Strategies in Decision Making in a Multiobjective Context: Integration

of DOE, NBI, and CFD in the Optimization of a Centrifugal Fan

Matheus C. Pereira1, Anderson P. Paiva1, Matheus B. Francisco1, Tiago Martins de Azevedo1

1
Industrial Engineering Institute, Federal University of Itajubá
BPS Avenue, 1303, Pinheirinho District, ZIP Code: 37550-903, Itajubá, Minas Gerais, Brazil
[email protected], [email protected], [email protected],
[email protected]

Abstract. This article evaluates the integration of Design of Experiments, Normal Boundary Intersection, and
Computational Fluid Dynamics (CFD) in the fluid dynamic assessment of a centrifugal fan. DOE is used for design
preparation, NBI for multiobjective optimization, and CFD for conducting experiments. The obtained results are
evaluated using Mahalanobis Distance. The proposed methodology is assessed through a case study involving a
centrifugal fan operating under high-speed and high-temperature conditions. The optimizations were confirmed
with high reliability, showing an error margin below 15%. This approach aims to contribute to the production of
high-performance and resource-efficient fans without extensive simulation and prototyping.

Keywords: Mulitobjective Optimization, Normal Boundary Intersection, Design of Experiments, Centrifugal

Fluid Dynamics, Fan.

1 Introduction
Using numerical methods and computational algorithms, Computational Fluid Dynamics (CFD) is employed
to solve problems involving fluids and their effects on solid objects. Despite numerous benefits of computationally
handling fluids without the need for numerous physical prototypes, there are challenges related to high
computational demands [1,2] and the significant time required for complex simulations, leading to long processing
times [3,4]. To avoid conducting numerous simulations without a well-defined methodology, Design of
Experiments (DOE) is used to guide the appropriate simulations, thereby reducing the number of simulations. With
DOE, it is possible to obtain a lot of information with fewer simulations, enabling more robust simulations in less
time and with greater reliability.
Once the simulations are performed, the next step is to carry out multiobjective optimization (MO). In this
case, the Normal Boundary Intersection (NBI) method is used, which handles MO problems with multiple and
often conflicting objectives. To find the appropriate solution, a Pareto Frontier is constructed, showing the set of
solutions. This allows verification of the solutions according to the objectives, and NBI has a strong capacity to
handle numerous objectives [5]. To evaluate the results obtained from NBI, the Mahalanobis Distance (MD) metric
is used, as proposed by Johnson & Wichern [6]. This metric is an "evolution" of the normalized Euclidean distance,
with the addition of the variance-covariance matrix.
With the best solution found, a new CFD simulation is conducted using Ansys software to verify if the result
obtained with NBI is reliable compared to the CFD simulation. The integration of DOE, CFD, and NBI techniques
aims to increase the precision and quality of the results.
The literature addresses the combination of methods such as CFD and MO, as evidenced by various studies,
including the performance and emission evaluation of marine combustion engines [7], combustion evaluation in
ultra-supercritical boilers [8], thermal management system design for batteries [9], efficiency evaluation of

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
Template file for CILAMCE-2024 full-length paper (double-click here to enter the short title of your paper)

centrifugal pumps [10], energy performance evaluation of parabolic trough collectors operating with nanofluids
[11], aircraft engine turbine evaluation [12], increasing wear resistance in turbines due to erosive particles [13],
and the study of airborne infectious disease transmission [14]. The use of NBI for MO with conflicting variables
can be seen in Naves et al. [15] Costa et al. [16,17], Lopes et al [18], Belinato et al. [19], Almeida et al [20], Bacci
et al. [21].
To demonstrate this approach, a case study of a centrifugal fan is used. This fan operates at high temperatures
and speeds within an industrial oven. The material used for the fan is AISI 304. The input variables for this fan are
Number of Blades (NB), Initial Angle (IA), Opening Angle (AO), and Blade Length (BL). The experiment
conducted was a Central Composite Design (CCD) with 24 experiments where conflicting responses, Mass Flow
Rate (MFR) and Torque (T), are evaluated. The goal is to produce a centrifugal fan with the highest MFR and the
lowest T, achieving a trade-off between the evaluated responses.

2 Methodology
Initially, it is crucial to define how the simulations will be conducted. A DOE is used to establish the design,
and the CCD is recommended. This involves determining the number of input factors, central points, axial points,
factorial points, and output variables. With the dependent and independent variables defined, Computer-Aided
Design (CAD) is used to design the centrifugal fan. For this activity, Ansys SpaceClaim is utilized. Following this,
it is necessary to generate the meshes and perform the CFD simulations, both of which are carried out using Ansys
Fluent. The entire process is managed using Ansys Workbench to define and organize the project optimally.
After obtaining the responses from the CFD simulations, the response coefficients are stored for subsequent
individual and MO. The individual optimization of the response variables is then performed, which is beneficial
for constructing the payoff matrix that contains the utopia and pseudo-nadir values related to the responses found.
A weight scheme is employed to evaluate the responses obtained from the NBI. This results in numerous
subproblems, with the number of weight variables corresponding to the number of response variables evaluated.
NBI is used to optimize the response variables, generating multiple optimal and viable solutions according to the
weights assigned. The general formulation of NBI, adapted from Das and Dennis, can be written as:
Max t Min − t
 (x, t )  (x, t )
S .t : Φβ + tnˆ = F (x ) S .t : F (x ) − Φβ − tnˆ = 0
 
 x  x (1)
 g j ( x)  0  g j ( x)  0
 
 h j ( x)  0  h j ( x)  0
 
Where: 𝑡 is a scalar perpendicular to the utopia line, 𝚽 ̅ is the normalized payoff matrix, 𝛃 is the weight vector, 𝐧
̂
̅
is the quasi-normal vector, and 𝐅(𝐱) is the vector of scaled objective functions. Limitations are associated with the
optimization algorithm, as gradient-based algorithms are sensitive to the initial starting point.
For each iteration performed, the MD is used to assess the quality of the results obtained. This allows for
classifying each response with a metric that indicates how good the solution is compared to others. Additionally,
the Pareto Frontier can be plotted to observe the behavior of two output variables, showing how one variable
changes with modifications to the other.
Once the optimal results are obtained, Ansys is used again to confirm the optimal result found. This involves
repeating the steps of geometry construction, meshing, and simulation to verify the accuracy of the results obtained
using the method.
Using this methodology, it is possible to verify the entire process, from experimental planning to validation
of the results obtained. This method is promising for product development, allowing for pre-construction
verification before prototyping, thereby saving time and resources. Figure 1 illustrates the step-by-step process.

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
 F. Author, S. Author, T. Author (double-click to edit author field)

Figure 1. Methodology framework

3 Case Study
To implement the proposed method and verify the benefits and quality of the obtained results, a case study
of an industrial centrifugal fan designed for air circulation in an oven is conducted. This fan is made from AISI
304 material, with a density of 7.80 g/m³, specific heat of 0.50 J/kg°C, and thermal conductivity of 16.20 W/m-K.
It operates at a rotational speed of 5,500 rpm.
To ensure the quality of the generated meshes, the criteria used are skewness being less than 0.950 and
orthogonal quality being greater than 0.150. To verify the convergence of the solutions during the simulation
iterations, equations of continuity, velocity in the x, y, and z axes, and turbulence (k and Ω) are considered.
Simulations were performed on 24 different designs, covering the entire process from geometry construction
to obtaining results. The dependent variables of the experiment are: Number of Blades (NB), Initial Angle (IA),
Opening Angle (OA), and Blade Length (BL). The independent variables are: Mass Flow Rate (MFR), the MFR
of the air (study fluid) moving through the fan's cross-section per unit of time, and Torque (T), the amount of
rotational force required to turn the fan blades.
The variables MFR and T are highly correlated, with a positive correlation of 0.876 and a covariance of
0.052. The optimization direction for MFR, given in kg/s, is maximization, while for T, given in N∙m, it is
minimization. The R² of the response surface regression model is 0.738 for MFR and 0.801 for T, with standard
deviations of 0.043 for MFR and 0.802 for T.
To facilitate understanding, Table 1 presents the results obtained from the DOE simulations, including the
coefficients that represent the contribution of each term to the total value of the objective function. Figure 2 shows
the dendrogram of clusters using the Ward Linkage Method and Euclidean distance measure, illustrating how the
experiments cluster into five groups. Contour plots are provided, showing the velocity magnitude of the fan as
represented by a central plane, with all experiments depicted on the same scale.

Figure 2. Cluster Dendrogram of Experiments

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
Template file for CILAMCE-2024 full-length paper (double-click here to enter the short title of your paper)

Table 1. Input and response variables and coefficients

NB IA OA BL MFR T COEF( MFR) COEF (T)
6 0.00 0.00 26.00 0.214 1.499 0.270 3.358
14 0.00 0.00 26.00 0.266 2.353 0.026 0.602
6 30.00 0.00 26.00 0.219 2.155 -0.005 0.402
14 30.00 0.00 26.00 0.258 3.509 0.012 0.266
6 0.00 90.00 26.00 0.223 1.681 0.011 0.161
14 0.00 90.00 26.00 0.313 3.308 0.016 0.145
6 30.00 90.00 26.00 0.232 2.412 -0.009 -0.237
14 30.00 90.00 26.00 0.339 5.180 -0.013 -0.355
6 0.00 0.00 40.00 0.245 1.850 -0.004 -0.128
14 0.00 0.00 40.00 0.279 2.408 0.000 0.129
6 30.00 0.00 40.00 0.246 2.873 0.018 0.370
14 30.00 0.00 40.00 0.213 2.389 -0.007 -0.170
6 0.00 90.00 40.00 0.263 2.249 0.011 0.192
14 0.00 90.00 40.00 0.320 3.416 -0.002 -0.008
6 30.00 90.00 40.00 0.263 3.158 0.006 0.127
14 30.00 90.00 40.00 0.387 5.793
18 15.00 45.00 33.00 0.354 4.715
10 -15.00 45.00 33.00 0.255 1.961
10 45.00 45.00 33.00 0.181 2.428
10 15.00 -45.00 33.00 0.226 2.169
10 15.00 135.00 33.00 0.176 1.279
10 15.00 45.00 19.00 0.210 2.175
10 15.00 45.00 47.00 0.270 3.092
10 15.00 45.00 33.00 0.270 3.358

4 Results

4.1 Individual Optimization

To understand how the variables behave when another is optimized, a Payoff Matrix is constructed, as shown
in Table 2. The main diagonal displays the utopia values for each response variable, while the secondary diagonal
shows the pseudo-nadir values. For conflicting variables, such as MFR and T, as one variable approaches its
optimum, the other will consequently move away from its optimum. This helps facilitate the MO stage. It is
important to note that the algorithm used for the optimizations was the Generalized Reduced Gradient (GRG).

Table 2. Payoff matrix

MFR T
MFR 0.400 0.194
T 5.509 1.061

4.2 Normal Boundary Intersection

To generate optimal Pareto solutions and evaluate results in scenarios with conflicting variables, the NBI
method is used. This method allows decision-makers to have a grid of options with a low computational cost,
especially compared to exhaustive CFD simulations, and aids in understanding the optimal solutions obtained.
A weight increment (w) of 5% was assigned in this process, with each weight associated with a response.

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
 F. Author, S. Author, T. Author (double-click to edit author field)

The grid of points was generated using a DOE Simplex-Lattice, where two weights are combined for each
objective. This results in 21 optimal and viable solutions. To verify the quality of these optimal solutions, the MD
is used to determine the distance between a point and the set of points found, taking into account the correlation
and covariance between variables. MD surpasses the Normalized Euclidean Distance (NED) because the best
results given by NED tend to zero out some of the weights, turning the problem from multiobjective to single-
objective, which is not the evaluation's purpose.
In gradient-based algorithms, the initial point influences the outcomes. Therefore, after individual
optimization, an average of the controllable variables was calculated and set as the initial point for each NBI
iteration.
Pareto Frontier represents the set of optimal and viable solutions in the context of objective optimization,
showing that one objective cannot be improved without worsening another. The results obtained with NBI are
illustrated in Figure 3, which shows the generated Pareto Frontier.
Using the MD metric to evaluate the best results, Table 3 displays the input configurations that should be
used to obtain the top three MD metrics. The goal is to optimize both variables by minimizing the distance to the
utopia point.

Figure 3. Pareto front of mass flow rate and torque responses

Table 3. Best results with Mahalanobis distance

w1 w2 NB IA OA BL MD
0.350 0.650 4 -3.70 28.85 35.87 0.625
0.300 0.700 6 -7.65 25.42 37.44 0.710
0.750 0.250 17 0.28 56.21 34.93 0.930

It is observed that the two smallest MD values prioritize the weight w2, which corresponds to T, while the
third smallest MD result prioritizes the weight w1, corresponding to the MFR. Thus, it is noted that the solution
favoring MFR has a higher NB and an IA, while the other variables do not significantly change relative to the
responses obtained.
By evaluating the standardized effects and performing an analysis of variance (ANOVA), the importance of
the variables is determined. Considering p-values less than 0.050, MFR shows greater importance for MFR (p-

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
Template file for CILAMCE-2024 full-length paper (double-click here to enter the short title of your paper)

value of 0.036), while T shows greater importance for NB (p-value of 0.013) and IA (p-value of 0.037).

4.3 Results of Computational Fluid Dynamics

After completing the optimization process and determining which solution to replicate, the results are
validated using Ansys software. A comparison is made between the optimization results and those obtained from
the simulation. Figure 4 illustrates the geometry, meshes, and contour lines related to particle movement. The
closer the value is to red, the higher it is, while the closer it is to blue, the lower it is.

Figure 4. Geometry, mesh, and pathlines of the ideal centrifugal fan

The results for MFR and T after 1,000 iterations and achieving the defined convergence criteria are as
follows: the MFR result was 0.238 kg/s, while the T was 2.214 N∙m. Both results were very similar to those
provided by the NBI method, achieving an accuracy of over 85.00% for both responses.
The variables have opposing optimization directions and natural dependencies, meaning that improving one
may worsen the other. Therefore, it is crucial to clearly define the design and performance objectives of the product
to assist in the decision-making process, which can impact the optimization process.
The replicated fan has 4 blades, an initial angle of -3.70º, an opening angle of 28.85º, and a blade length of
35.87 mm. The blade thickness and width are fixed variables with respective values of 0.80 mm and 44.60 mm.

5 Conclusions
This article aims to combine DOE techniques to plan and conduct experiments, and NBI for MO, utilizing
CFD for fluid dynamics simulations. This approach is used to simulate the experiments generated by the design
and to confirm the optimal and viable result with the lowest MD value.
By employing and integrating these techniques, it becomes possible to identify the best parameters, determine
the factors that have the most influence on the responses, and simultaneously optimize the criteria. This creates a
robust method for achieving these objectives, and the proposed approach is applied in a case study involving a
centrifugal fan to verify the precision of the method.
In the case study, the proposed method performed well, as the results obtained from validation were accurate
and showed better-than-expected performance. This validation can lead to improved fan performance without the
need for numerous simulations or physical prototypes, thus saving resources and time.
For future work, the initialization of the algorithm will be carried out with multiple starting points to
understand how this influences the algorithm's performance. Additionally, optimization will be performed using
the Non-dominated Sorting Genetic Algorithm II (NSGA-II) for comparison with the NBI.

CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024
 F. Author, S. Author, T. Author (double-click to edit author field)

Acknowledgements. Thanks are expressed to CAPES, CNPq, and FAPEMIG for the support provided to this
work. This research was also made possible by the support of NOMATI-UNIFEI, which provided access to their
laboratories, materials, and expertise.

Authorship statement. The authors hereby confirm that they are the sole liable persons responsible for the
authorship of this work, and that all material that has been herein included as part of the present paper is either the
property (and authorship) of the authors, or has the permission of the owners to be included here.

References
[1] Raman V, Hassanaly M. Emerging trends in numerical simulations of combustion systems. Proc Combust Inst
2019;37:2073–89. https://doi.org/10.1016/j.proci.2018.07.121.
[2] Li G, Wang H, Zhang M, Tupin S, Qiao A, Liu Y, et al. Prediction of 3D Cardiovascular hemodynamics before and
after coronary artery bypass surgery via deep learning. Commun Biol 2021;4:99. https://doi.org/10.1038/s42003-020-01638-
1.
[3] Liu Y, Lu Y, Wang Y, Sun D, Deng L, Wang F, et al. A CNN-based shock detection method in flow visualization.
Comput Fluids 2019;184:1–9. https://doi.org/10.1016/j.compfluid.2019.03.022.
[4] Casas CQ, Arcucci R, Wu P, Pain C, Guo YK. A Reduced Order Deep Data Assimilation model. Phys D Nonlinear
Phenom 2020;412:132615. https://doi.org/10.1016/j.physd.2020.132615.
[5] Das I, Dennis JE. Normal-Boundary Intersection: A New Method for Generating the Pareto Surface in Nonlinear
Multicriteria Optimization Problems. SIAM J Optim 1998;8:631–57. https://doi.org/10.1137/S1052623496307510.
[6] Johnson RA, Wichern DW. Applied multivariate statistical analysis. 1982.
[7] Tan D, Wu Y, Lv J, Li J, Ou X, Meng Y, et al. Performance optimization of a diesel engine fueled with
hydrogen/biodiesel with water addition based on the response surface methodology. Energy 2023;263:125869.
https://doi.org/10.1016/j.energy.2022.125869.
[8] Shi Y, Zhong W, Chen X, Yu AB, Li J. Combustion optimization of ultra supercritical boiler based on artificial
intelligence. Energy 2019;170:804–17. https://doi.org/10.1016/j.energy.2018.12.172.
[9] Li W, Peng X, Xiao M, Garg A, Gao L. Multiobjective design optimization for mini-channel cooling battery
thermal management system in an electric vehicle. Int J Energy Res 2019;43:3668–80. https://doi.org/10.1002/er.4518.
[10] Wang C-N, Yang F-C, Nguyen VTT, Vo NTM. CFD Analysis and Optimum Design for a Centrifugal Pump Using
an Effectively Artificial Intelligent Algorithm. Micromachines 2022;13:1208. https://doi.org/10.3390/mi13081208.
[11] Abubakr M, Amein H, Akoush BM, El-Bakry MM, Hassan MA. An intuitive framework for optimizing energetic
and exergetic performances of parabolic trough solar collectors operating with nanofluids. Renew Energy 2020;157:130–49.
https://doi.org/10.1016/j.renene.2020.04.160.
[12] Wang Y, Liu T, Zhang D, Xie Y. Dual-convolutional neural network based aerodynamic prediction and
multiobjective optimization of a compact turbine rotor. Aerosp Sci Technol 2021;116:106869.
https://doi.org/10.1016/j.ast.2021.106869.
[13] Aponte RD, Teran LA, Grande JF, Coronado JJ, Ladino JA, Larrahondo FJ, et al. Minimizing erosive wear through
a CFD multiobjective optimization methodology for different operating points of a Francis turbine. Renew Energy
2020;145:2217–32. https://doi.org/10.1016/j.renene.2019.07.116.
[14] Arjmandi H, Amini R, khani F, Fallahpour M. Minimizing the respiratory pathogen transmission: Numerical study
and multiobjective optimization of ventilation systems in a classroom. Therm Sci Eng Prog 2022;28:101052.
https://doi.org/10.1016/j.tsep.2021.101052.
[15] Naves FL, Paula TI de, Balestrassi PP, Braga WLM, Sawhney RS, Paiva AP de. Multivariate Normal Boundary
Intersection based on rotated factor scores: A multiobjective optimization method for methyl orange treatment. J Clean Prod
2017;143:413–39. https://doi.org/10.1016/j.jclepro.2016.12.092.
[16] Costa DMC, Brito TG, Paiva AP de, Leme RC, Balestrassi PP. A normal boundary intersection with multivariate
mean square error approach for dry end milling process optimization of the AISI 1045 steel. J Clean Prod 2016;135:1658–
72. https://doi.org/10.1016/j.jclepro.2016.01.062.
[17] Costa DMD, Paula TI, Silva PAP, Paiva AP. Normal boundary intersection method based on principal components
and Taguchi’s signal-to-noise ratio applied to the multiobjective optimization of 12L14 free machining steel turning process.
Int J Adv Manuf Technol 2016;87:825–34. https://doi.org/10.1007/s00170-016-8478-7.
[18] Lopes LGD, Brito TG, Paiva AP, Peruchi RS, Balestrassi PP. Robust parameter optimization based on multivariate
normal boundary intersection. Comput Ind Eng 2016;93:55–66. https://doi.org/10.1016/j.cie.2015.12.023.
[19] Belinato G, Almeida FA de, Paiva AP de, Gomes JH de F, Balestrassi PP, Rosa PARC. A multivariate normal
boundary intersection PCA-based approach to reduce dimensionality in optimization problems for LBM process. Eng
Comput 2019;35:1533–44. https://doi.org/10.1007/s00366-018-0678-3.
[20] Almeida FA, Santos ACO, Paiva AP de, Gomes GF, Gomes JH de F. Multivariate Taguchi loss function
optimization based on principal components analysis and normal boundary intersection. Eng Comput 2022;38:1627–43.
https://doi.org/10.1007/s00366-020-01122-8.
[21] Bacci LA, Mello LG, Incerti T, Paiva AP de, Balestrassi PP. Optimization of combined time series methods to
forecast the demand for coffee in Brazil: A new approach using Normal Boundary Intersection coupled with mixture designs
of experiments and rotated factor scores. Int J Prod Econ 2019;212:186–211. https://doi.org/10.1016/j.ijpe.2019.03.001.
CILAMCE-2024
Proceedings of the joint XLV Ibero-Latin-American Congress on Computational Methods in Engineering, ABMEC
Maceió, Brazil, November 11-14, 2024