This was closely fol-
way down to 1 Pa. lowed by the GA with Bézier fin. The fastest converging algorithm
The resulting optimal design is considerably thinner and longer was the PS with FFD, reaching 95% of the final objective value in
than the initial design of a rectangular fin, Fig. 10. The objec- 0.2 hrs. Further details are reported in Table 2.
tive value for the optimum is 152% better then the original value, Inferring from the proposed optimal designs, it seems that
achieving lower pressure drop and a marginally improved heat there are two global ”best” designs, symmetrical about the center-
transfer rate, Table 2. line, Fig. 11. These designs extend diagonally from either the top-
left corner to the bottom-right corner (GA Bézier) or from the
4.6. PS with composite Bézier bottom-left corner to the top-right corner (GA with FFD), and
achieved almost identical objective values. Intuitively this makes
The particle swarm optimization algorithm and composite sense because these designs would have the largest surface area
Bézier geometry method are combined to optimize the base heat with the smallest cross-sectional area without significant flow
exchanger fin. 50 particles are used to optimize the 10 control turning or flow separation. Having a larger cross-sectional area will
points defining the fin shape. During initialization of the particle accelerate the flow resulting in higher pressure drops. Likewise,
positions, a random displacement was added to the original con- turning the flow will increase the pressure drop.
trol point positions. This random displacement was restricted to a Comparing the optimization techniques, the stereotypical fight
maximum of 10% of the domain length, ±1 mm in the x direction between exploration and convergence to a optimal is observed.
and ±0.3 mm in the y direction. The PS hyper parameters were The genetic algorithm does better at exploring the design space
kept at the default values of w = 0.5, c1 = 1, and c2 = 2. for all geometry methods, ultimately resulting in the most perfor-
A total of 39,801 successful simulations were completed in 48 mant design, at the cost of time to solution. The wide breadth of
hrs, averaging 829 simulations per hour. The optimization cam- proposed designs are evident in the Pareto plot, Fig. 9. The parti-
paign quickly finds a design in only 0.7 hrs which has an objective cle swarm algorithm converges exceptionally fast to a good design.
value within 95% of the final optimal objective value, Fig. 8. Although the PS algorithm is not using local gradients, it is using
Since the PS algorithm converges quickly, the variation in de- the global best position and the particle best position as weighted
signs is not as large as the GA methods, Fig. 9. The final optimized global gradients to help it converge to a solution. However, the PS
design reaches an objective function value of 0.165. The optimized optimal objective function value was about 5% lower than the GA
shape is fairly similar to the GA with Bézier, but the objective func- solution. The PS algorithm does not explore the design space as
tion for the PS algorithm is about 5% lower than the optimum well, quickly converging to a fin design and getting stuck. This sug-
found using the GA method, Fig. 10. This comparison demonstrates gests that there are local minimas in the design space.
the typical fight between exploration and speed with these meta- There are numerous hyper-parameters that can be adjusted to
heuristic black box algorithms. affect these observations, however it is difficult to select these
The resulting optimal design is still better than the baseline de- hyper-parameters a priori since they are sensitive to the specific
sign, with an objective value 165% better than the original. The problem. Due to this, the hyper-parameters were not explored and
pressure drop is 70% of the baseline, and the heat transfer rate is kept at the default values.
147% of the baseline, Table 2. Comparing the geometry methods, the free form deformation
technique had both the fastest time to convergence and the best
performing designs. Both the FFD and composite Bézier geometries
4.7. PS with FFD use a total of 20 continuous variables. Combined with the GA and
PS optimization techniques, these approaches allow for the geom-
The particle swarm optimization algorithm using 50 particles, etry to be varied significantly between iterations throughout the
combined with the free form deformation technique, was used design space. Conversely, the binary level set approach uses a total
to optimize the original geometry. Matching the GA optimization of 2500 binary variables to represent the design space. Since the
campaign, a 5 x 2 lattice was fit to the geometry. During initializa- GA mutation operator is only flipping variables along the gas/solid
tion of the particle positions, a random displacement, up to a max- interface, the fin geometry does not change significantly between
imum of 10% of the axis length, was added to the original lattice iterations, requiring many mutations to achieve significant design
positions. The PS hyper parameters were kept at w = 0.5, c1 = 1, deviations. This results in the binary level set approach being the
and c2 = 2. lowest of the geometry approaches.
The optimization process was allowed to run for a total of 48 Additionally, the FFD and composite Bézier geometries allow for
hrs, with a maximum of 50 concurrent simulations, restricted by generally smoother and more continuous geometries resulting in
the number of particles. A total of 37,728 successful simulations lower pressure drops. This helps the optimization algorithm sug-
were completed, averaging about 786 simulations per hour. This gest more aerodynamic shapes opposed to the level set method
optimization run converged the fastest, with a design reaching 95% which typically results in a surface with protruding features due
of final optimal design in just 0.2 hrs, Fig. 8. to the underlying discretization of the geometry. If the underlying
Again, the PS algorithm does not seem to explore a large design discretization was finer, then the surface would be smoother. Fi-
space, quickly converging to an optimum design, Fig. 9. The result- nally, there are no constraints on the Bézier curves. As the geom-
ing optimal design is similar to the others, becoming very thin and etry gets thinner, the chance of the Bézier curves self-intersecting
long. The resulting design has an objective value of 0.166 W/Pa1/3 , increases, often resulting in failed meshes and strange geometries.
165% better than the base design. The pressure drop is 59% of the Further, since the nodes of the Bézier curves are not ordered, mul-
original with the heat transfer increasing to 138% of the base de- tiple sets of points could create the same shape. With an N-point
sign. cyclic-Bézier curve there are at least N duplicate shapes since the
starting point is not unique. This could confound the optimization
5. Comparison of optimal designs algorithms.
The FFD approach seems to be the best geometry approach for
Comparing all the optimal designs, the GA with FFD resulted in this test problem because it offers a high degree of variability with
the best fin design, receiving the maximum objective value of 0.175 a low probability of self intersection. Further, there are not mul-
W/Pa1/3 . The pressure drop is 68% of the baseline while the heat tiple solutions that can define the same shape. Which all comes

9
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Fig. 11. Base and optimal fin designs of the 2 optimization algorithms and 3 geometry representations.

from the elegance of the technique, by utilizing a lattice of control

points to deform the original geometry. However, FFD can be re-
strictive, preventing small features. For this test problem, that does
not seem to be an issue.

6. Comparison to other works

There are currently two other works in the open literature that
have performed optimization on a similar test problem with two
different optimization algorithms. However, it is difficult to directly
compare the final objective values across all works due to slight
differences in problem descriptions and missing details. Regardless,
assumptions were made and values were extracted from the pro-
vided plots.
Mekki et al. was first to suggest this test problem specifically
for heat exchangers [23]. Mekki et al. used a binary genetic al- Fig. 12. Optimal design suggested by Mekki at al., extracted from Figure 18 [23],
gorithm with similar mutation and cross-over operators. However and one of the best designs suggested by Keramati et al., extracted from Figure 15b
Mekki et al. did not use an asynchronous GA algorithm and did [26].
not use a level set operator to smooth the resulting geometries.
This resulted in stair-stepped fins. Mekki et al. also used Open-
FOAM with the same bouyantSimpleFoam solver, and the same per- solver, is used to determine the objective function with the same
formance metric. performance metric [41]. One of the best performing shapes is
Mekki et al. explored different parameters of the GA algorithm, closer to this work, reducing the pressure drop by 60% and in-
however the most performant single fin design was significantly creasing the heat transfer by 30%. However the resulting geoemtry
different, Fig. 12. The design drastically increased the pressure drop is significantly shorter, Fig. 12 [26]. Because the starting geometry
by 5.3 times and increased the heat transfer by 2.6 times, result- is not consistent, the objective improvement can not be compared.
ing in an overall objective improvement of 45% [23]. It is unknown However, the resulting optimal fin design is still different than the
if the different design is due to slight differences in the GA algo- optimal designs in this work. It is not clear if this difference is due
rithm, the geometry approach, OpenFOAM settings, or potential is- to the algorithm, FEniCS solver, the different starting fin geome-
sues elsewhere in the implementation. Additionally, no timing in- try, or other details omitted. The optimization process reached a
formation was provided and is unclear how much human involve- steady state in approximately 200 episodes, with 50 simulations
ment was required during the process. However total simulation per episode, or a total of 10,0 0 0 simulations [26]. No timing data
counts to 95% of the optimum can be estimated. Mekki et al.’s sin- was provided.
gle fin optimization required 60 generations with a population size Comparing the five algorithmic combinations presented in this
of 95, or 5700 total OpenFOAM CFD evaluations, to reach a steady work to the two currently available in the literature, all five outper-
state [23]. form the objective value improvement and do so with significantly
Keramati et al. also used a similar test problem, but with a fewer evaluations of the expensive CFD solvers. Unfortunately, real
slightly different starting design and a deep reinforcement learn- world time to solutions are not available for the other works. The
ing optimization technique with Bézier curves defining the geom- performance improvement, number of design iterations, and the
etry [26]. Instead of OpenFOAM, FEniCS, a Finite Element Method time required to achieve 95% of the optimum are listed in Table 3.

10
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

Table 3
Comparison of optimization results to other works. Note: Assumptions were made and
values extracted from plots to estimate values.

GA FFD PS FFD Mekki [23] Keramati [26]

f ob jective improvement 75 65 45 - %
dP reduction 32 41 -427 60 %
Q improvement 53 38 164 30 %
Evaluations to 95% 742 157 5700 10,000
Time to 95% 1.3 0.2 - - hr

7. Conclusion factured and withstand operation. To move in this direction, fu-

ture work will focus on small, but realistic shape optimization
This paper reports results from three geometry representations problems, working toward optimizing cooling structures in tur-
and two population based metaheuristic optimization algorithms. bine blades. Additionally, other optimization algorithms, machine
A test problem to optimize a single fin of a heat exchanger was learning approaches, and geometry representations such as the
investigated to extend and compare to recent work [23,26]. As de- Harmonic Coordinates mesh deformation technique developed by
scribed by Dbouk [3], some of the major issues facing topology op- Pixar [42] will be investigated to further accelerate the optimiza-
timizations for thermal system design are optimization algorithms, tion process, allowing for the optimization of real designs to be-
3D mesh generation, code parallelization, and overall computa- come a tractable solution.
tional time. This work proposes several modifications to existing
algorithms and workflows to overcome some of these issues in- 8. Disclaimer
cluding:
This project was funded by the United States Department of En-
1. Construction of a client-server infrastructure to maximize ob- ergy, National Energy Technology Laboratory. Neither the United
jective function evaluation throughput on queue managed States Government nor any agency thereof, nor any of their em-
workload systems. ployees, nor the support contractor, nor any of their employees,
2. Smoothing of the binary geometry representation using a level makes any warranty, express or implied, or assumes any legal li-
set to remove stair-stepped geometry. ability or responsibility for the accuracy, completeness, or useful-
3. Reformulation of the GA and PS algorithms to allow for com- ness of any information, apparatus, product, or process disclosed,
pletely asynchronous evaluations of the objective function, or represents that its use would not infringe privately owned
ensuring that all allocated computational resources are fully rights. Reference herein to any specific commercial product, pro-
loaded. cess, or service by trade name, trademark, manufacturer, or oth-
4. Using a solution checking methodology to reject poor quality erwise does not necessarily constitute or imply its endorsement,
meshes and failed convergence with fault tolerant optimization recommendation, or favoring by the United States Government or
algorithms. any agency thereof. The views and opinions of authors expressed
herein do not necessarily state or reflect those of the United States
It is these improvements over the existing literature that di-
Government or any agency thereof.
rectly address issues raised by Dbouk [3], allowing for 210,810
OpenFOAM simulations to be completed in 48 hrs. Declaration of competing interest
The optimization campaigns found the best design to be much
thinner and extend between opposite corners of the design space, The authors declare that they have no known competing finan-
increasing surface area and heat transfer while simultaneously re- cial interests or personal relationships that could have appeared to
ducing the pressure drop. The best design was found using the FFD influence the work reported in this paper.
geometry technique combined with the GA algorithm, increasing
the objective function compared to the baseline design 175%. The Data availability
PS algorithm converged the fastest, while the GA did a better job
exploring the design space. Finally the FFD technique was the best Data will be made available on request.
geometry method, resulting in the best design (with the GA) and
the fastest convergence (with the PS). CRediT authorship contribution statement
Further, the five combinations of optimization algorithms and
geometry representations outperformed similar works in the avail- Justin Weber: Conceptualization, Methodology, Software, Writ-
able literature, finding better designs with higher objective values ing – original draft. E David Huckaby: Methodology, Software,
and doing so with fewer evaluations of the expensive CFD objec- Writing – review & editing. Douglas Straub: Writing – review &
tive evaluations [23,26]. editing.
With modern parallel computing infrastructure, both local clus-
ters and cloud resources, optimization of a single heat exchanger Acknowledgements
fin is a tractable problem. The techniques here generated improved
designs in less than 48 hrs, all reaching 95% of the final optimal This work was performed in support of the U.S. Department of
design in just over 4 hrs. For this specific problem, the shape op- Energy’s Fossil Energy and Carbon Management’s Program and ex-
timization process is fast enough to be integrated into the design ecuted through the National Energy Technology Laboratory (NETL)
process, resulting in higher performing systems. Research & Innovation Center’s Turbines Program.
However the example problem here is quite small compared
to the eventual goal of this work, optimizing the internal cool- References
ing structures of a complete turbine blade. Further, conjugate heat
[1] S. Colburn, A. Majumdar, Inverse design and flexible parameterization of meta-
transfer, material constraints and manufacturing constraints need optics using algorithmic differentiation, Commun. Phys. 4 (1) (2021) 65, doi:10.
to be incorporated so that the optimal designs can be manu- 1038/s42005- 021- 00568- 6.

11
J. Weber, E.D. Huckaby and D. Straub International Journal of Heat and Mass Transfer 207 (2023) 124003

[2] Z. Nie, T. Lin, H. Jiang, L. B. Kara, Topologygan: Topology optimization using [21] F. Feppon, G. Allaire, C. Dapogny, P. Jolivet, Body-fitted topology optimization of
generative adversarial networks based on physical fields over the initial do- 2D and 3D fluid-to-fluid heat exchangers, Comput. Methods Appl. Mech. Eng.
main, 2020, 10.48550/ARXIV.2003.04685 376 (2021) 113638, doi:10.1016/j.cma.2020.113638.
[3] T. Dbouk, A review about the engineering design of optimal heat trans- [22] M. Yoshimura, K. Shimoyama, T. Misaka, S. Obayashi, Topology optimization of
fer systems using topology optimization, Appl. Therm. Eng. 112 (2017) 841– fluid problems using genetic algorithm assisted by the kriging model, Int. J.
854, doi:10.1016/j.applthermaleng.2016.10.134. https://www.sciencedirect.com/ Numer. Methods Fluids 109 (2017) 514–532, doi:10.1002/nme.
science/article/pii/S135943111632645X [23] B.S. Mekki, J. Langer, S. Lynch, Genetic algorithm based topology optimization
[4] J. Alexandersen, C.S. Andreasen, A review of topology optimisation for fluid- of heat exchanger fins used in aerospace applications, Int. J. Heat Mass Transf.
based problems, Fluids 5 (1) (2020), doi:10.3390/fluids5010029. https://www. 170 (2021) 121002, doi:10.1016/j.ijheatmasstransfer.2021.121002.
mdpi.com/2311-5521/5/1/29 [24] K. Shimoyama, A. Komiya, Multi-objective bayesian topology optimization of
[5] A. Fawaz, Y. Hua, S. Le Corre, Y. Fan, L. Luo, Topology optimization of heat a lattice-structured heat sink in natural convection, Struct. Multidiscip. Optim.
exchangers: a review, ENERGY 252 (2022), doi:10.1016/j.energy.2022.124053. 65 (1) (2022) 1–15, doi:10.10 07/s0 0158- 021- 03092- x.
[6] R.L. Webb, Performance evaluation criteria for use of enhanced heat trans- [25] Voxel-Based Topology Optimization of Heat Exchanger Fins, 2021.
fer surfaces in heat exchanger design, Int. J. Heat Mass Transf. 24 (4) 10.2514/6.2022-2445
(1981) 715–726, doi:10.1016/0017-9310(81)90015-6. http://www.sciencedirect. [26] H. Keramati, F. Hamdullahpur, M. Barzegari, Deep reinforcement learning for
com/science/article/pii/0 01793108190 0156 heat exchanger shape optimization, Int. J. Heat Mass Transf. 194 (2022) 123112,
[7] R.L. Webb, M.J. Scott, A parametric analysis of the performance of internally doi:10.1016/j.ijheatmasstransfer.2022.123112.
finned tubes for heat exchanger application, J. Heat Transfer 102 (1) (1980) [27] E.V. Sikin, A.I. Plis, Handbook on splines for the user, CRC Press, 1995.
38–43, doi:10.1115/1.3244245. [28] T.W. Sederberg, S.R. Parry, Free-form deformation of solid geometric models,
[8] R. Webb, E. Eckert, Application of rough surfaces to heat exchanger design, in: Proceedings of the 13th Annual Conference on Computer Graphics and In-
Int. J. Heat Mass Transf. 15 (9) (1972) 1647–1658, doi:10.1016/0017-9310(72) teractive Techniques, in: SIGGRAPH ’86, Association for Computing Machinery,
90095-6. New York, NY, USA, 1986, p. 151160, doi:10.1145/15922.15903.
[9] D.L. Gee, R.L. Webb, Forced convection heat transfer in helically rib- [29] Pallets, Flask, 2022, https://www.palletsprojects.com/p/flask/.
roughened tubes, Int. J. Heat Mass Transf. 23 (8) (1980) 1127–1136, doi:10. [30] P.T. Inc., Dash, 2022, https://www.dash.plotly.com/.
1016/0017-9310(80)90177-5. http://www.sciencedirect.com/science/article/pii/ [31] M. Mitchell, An introduction to genetic algorithms, MIT Press, 1996.
0017931080901775 [32] Z. Michalewicz, C.Z. Janikow, Genetic algorithms for numerical optimization,
[10] A. Gersborg-Hansen, M.P. Bendsøe, O. Sigmund, Topology optimization of heat Stat. Comput. 1 (2) (1991) 75–91, doi:10.1007/BF01889983.
conduction problems using the finite volume method, Struct. Multidiscip. Op- [33] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceedings of
tim. 31 (4) (2006) 251–259, doi:10.1007/s00158- 005- 0584- 3. ICNN’95 - International Conference on Neural Networks, volume 4, 1995,
[11] G. Marck, M. Nemer, J.L. Harion, Topology optimization of heat and mass trans- pp. 1942–1948 vol.4, doi:10.1109/ICNN.1995.488968.
fer problems: laminar flow, Numer. Heat Transfer, Part B: Fundamentals 63 (6) [34] T.K. Truong, Different transfer functions for binary particle swarm optimization
(2013) 508–539, doi:10.1080/10407790.2013.772001. with a new encoding scheme for discounted {0-1} knapsack problem, Math.
[12] G. Marck, Y. Privat, On some shape and topology optimization problems in Probl. Eng. 2021 (2021) 2864607, doi:10.1155/2021/2864607.
conductive and convective heat transfers, OPT-i 2014 - 1st International Con- [35] P. Virtanen, R. Gommers, T.E. Oliphant, M. Haberland, T. Reddy, D. Courna-
ference on Engineering and Applied Sciences Optimization, Proceedings (2014) peau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S.J. van der Walt,
1640–1657. M. Brett, J. Wilson, K.J. Millman, N. Mayorov, A.R.J. Nelson, E. Jones, R. Kern,
[13] T.V. Oevelen, M. Baelmans, Application of Topology Optimization in a Conju- E. Larson, C.J. Carey, İ. Polat, Y. Feng, E.W. Moore, J. VanderPlas, D. Laxalde,
gate, in: OPT-i 2014-1st International Conference on Engineering and Applied J. Perktold, R. Cimrman, I. Henriksen, E.A. Quintero, C.R. Harris, A.M. Archibald,
Sciences Optimization, 2014, pp. 532–577. A.H. Ribeiro, F. Pedregosa, P. van Mulbregt, SciPy 1.0 Contributors, Scipy 1.0:
[14] E.A. Kontoleontos, E.M. Papoutsis-Kiachagias, A.S. Zymaris, D.I. Papadimitriou, fundamental algorithms for scientific computing in python, Nat. Methods 17
K.C. Giannakoglou, Adjoint-based constrained topology optimization for vis- (2020) 261–272, doi:10.1038/s41592- 019- 0686- 2.
cous flows, including heat transfer, Eng. Optim. 45 (8) (2013) 941–961, doi:10. [36] A.F. Gad, Pygad: An intuitive genetic algorithm python library, 2021, (????).
1080/0305215X.2012.717074. 2106.06158
[15] V. Subramaniam, T. Dbouk, J.L. Harion, Topology optimization of conjugate [37] D.H. Ackley, Empirical demonstrations, Springer US, Boston, MA, 1987,
heat transfer systems: a competition between heat transfer enhancement and pp. 71–102.
pressure drop reduction, Int. J. Heat Fluid Flow 75 (January) (2019) 165–184, [38] S.N.A. Yusof, N.M. Muhammad, W.M.A. Aziz Japar, Y. Asako, C. Hong, L.K. Tan,
doi:10.1016/j.ijheatfluidflow.2019.01.002. N.A. Che Sidik, Validity of performance factors used in recent studies on heat
[16] T. Bruns, Topology optimization of convection-dominated, steady-state heat transfer enhancement by surface modification or insert devices, Int. J. Heat
transfer problems, Int. J. Heat Mass Transf. 50 (15) (2007) 2859–2873, doi:10. Mass Transf. 186 (2022) 122431, doi:10.1016/j.ijheatmasstransfer.2021.122431.
1016/j.ijheatmasstransfer.2007.01.039. [39] J. Fan, W. Ding, J. Zhang, Y. He, W. Tao, A performance evaluation plot of en-
[17] K. Yaji, T. Yamada, S. Kubo, K. Izui, S. Nishiwaki, A topology optimiza- hanced heat transfer techniques oriented for energy-saving, Int. J. Heat Mass
tion method for a coupled thermal fluid problem using level set bound- Transf. 52 (1) (2009) 33–44, doi:10.1016/j.ijheatmasstransfer.20 08.07.0 06.
ary expressions, Int. J. Heat Mass Transf. 81 (2015) 878–888, doi:10.1016/j. [40] Joule 2.0 - hp proliant xl170r/xl190r/dl360 gen10, xeon gold 6148 20c 2.4ghz,
ijheatmasstransfer.2014.11.005. intel omni-path, nvidia tesla p100, 2022, https://www.top500.org/system/
[18] P. Coffin, K. Maute, Level set topology optimization of cooling and heating de- 179599/.
vices using a simplified convection model, Struct. Multidiscip. Optim. 53 (5) [41] M.S. Alnaes, J. Blechta, A.J. J. Hake, B. Kehlet, A. Logg, J.R. C. Richardson,
(2016) 985–1003, doi:10.10 07/s0 0158-015- 1343- 8. M.E. Rognes, G.N. Wells, The FEnics project version 1.5, Arch. Numer. Softw.
[19] M.S. Santhanakrishnan, T. Tilford, C. Bailey, Performance assessment of den- 3 (2015), doi:10.11588/ans.2015.100.20553.
sity and level-set topology optimisation methods for three dimensional heat [42] P. Joshi, M. Meyer, T. DeRose, B. Green, T. Sanocki, Harmonic coordinates for
sink design, J. Algor. Comput. Technol. 12 (3) (2018) 273–287, doi:10.1177/ character articulation, ACM Trans. Graph. 26 (3) (2007) 71es, doi:10.1145/
1748301818779019. 1276377.1276466.
[20] F. Dugast, Y. Favennec, C. Josset, Y. Fan, L. Luo, Topology optimization of ther-
mal fluid flows with an adjoint lattice boltzmann method, J. Comput. Phys. 365
(2018) 376–404, doi:10.1016/j.jcp.2018.03.040.

12