Professional Documents
Culture Documents
Measurement
journal homepage: www.elsevier.com/locate/measurement
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.
⁎
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. Measurement 116 (2018) 495–506
Therefore, in this work, the experimental study was firstly con- each input parameters on the two mechanical properties of FDM
ducted by deploying the Stratasys FDM technology to fabricate hex- printed structures. The experimental finding of the present study was
agonal honeycomb structures with different cell sizes and wall thick- then validated by the obtained best models.
ness (input parameters). Based on the different values of these two
input parameters, the two mechanical properties such as the yield 2. FDM experimental details
strength and stiffness (modulus of elasticity) of FDM printed structures
was measured. Further, three CI methods such as GP, ANS and RSR 2.1. The experimental details of the FDM process
were applied for modeling the two mechanical properties (output) of
FDM fabricated cellular structures. 2-D. 3-D surface and sensitivity An essential procedure of FDM is the nozzle depositing heated
analysis were performed on the best method to study the main effects of thermoplastic sequentially on the base plate or previously re-melted
496
B. Panda et al. Measurement 116 (2018) 495–506
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 width and air gap were fixed at their most efficient levels. Only para-
to the previous layer and finally solidify. This process goes on till the meters related to hexagonal honeycomb structure such as cell size and
complete part body gets printed. In the present work, Dimension SST wall thickness were varied at regular interval to study their effect on
1200es from Stratasys® was used to produce honeycomb cellular yield strength and modulus of elasticity followed by model formulation
structure directly from the computer aided design (CAD) models with using CI methods. The complete process of experimentation and mod-
different cell sizes and wall thickness. All the machine related process eling is shown in Fig. 1.
parameters such as layer thickness, orientation, raster angle, raster
497
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.
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
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
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
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
501
B. Panda et al. Measurement 116 (2018) 495–506
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. Measurement 116 (2018) 495–506
Fig. 8. Relationships between the modulus of elasticity and design variables of 3-D printed honeycomb cellular structures.
503
B. Panda et al. Measurement 116 (2018) 495–506
varying both the inputs simultaneously. Plots resulting from the 2D and degree of variation in yield strength and modulus of elasticity can be
3D surface analysis is shown in Figs. 8 and 9. Fig. 8 clearly shows that obtained. It was found from Fig. 10 that the values of yield strength and
the increase in cell size decreases the yield strength and modulus of modulus of elasticity have a non-linear increase to a saturation point of
elasticity values while the increment in wall thickness results in non- wall thickness of 3 mm and then it starts decreasing after this point.
linear increase of yield strength and modulus of elasticity. Interaction Thus, for achieving the higher strength and modulus of elasticity of the
effects of the inputs on the yield strength and modulus of elasticity 3-D printed honeycomb cellular structures, the values for the cell size
(Fig. 9) clearly shows the non-linear complex nature of the process. It should be kept as low as 4 mm while the wall thickness should be kept
implies that by varying both the inputs simultaneously, the greater around 3 mm.
504
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
Fig. 10. Pareto front of Optimization of (a) Yield strength and (b) Modulus of elasticity.
6. Conclusions
Appendix A
505
B. Panda et al. Measurement 116 (2018) 495–506
506