Measurement 116 (2018) 495–506

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Experimental and numerical modelling of mechanical properties of 3D T

printed honeycomb structures
⁎
Biranchi Pandaa, Marco Leitea, Bibhuti Bhusan Biswalb, Xiaodong Niuc, Akhil Gargc,
a
IDMEC, Instituto Superior Técnico, Universidade de Lisboa, Av. Rovisco Pais, 1040-001 Lisboa, Portugal
b
Department of Industrial Design, National Institute of Technology, Rourkela, India
c
Intelligent Manufacturing Key Laboratory of Ministry of Education, Shantou University, Shantou, China

A R T I C L E I N F O A B S T R A C T

Keywords: In recent years, 3-D printing experts have laid emphasis on designing and printing the cellular structures, since
3D printing (3DP) the key advantages (high strength to weight ratio, thermal and acoustical insulation properties) offered by these
Cellular structures structures makes them highly versatile to be used in aerospace and automotive industries. In the present work,
Computational intelligence (CI) an experimental study is firstly conducted to study the effects of the design parameters (wall thickness and cell
Mechanical Properties
size) on the mechanical properties i.e yield strength and modulus of elasticity (stiffness) of honeycomb cellular
structures printed by fused deposition modelling (FDM) process. Further, three promising numerical modelling
methods based on computational intelligence (CI) such as genetic programming (GP), automated neural network
search (ANS) and response surface regression (RSR) were applied and their performances were compared while
formulating models for the two mechanical properties. Statistical analysis concluded that the ANS model per-
formed the best followed by GP and RSR models. The experimental findings were validated by performing the 2-
D, 3-D surface analysis on formulated models based on ANS.

1. Introduction the fabricated prototype can improve [7–10].

With the increasing demand for sustainable product development,
scientists have shifted their focus on producing cellular structures
having minimum weight with high mechanical strength. In this context,
conventional manufacturing operations such as casting, machining,
forming may not be suitable because it takes longer time and need
specific tooling for fabrication of these complex structures [1–3]. Over
the years, these manufacturing operations have evolved into additive
manufacturing (AM) process, which has been proven to be a suitable
fabrication technology for designing the complex shaped parts in short
span of time [4,5]. Among AM processes, fused deposition modeling
(FDM) is a popular extrusion based process which fabricates functional
prototypes by depositing layer by layer of polymer materials inside a
temperature controlled environment [6]. It is important to realize the
relations between the process parameters and characteristics of the
FDM printed prototypes, which could be useful in improving the
functional quality of the prototypes. One way to achieve this is to de-
velop new materials having superior characteristics than conventional
materials, for which complete understanding of material science is
necessary and another convenient approach is by suitably adjusting the
process parameters during the fabrication stage so that the properties of
Literature reveals that the proper adjustment of input parameters
(related to both machine and geometry) can improve the properties of
the FDM functional prototypes [11–14]. Many researchers have de-
ployed physics-based models to improve the properties cellular proto-
type [15–19], but their formulation needs in-depth knowledge of the
process, which may be a cumbersome task. One promising alternative
to this could be to drive the formulation of models from only the given
data and perhaps this could be an interesting means of understanding
the complex and dynamic nature of the process. In this context, several
well-known computational intelligence (CI) methods such as genetic
programming (GP), automated neural networks search (ANS), fuzzy
logic and support vector regression can be applied for formulating
models of the properties of the 3-D printed parts [20–24]. Other class of
methods include the statistical methods such as response surface re-
gression (RSR) can also be applied in circumstance of smaller set of data
samples. Due to complex mechanism of the FDM process, the pre-as-
sumption of the model structure is always a difficult task. To the best of
author’s knowledge there is hardly any research that focuses on the
combined experimental and numerical investigation in developing
functional expressions/models for studying the mechanical of the FDM
printed honeycomb cellular structures.

⁎
Corresponding author.
E-mail address: akhil@stu.edu.cn (A. Garg).

https://doi.org/10.1016/j.measurement.2017.11.037
Received 25 July 2017; Received in revised form 8 October 2017; Accepted 13 November 2017
Available online 16 November 2017
0263-2241/ © 2017 Elsevier Ltd. All rights reserved.
B. Panda et al.
Therefore, in this work, the experimental study was firstly con-
ducted by deploying the Stratasys FDM technology to fabricate hex-
agonal honeycomb structures with different cell sizes and wall thick-
ness (input parameters). Based on the different values of these two
input parameters, the two mechanical properties such as the yield
strength and stiffness (modulus of elasticity) of FDM printed structures
was measured. Further, three CI methods such as GP, ANS and RSR
were applied for modeling the two mechanical properties (output) of
FDM fabricated cellular structures. 2-D. 3-D surface and sensitivity
analysis were performed on the best method to study the main effects of
496
Measurement 116 (2018) 495–506
Fig. 1. Process flow for modelling mechanical proper-
ties of honeycomb cellular structure.
each input parameters on the two mechanical properties of FDM
printed structures. The experimental finding of the present study was
then validated by the obtained best models.
2. FDM experimental details
2.1. The experimental details of the FDM process
An essential procedure of FDM is the nozzle depositing heated
thermoplastic sequentially on the base plate or previously re-melted
B. Panda et al.
Fig. 2. (a) 3D CAD models and (b) process of generating honeycomb cellular structure.
layer as specified in the digital model. As each layer is extruded, bonds
to the previous layer and finally solidify. This process goes on till the
complete part body gets printed. In the present work, Dimension SST
1200es from Stratasys® was used to produce honeycomb cellular
structure directly from the computer aided design (CAD) models with
different cell sizes and wall thickness. All the machine related process
parameters such as layer thickness, orientation, raster angle, raster
497
Measurement 116 (2018) 495–506
width and air gap were fixed at their most efficient levels. Only para-
meters related to hexagonal honeycomb structure such as cell size and
wall thickness were varied at regular interval to study their effect on
yield strength and modulus of elasticity followed by model formulation
using CI methods. The complete process of experimentation and mod-
eling is shown in Fig. 1.
B. Panda et al. Measurement 116 (2018) 495–506

Table 1 generations, and the objective function. The objective function used in this
Statistics of the experimental data extracted from FDM made honeycomb structrues. study is structural risk minimization (SRM). The genetic operators are used
to generate the new set of population (new generation) from the initial
Parameter Cell size Wall thickness Yield Modulus of
(in mm) (in mm) strength (in elasticity (in population. In this study, the maximum number of simulation runs are kept
MPa) MPa) as the stopping criterion for the GP algorithm. More details about the GP
used can be referred from the study [28–30]. The population size, number
Mean 11.25 1.875 11.20 0.311
of generations, depth of the tree, tournament size, number of iterations set
Standard error 0.89 0.15 0.57 0.02
Median 11.25 1.75 11.10 0.322 at 500, 150, 8, 6 and 10 respectively. The best GP models (Eqs. (A1), (A2) in
Standard 4.361 0.755 2.826 0.576 appendix A) are selected based on minimum root means square error
deviation (RMSE, Eq. (1)) on the training data from all runs (Fig. 5) and the perfor-
Kurtosis −1.280 −1.137 −1.98 −0.74 mance is discussed in the following Section 4.
Skewness 0 0.464 0.075 −0.331
Root mean square error
Minimum 5 1 6.059 0.10
Maximum 17.5 3 16.51 0.48 N
∑ |Mi−Ai |2
i=1
(RMSE ) =
N (1)
2.2. Design of cellular lattice structures
where Mi and Ai are the predicted and actual values respectively, Mi and
The CAD models of the honeycomb cellular structure were gener- Ai are the average values of the predicted and actual respectively and n
ated by repeating the unit cell as shown in Fig. 2(a). CATIA V5®, is the number of training samples.
commercial CAD package was used to generate this structure while the
cells size (5–15 mm) and wall thickness (1–3 mm) were parametrically 3.2. Automated neural network models
controlled to avoid complexity of the design process. Among wide range
of cellular structures, honeycomb pattern was selected in this study Automated neural network search (ANS) method is a popular CI
because of high strength/weight ratio which can be benefit to aerospace method for modeling the processes in the circumstance of partial in-
and automobile industries. [25,26]. Two stage design process was formation about it. The architecture (network) of the ANS consists of
adopted (Fig. 2(b)) to generate honeycomb cellular structure. Using this the three layers such as the input layer, the hidden layer and the output
promising approach, most complex solid objects can be further filled layer. The input layer consists of the number of neurons which is same
with different cellular patterns according to the require cell size or infill as the number of the input variables. The hidden layer consists of the
percentages [27]. number of neurons which is chosen based on the training algorithm.
The output layer consists of the number of neurons which is same as the
2.3. Material and measurement number of outputs. The network chosen for each of the output may be
different since the data possesses different nature of complexity. The
The material used for test specimen fabrication was acrylonitrile activation function in neurons of the hidden and output layers plays an
butadiene styrene (ABS P400) which is manufactured and supplied by important role in prediction of the accurate values [31].
Stratasys® Pvt. Ltd., USA. As a raw material, density of full dense Data was partitioned into training and testing in the standard ratio
ABSP400 was near about 1 gm/cm3. However, experimentally the 4:5. The setting for data partition is shown in Fig. 6(a). ANS metho-
density of honeycomb structures made with ABS is found to be dology automates the selection of the activation function to be required
0.98 gm/cm3. This density was measured according to Archimedes’ in neurons of the hidden and output layer. The activation functions
principle. The deviation between experimental and actual density value used in this work is shown in Fig. 6(b). Two types of ANS methods were
can be explained by the layer by layer deposition strategy of FDM used in this work i.e. Radial basis function (RBF) and Multilayer per-
process, which makes low dense specimens by generating air-gap inside ceptron (MLP) networks. The settings for these two networks are shown
between the layers. in Fig. 6(c). The simulation was run on the STATISTICA 8 software [32]
To determine yield strength (σ ), uniaxial compression tests were and the performance of ANS networks for yield strength and modulus of
carried out using Instron 5582 at 1.0 mm/min loading rate. From the elasticity is given in Tables 2 and 3 respectively. From Tables 2 and 3,
obtained stress- strain graph compressive modulus (∈) was calculated the MLP network 2-4-1 (bold in Table 2) for yield strength and the MLP
for individual experimental runs. In total, 24 data samples were gen- network 2-5-1 1 (bold in Table 3) was chosen for the analysis. The
erated by varying input parameters and their respective outputs were statistical performance of these networks was discussed in Section
recorded in Table 1. 4.where
In this work, 80% of the total samples were chosen randomly for
training the three models while the remaining was used for testing the 1. Performance is coefficient of determination (R2).
robustness of the models. The data points generated from the experi- 2. BFGS is Broyden–Fletcher–Goldfarb–Shanno algorithm for opti-
ment is shown by 3-D surface (Fig. 3). It clearly shows the data pos- mizing networks and updating weights.
sesses higher non-linearity in the process. 3. MLP is multi-layer perceptron ANS type.
4. RBF is radial basis function ANS type.
3. Computational intelligence methods 5. RBFT is Redundant Byzantine Fault Tolerance algorithm.

3.1. Genetic programming (GP) 3.3. Response surface regression (RSR)

The GP belongs to class of evolutionary algorithms. The mechanism of The response surface regression is the polynomial based equation
GP is shown in Fig. 4. The implementation of GP involves firstly the settings whose coefficients are estimated using the orthogonal least squares
of its parameters such as the functional set, terminal set, population size, method [33]. The polynomial of n degree is assumed and is fitted to the

498
B. Panda et al. Measurement 116 (2018) 495–506

Fig. 3. 3-D surface plot of the data points gener-

ated from FDM experiment for yield strength and
modulus of elasticity.

499
B. Panda et al. Measurement 116 (2018) 495–506

Experimental
data

Population
size,
generations,
Settings terminal and
functional set
and objective
function

Genetic
evolution for
new population

Selection and Best model

validation

Fig. 4. Schematic diagram showing the flow of GP approach.

Fig. 6. Settings of ANS includes (a) Data partition (b) Activation functions (c) Types of
ANS used in the study.

Fig. 5. RMSE of best GP models (a) Yield strength (b) Modulus of elasticity for individual
run.

500
B. Panda et al. Measurement 116 (2018) 495–506

Table 2
Performance of ANS models for Yield strength (MPa) prediction.

Network name Training performance Testing performance Training algorithm Error function Hidden activation Output activation

RBF 2-18-1 0.965437 0.998130 RBFT SOS Gaussian Identity

RBF 2-17-1 0.990062 0.974687 RBFT SOS Gaussian Identity
RBF 2-10-1 0.758414 0.981022 RBFT SOS Gaussian Identity
MLP 2-5-1 0.978691 0.967480 BFGS 9 SOS Sine Exponential
MLP 2-27-1 0.958828 0.991690 BFGS 9 SOS Sine Exponential
RBF 2-19-1 0.999203 0.994020 RBFT SOS Gaussian Identity
MLP 2-25-1 0.999730 0.999447 BFGS 105 SOS Tanh Logistic
MLP 2-28-1 0.996733 0.995962 BFGS 64 SOS Logistic Exponential
RBF 2-11-1 0.935041 0.991639 RBFT SOS Gaussian Identity
RBF 2-16-1 0.984672 0.996458 RBFT SOS Gaussian Identity
MLP 2-30-1 0.988363 0.998375 BFGS 49 SOS Sine Tanh
RBF 2-20-1 1.000000 0.998193 RBFT SOS Gaussian Identity
MLP 2-6-1 0.998588 0.997940 BFGS 64 SOS Exponential Exponential
MLP 2-19-1 0.996440 0.997094 BFGS 56 SOS Logistic Exponential
MLP 2-28-1 0.998759 0.999992 BFGS 45 SOS Tanh Exponential
MLP 2-23-1 0.997922 0.995644 BFGS 94 SOS Exponential Logistic
RBF 2-16-1 0.996221 0.970688 RBFT SOS Gaussian Identity
MLP 2-25-1 0.999444 0.999713 BFGS 119 SOS Logistic Tanh
MLP 2-11-1 0.997867 0.997592 BFGS 80 SOS Exponential Logistic
RBF 2-17-1 0.993667 0.985404 RBFT SOS Gaussian Identity
RBF 2-18-1 0.991450 0.999241 RBFT SOS Gaussian Identity
RBF 2-10-1 0.955060 0.971567 RBFT SOS Gaussian Identity
RBF 2-14-1 0.995545 0.986276 RBFT SOS Gaussian Identity
MLP 2-4-1 0.998521 0.999687 BFGS 74 SOS Tanh Tanh
RBF 2-12-1 0.968608 0.974707 RBFT SOS Gaussian Identity
MLP 2-26-1 0.998333 0.999579 BFGS 28 SOS Exponential Exponential
RBF 2-12-1 0.996499 0.999328 RBFT SOS Gaussian Identity
RBF 2-14-1 0.967565 0.979313 RBFT SOS Gaussian Identity
MLP 2-26-1 0.997989 0.998774 BFGS 40 SOS Exponential Exponential
RBF 2-14-1 0.988350 0.966999 RBFT SOS Gaussian Identity

Table 3
Performance of ANS models for modulus of elasticity prediction.

Network name Training performance Testing performance Training algorithm Error function Hidden activation Output activation

MLP 2-14-1 0.958407 0.989770 BFGS 4 SOS Identity Logistic

MLP 2-22-1 0.959378 0.987057 BFGS 4 SOS Identity Logistic
MLP 2-15-1 0.989368 0.997576 BFGS 58 SOS Tanh Logistic
MLP 2-11-1 0.948967 0.990109 BFGS 2 SOS Logistic Sine
RBF 2-6-1 0.973681 0.986178 RBFT SOS Gaussian Identity
RBF 2-17-1 0.984168 0.986492 RBFT SOS Gaussian Identity
RBF 2-16-1 0.983245 0.999548 RBFT SOS Gaussian Identity
MLP 2-19-1 0.993893 0.999927 BFGS 81 SOS Exponential Tanh
RBF 2-18-1 0.958881 0.998799 RBFT SOS Gaussian Identity
RBF 2-17-1 0.978440 0.998058 RBFT SOS Gaussian Identity
MLP 2-7-1 0.989776 0.999991 BFGS 39 SOS Tanh Sine
RBF 2-13-1 0.994932 0.992524 RBFT SOS Gaussian Identity
RBF 2-18-1 0.999389 0.987135 RBFT SOS Gaussian Identity
RBF 2-14-1 0.941495 0.999931 RBFT SOS Gaussian Identity
MLP 2-4-1 0.978678 0.987541 BFGS 36 SOS Exponential Identity
MLP 2-17-1 0.990122 0.994365 BFGS 30 SOS Logistic Identity
RBF 2-17-1 0.982527 0.985336 RBFT SOS Gaussian Identity
RBF 2-12-1 0.969691 0.983399 RBFT SOS Gaussian Identity
RBF 2-15-1 0.952966 0.997159 RBFT SOS Gaussian Identity
MLP 2-23-1 0.987094 0.996138 BFGS 54 SOS Tanh Exponential
RBF 2-15-1 0.969921 0.989991 RBFT SOS Gaussian Identity
MLP 2-14-1 0.993144 0.995406 BFGS 62 SOS Tanh Exponential
RBF 2-6-1 0.709994 0.993906 RBFT SOS Gaussian Identity
MLP 2-15-1 0.987173 0.998901 BFGS 36 SOS Tanh Sine
MLP 2-16-1 0.946139 0.990439 BFGS 4 SOS Identity Identity
MLP 2-14-1 0.985963 0.995489 BFGS 416 SOS Logistic Sine
MLP 2-5-1 0.987210 0.992479 BFGS 18 SOS Tanh Tanh
MLP 2-5-1 0.976965 0.994874 BFGS 11 SOS Exponential Tanh
RBF 2-18-1 0.989251 0.982037 RBFT SOS Gaussian Identity
MLP 2-19-1 0.990335 0.992391 BFGS 49 SOS Exponential Tanh

501
B. Panda et al. Measurement 116 (2018) 495–506

Table 4 Yield Strength (MPa) (Target) vs. Yield Strength (MPa)

Statistical metrics of the three models. (Output)
Samples: Train, Test
R2 RMSE (%) MAPE (%) 18

Yield Strength (MPa) (Output)

17
Models Training Testing Training Testing Training Testing
16
phase phase phase phase phase phase
15
Yield Strength (MPa) 14
GP 0.99 0.95 0.04 0.15 0.34 6.21 13
ANS 0.99 0.99 0.01 0.02 0.02 1.02 12
RSR 0.92 0.86 3.45 7.61 5.34 12.56 11
10
Modulus of elasticity (MPa)
9
GP 0.96 0.93 0.030 0.102 0.23 0.73
ANS 0.98 0.99 0.002 0.001 0.03 0.02
8
RSR 0.94 0.88 2.36 6.52 3.79 8.92 7
6
5
Table 5 4
Best automated neural network model prediction values against the actual output. 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Yield Strength (MPa) (Target)
Sample Cell size Wall thickness Yield strength Yield Strength
(mm) (mm) (Actual) (predicted) MLP 2-4-1 (a)
Train 5 1 13.00011 13.09981
Train 7.5 1 10.00014 9.69620
Train 10 1 8.50017 8.61932
Train 12.5 1 7.50017 7.77484
Train 15 1 6.70020 6.79782
Train 17.5 1 6.05900 5.77503
Test 5 1.5 15.23038 15.16403
Train 7.5 1.5 11.88206 11.97615
Train 10 1.5 10.35333 10.49248
Train 12.5 1.5 9.67496 9.63675
Train 15 1.5 8.78599 8.86551
Train 17.5 1.5 8.06886 8.13634
Train 5 2 15.89000 15.98687
Train 7.5 2 13.89935 13.77446
Train 10 2 12.57939 12.35402
Train 12.5 2 11.41247 11.53826
Test 15 2 10.80520 10.85112
Test 17.5 2 10.23100 10.21210
Train 5 3 16.51017 16.33777
Train 7.5 3 14.48020 14.67792
Train 10 3 13.08021 13.05363
Train 12.5 3 12.14113 12.11377
Test 15 3 11.50829 11.41002
Train 17.5 3 10.70694 10.80869
Fig. 7. Statistical fit of the (a) Yield Strength (b) Modulus of elasticity models on training
Table 6 and testing phase. (*Dots shows the plot between actual and predicted values obtained
Best automated neural network model prediction values against the actual output. from Tables 5 and 6 respectively. Solid Line shows the fit between the predicted and
actual values at 45 degree).
Sample Cell size Wall thickness Modulus of Modulus of elasticity
(mm) (mm) elasticity (Predicted) MLP 2-5-1
(Actual)

Train 5 1 0.370000 0.376225 experimental data measured and discussed in Section 2. The degree for
Train 7.5 1 0.360000 0.336745 which the best fitting (training accuracy) of the polynomial with the
Train 10 1 0.290000 0.295130
actual data was obtained is selected for the analysis. The performance
Train 12.5 1 0.240000 0.254252
Train 15 1 0.200000 0.215956 of the RSR models (Eqs. (2) and (3)) is discussed in Section 4.
Train 17.5 1 0.190230 0.181134
Train 5 1.5 0.398823 0.407553 Yield Strength=11.0100632881 + 8.32880519224
Train 7.5 1.5 0.371959 0.364114
Train 10 1.5 0.294655 0.309454 *"“"Wall thickness(mm)"”"−1.51618707211
Train 12.5 1.5 0.259944 0.250068 *"“"Wall thickness(mm)"”"2−1.27763188127
Train 15 1.5 0.217643 0.192893
Test 17.5 1.5 0.108749 0.142292 *"“"Cell size(mm)"”" + 0.035323974247
Train 5 2 0.450000 0.436796
*"“"Cell size(mm)"”"2 (2)
Test 7.5 2 0.420000 0.406732
Train 10 2 0.350000 0.358582
Test 12.5 2 0.270000 0.291924
Train 15 2 0.210000 0.214849 Modulus of elasticity=0.493523273012−0.0131550858481
Train 17.5 2 0.130000 0.139965 *"“"Cell size(mm)"”"−0.000281639611644
Train 5 3 0.480000 0.447053
Train 7.5 3 0.420000 0.435889 *"“"Cell size(mm)"”"2−0.0612911234404
Train 10 3 0.400000 0.418441
*"“"Wall thickness(mm)"”" + 0.0298364519899
Train 12.5 3 0.380000 0.390500
Train 15 3 0.360000 0.346018 *"“"Wall thickness(mm)"“"2 (3)
Test 17.5 3 0.292938 0.279204

502
B. Panda et al.
Fig. 8. Relationships between the modulus of elasticity and design variables of 3-D printed honeycomb cellular structures.
4. Performance analysis of the best model
In this section, the performance analysis of the three best models
(Eq. (A1), (A2) in the appendix A) formulated from GP, ANS and RSR
was demonstrated against the actual data (Table 1). Table 4 shows the
values of error metrics (R2, MAPE and RMSE) of the three models for
yield strength and modulus of elasticity on the training and testing data.
It clearly shows that all the models have good training accuracy. For the
testing data, among these three models, the ANS models have per-
formed better than those of GP and RSR models with coefficient of
determination obtained as high as 0.99. The predictions obtained for
yield strength and modulus of elasticity from the best networks chosen
for ANS models is given in Tables 5and 6 respectively. It can be seen
that, the predictions obtained from the ANS models are very close to the
actual data on both the training and the testing phase. Similar inference
is drawn from Fig. 7(a) and (b), where the statistical fit of the actual and
predicted values for yield strength and modulus of elasticity was
503
Measurement 116 (2018) 495–506
approximate to ideal one.
The statistical analysis of the models formulated from the three
methods concludes that the ANS models have performed the best fol-
lowed by GP and RSR. It would also be interesting to explore the nature
of the relationships (main and interactive effects) between the process
parameters. Thus, the following section introduces the 2D, 3D surface
and sensitivity analysis for the ANS models.
5. 2-D, 3-D surface and global sensitivity analysis of the ANS
models
The 2-D, 3D surface and sensitivity analysis on the ANS models was
performed based on the procedures including the mathematical for-
mulae discussed in the study by Panda et al [34,35]. The main effect of
the inputs on the yield strength and modulus of elasticity is evaluated
by keeping one of the input at mean value and varying the other. The
interaction effect (3D) between the two inputs was evaluated by
B. Panda et al.
varying both the inputs simultaneously. Plots resulting from the 2D and
3D surface analysis is shown in Figs. 8 and 9. Fig. 8 clearly shows that
the increase in cell size decreases the yield strength and modulus of
elasticity values while the increment in wall thickness results in non-
linear increase of yield strength and modulus of elasticity. Interaction
effects of the inputs on the yield strength and modulus of elasticity
(Fig. 9) clearly shows the non-linear complex nature of the process. It
implies that by varying both the inputs simultaneously, the greater
504
Measurement 116 (2018) 495–506
Fig. 9. Interaction effect of the inputs on the
Yield strength and Modulus of elasticity.
degree of variation in yield strength and modulus of elasticity can be
obtained. It was found from Fig. 10 that the values of yield strength and
modulus of elasticity have a non-linear increase to a saturation point of
wall thickness of 3 mm and then it starts decreasing after this point.
Thus, for achieving the higher strength and modulus of elasticity of the
3-D printed honeycomb cellular structures, the values for the cell size
should be kept as low as 4 mm while the wall thickness should be kept
around 3 mm.
B. Panda et al. Measurement 116 (2018) 495–506

design variables (cell size and wall thickness) on the two mechanical
properties (yield strength and modulus) of 3-D printed honeycomb
structures. Three computational intelligence (CI) methods were applied
and compared while modelling the mechanical properties based on our
experimental findings. The results of statistical validation concluded
that, the ANS models performed the best followed by the GP and RSR
models. The higher generalization ability of the ANS models obtained
from statistical analysis is useful for studying the cellular structure
dynamics in uncertain input conditions. The conducted 2D, 3D surface
and sensitivity analysis validates the robustness of the ANS models by
unveiling the hidden relationships of yield strength and modulus of
elasticity with respect to the two input design variables (cell size and
wall thickness). The generalized ANS models obtained can easily be
optimized analytically in determining the appropriate inputs settings
that maximizes the two mechanical properties of honeycomb structure.
Future work should focus on the study of different type of cellular
structures and printable materials that are subjected to combined
buckling and compression load while changing the default machine
parameters such as air gap, layer thickness, part orientation and raster
width etc.

Acknowledgements

This work was supported by FCT, through IDMEC, under LAETA,

project UID/EMS/50022/2013 and Shantou University Scientific
Research Fund (NTF 16002). This work is also supported by a grant of
an International Science & Technology Collaboration project between
China and Israel (No: 2017A050501061) founded by Guangdong
Science and Technology Department.

Fig. 10. Pareto front of Optimization of (a) Yield strength and (b) Modulus of elasticity.

6. Conclusions

The present study emphasizes the need of investigating the effects of

Appendix A

Unconfined honeycomb strength=21.0038 + (0.07209). ∗ ((((cos(plog((x1). ∗ (x1))))−((7.122794))) + (((tan(x2)) + (tanh(x2))). ∗ (tan(plog(x1)))))

−(cos((cos(exp(x2))). ∗ (x1)))) + (−0.12837). ∗ (cos(((tanh(x1))−(x1)). ∗ (plog(tan(plog(x1)))))) + (−0.075038)
. ∗ (((cos(x1)) + (((tan(x2))+(tanh(plog(x1)))). ∗ (tan(plog(x1)))))
−(exp(plog((4.099477))))) + (−11.7878). ∗ (plog(plog((x1) + (plog(x2))))) + (−4.9613). ∗ (tanh(cos(x2)))
+ (0.0061129). ∗ (((cos(x1)) + (((tan(x2)) + (tanh(tan(x2)))). ∗ (tan(cos(x1)))))−(exp(x2))) (A1)

505
B. Panda et al.
Unconfined honeycomb stiffness=0.53086 + (0.0065228). ∗ (cos((plog(cos(x2)))−(tan(plog(((x2). ∗ ((−7.526176))) + (x1)))))) + (−0.018132).
∗ (cos((plog(tan(tan((x2). ∗ (x1)))))−(tan(x2)))) + (0.0092228). ∗ (tan(tan(cos(exp(x2))))) + (−0.0002591).
∗ (((x2). ∗ (x1)). ∗ (x1)) + (−0.016805).
∗ ((((((x1) + ((8.100013)))−(exp(x2))) + (sin(tanh(x2)))). ∗ ((tanh(sin(x2))). ∗ (exp((x1)(x1))))) + (tanh
(sin(cos((x2) + (x1)))))) + (1.1563e−08). ∗ ((exp(tan(plog(((x2). ∗ ((−7.149202))) + (((6.373906)). ∗ (x1))))
))−(plog((tan((x2). ∗ (x1))). ∗ (x2))));
References
[1] C.K. Chua, K.F. Leong, 3D Printing and Additive Manufacturing: Principles and
Applications, World Scientific Publishing Company, Fifth Edition of Rapid
Prototyping, 2016.
[2] I. Gibson, D. Rosen, B. Stucker, Additive manufacturing technologies: 3D printing,
rapid prototyping, and direct digital manufacturing, Springer, 2014.
[3] L.J. Gibson, M.F. Ashby, Cellular Solids: Structure and Properties, Cambridge
University Press, 1999.
[4] B. Panda, S.C. Paul, L.J. Hui, Y.W.D. Tay, M.J. Tan, Additive manufacturing of
geopolymer for sustainable built environment, J. Clean. Product. 167 (2017)
281–288.
[5] B. Berman, 3-D printing: The new industrial revolution, Business Horizons 55 (2)
(2012) 155–162.
[6] S.H. Masood, W.Q. Song, Development of new metal/polymer materials for rapid
tooling using fused deposition modelling, Mater. Des. 25 (2004) 587–594.
[7] B. Panda, S.C. Paul, N.A.N. Mohamed, Y.W.D. Tay, M.J. Tan, Measurement of
tensile bond strength of 3D printed geopolymer mortar, Measurement 113 (2018)
108–116.
[8] I. Yadroitsev, P. Bertrand, I. Smurov, Parametric analysis of the selective laser
melting process, Appl. Surf. Sci. 253 (19) (2007) 8064–8069.
[9] R. Anitha, S. Arunachalam, P. Radhakrishnan, Critical parameters influencing the
quality of prototypes in fused deposition modelling, J. Mater. Process. Technol. 118
(2001) 385–388.
[10] A.K. Sood, R.K. Ohdar, S.S. Mahapatra, Parametric appraisal of mechanical property
of fused deposition modelling processed parts, Mater. Des. 31 (1) (2010) 287–295.
[11] P.M. Pandey, N. Venkata Reddy, S.G. Dhande, Improvement of surface finish by
staircase machining in fused deposition modeling, J. Mater. Process. Technol. 132
(2003) 323–331.
[12] Z.A. Khan, B.H. Lee, J. Abdullah, Optimization of rapid prototyping parameters for
production of flexible ABS object, J. Mater. Process. Technol. 169 (1) (2005) 54–61.
[13] B.N. Panda, M.R. Bahubalendruni, B.B. Biswal, A general regression neural network
approach for the evaluation of compressive strength of FDM prototypes, Neural
Comput. Appl. 26 (5) (2015) 1129–1136.
[14] B.N. Panda, K. Shankhwar, A. Garg, Z. Jian, Performance evaluation of warping
characteristic of fused deposition modelling process, Int. J. Adv. Manuf. Technol. 88
(5–8) (2017) 1799–1811.
[15] H.S. Byun, K.H. Lee, Determination of the optimal build direction for different rapid
prototyping processes using multi-criterion decision making, Robot. Comput.-
Integrat. Manuf. 22 (2006) 69–80.
[16] D. Ahn, J.H. Kweon, S. Kwon, J. Song, S. Lee, Representation of surface roughness
in fused deposition modeling, J. Mater. Process. Technol. 209 (2009) 5593–5600.
[17] Chunze Yan, et al., Evaluations of cellular lattice structures manufactured using
selective laser melting, Int. J. Mach. Tools Manuf 62 (2012) 32–38.
[18] D.Y. Chang, B.H. Huang, Studies on profile error and extruding aperture for the RP
parts using the fused deposition modelling process, Int. J. Adv. Manuf. Technol. 53
506
Measurement 116 (2018) 495–506
(A2)
(2011) 1027–1037.
[19] F. Brenne, T. Niendorf, H.J. Maier, Additively manufactured cellular structures:
impact of microstructure and local strains on the monotonic and cyclic behaviour
under uniaxial and bending load, J. Mater. Process. Technol. 213 (2013)
1558–1564.
[20] A.K. Sood, R.K. Ohdar, S.S. Mahapatra, Experimental investigation and empirical
modelling of FDM process for compressive strength improvement, J. Adv. Res. 3 (1)
(2012) 81–90.
[21] A. Boschetto, V. Giordano, F. Veniali, Surface roughness prediction in fused de-
position modelling by neural networks, Int. J. Adv. Manuf. Technol. (2013) 67.
[22] C.L. Li, G.Y. Fu, K.B. Guo, Study on forecast of forming temperature of ABS resign
during fused deposition manufacturing by fuzzy comprehensive evaluation, Key
Eng. Mater. 464 (2011) 264–267.
[23] L.N. Carter, K. Essa, M.M. Attallah, Optimisation of selective laser melting for a high
temperature Ni-superalloy, Rapid Prototyping J. 21 (4) (2015) 423–432.
[24] E. Vahabli, S. Rahmati, Application of an RBF neural network for FDM parts’ surface
roughness prediction for enhancing surface quality, Int. J. Precision Eng. Manuf. 17
(12) (2016) 1589–1603.
[25] S.D. Sangle, Design and Testing of Scalable 3D-Printed Cellular Structures
Optimized for Energy Absorption, Doctoral dissertation Wright State University,
2017.
[26] P. Zhang, J. Toman, Y. Yu, E. Biyikli, M. Kirca, M. Chmielus, A.C. To, Efficient
design-optimization of variable-density hexagonal cellular structure by additive
manufacturing: theory and validation, J. Manuf. Sci. Eng. 137 (2) (2015) 021004.
[27] B.N. Panda, Design and Development of Cellular Structure for Additive
Manufacturing (Doctoral dissertation). 2015.
[28] J.R. Koza, Genetic Programming II: Automatic Discovery of Reusable Programs,
1994.
[29] M. Hinchliffe, H. Hiden, B. Mckay, M. Willis, M. Tham, G. Barton, Modell. Chem.
Process Syst. Using A Multi-Gene Genetic Program. Algorithm (1996) 28–31.
[30] D.P. Searson, D.E. Leahy, M.J. Willis, GPTIPS: an open source genetic programming
toolbox for multi gene symbolic regression, Int. Multi Conf. Eng. Comput. Scientists
1 (2010) 77–80.
[31] S.C. Paul, B. Panda, A. Grag, A novel approach in modelling of concrete made with
recycled aggregates, Measurement 115 (2017) 64–72.
[32] Hilbe, M. Joseph, STATISTICA 7. The American Statistician, 2012 (accessed on 28th
August 2016).
[33] A.J. Makadia, J.I. Nanavati, Optimization of machining parameters for turning
operations based on response surface methodology, Measurement 46 (4) (2013)
1521–1529.
[34] B.N. Panda, S.A. Vendan, A. Garg, Experimental-and numerical-based studies for
magnetically impelled arc butt welding of T11 chromium alloy tubes, Int. J. Adv.
Manuf. Technol. 88 (9–12) (2017) 3499–3506.
[35] A. Garg, J.S.L. Lam, B.N. Panda, A hybrid computational intelligence framework in
modelling of coal-oil agglomeration phenomenon, Appl. Soft Comput. 55 (2017)
402–412.