The International Journal of Advanced Manufacturing Technology (2020) 107:4481–4492

https://doi.org/10.1007/s00170-020-05270-5

ORIGINAL ARTICLE

A parametric study of 3D printed polymer gears

Ye Zhang 1 & Ken Mao 1 & Simon Leigh 1 & Akeel Shah 1 & Zhiming Chao 1 & Guotao Ma 1

Received: 12 September 2019 / Accepted: 30 March 2020 / Published online: 28 April 2020
# The Author(s) 2020

Abstract
The selection of printing parameters for 3D printing can dramatically affect the dynamic performance of components such as
polymer spur gears. In this paper, the performance of 3D printed gears has been optimised using a machine learning process. A
genetic algorithm (GA)–based artificial neural network (ANN) multi-parameter regression model was created. There were four
print parameters considered in 3D printing process, i.e. printing temperature, printing speed, printing bed temperature and infill
percentage. The parameter setting was generated by the Sobol sequence. Moreover, sensitivity analysis was carried out in this
paper, and leave-one cross validation was applied to the genetic algorithm-based ANN which showed a relatively accurate
performance in predictions and performance optimisation of 3D printed gears. Wear performance of 3D printed gears increased
by 3 times after optimised parameter setting was applied during their manufacture.

Keywords 3D printing . Nylon . Gears . Wear . Polymer gears . Machine learning

1 Introduction are many investigations into the characteristics of wear and

thermal behaviour of injection-moulded gears. Mao et al. [2]
For applications such as automotive and aerospace engineer- carried out an analysis of the friction and wear behaviour of
ing, polymer gears have unique advantages over metal gears: acetal and nylon gears including characterising the failure
low cost and weight, high efficiency, quietness of operation, mechanism and thermal analysis. The results showed the op-
functioning without external lubrication, etc. The performance erational time of polymer spur gears under different circum-
of 3D printed gear has been investigated previously. stances. Hu and Mao [3] investigated the effects of different
According to Ye et al. [1], 5 different 3D printing nylon ma- misalignments on the fatigue of polymer gears during use.
terials have been compared; result shows Nylon 618 has out- Hooke et al. [4] proved that increases in the surface tempera-
standing performance compared with other nylon materials, ture can dramatically increase the wear rate of the gear tooth.
including 23% carbon fibre reinforced nylon filament. There Moreover, Gauvin et al. [5] carried out an investigation into
the maximum surface temperature experienced by polymer
gears without lubrication. Mao et al. [6] introduced a new
* Ye Zhang
[email protected]
method to predict the surface temperature of acetal gears and
found the correlation between fatigue life and tooth size.
Ken Mao Additive manufacturing (AM) and 3D printing processes have
[email protected] become increasingly popular; the applications of 3D printing
Simon Leigh are usually suitable for relatively low production volumes,
[email protected] small size parts and complex designs. It is generally under-
stood that 3D printing is cost effective if production volumes
Akeel Shah
[email protected] are below 1000 units in comparison with plastic injection
moulding [7]. The technology has been applied in wide range
Zhiming Chao
[email protected]
of industries, including the automotive industry, aerospace,
medical and architectural [8]. There was limited research on
Guotao Ma dynamic performance of 3D printed polymer parts; however,
[email protected]
there are several investigations regarding the parameters
1
School of Engineering, The University of Warwick, Coventry CV4 which affect the mechanical and thermal properties. Chacon
7AL, UK et al. [9] has investigated the effect of process parameters on
4482 Int J Adv Manuf Technol (2020) 107:4481–4492

mechanical performance of PLA in terms of on-edge orienta- 2 Methods and experiments

tion, layer thickness and feed rate. It has been shown that
higher printing speeds can increase the mechanical perfor- All 3D printing parameters were set as default and printed with
mance of printed parts. Giovanni [10] carried out Taguchi’s manufacturer-recommended printing temperature (250 °C), bed
experimental design for fatigue analysis of PLA and claimed temperature (30 °C) and speed (45 mm/s) apart from infill per-
that infill percentage had the most influence on fatigue life. centage, which was set to 60% for both printer systems. Under
Kuznetsov et al. [11] claimed that printing temperature and initial setting which printing temperature was 245°C, printing
printing speed could dramatically dominate the mechanical speed was 45°C, infill percentage was 60% and bed temperature
properties of the 3D printed part. Moreover, the thermal con- was 45°C gear running under 10 Nm torque and last around 16 h.
ductivity of 3D printing filaments can also affect the There are several stages required in order to complete dynamic
properties of the object [12], increasing or decreasing performance optimisation of 3D printed gears. First of all, the use
the bonding quality between each layer during fused of a Sobol sequence was employed to generate Sobol random
deposition modelling [13, 14]. In order to understand parameters with multiple data points per parameter. The Sobol
the complicated interplay between these different pro- sequence is a low discrepancy quasi-random sequence. Using
cess parameters and to select the most appropriate pa- this method, a set of test data comprising of 50 points was gen-
rameter set for the production of 3D printed gears, a erated, including a printing temperature range (230–275 °C),
multiple regression process is required. printing speed range (20–75 mm/s), infill percentage (20–80%)
There was very limited research regarding of machine and bed temperature range (30–70 °C). Bed temperature refers to
learning associated with predicting performance of gears and the temperature of printing surface which will affect the first few
only some on its application to 3D printing processes. Fracture layers during printing. Infill percentage represents how ‘hollow’
behaviour of 3D printed material has shown dramatically dif- the gear is, with the aim of reducing the infill percentage and
ferent compared with other materials [15]. Deng et al. [16] hence reducing the weight and inertia during operation. Each
introduced optimisation methods to the multi-factor printing parameter was increased by factor of one, for example printing
of a ceramic slurry by using artificial neural networks. Koeppe temperature was increased from 230 to 275 °C and the Sobol
et al. [17] used neural networks to analyse load distribution in sequence covers the entire range, hence there was a total of 45
3D printed lattice cell structures. Delli and Chang [18] used data points for printing temperature. Figure 1 shows the specifi-
supervised machine learning to do real time monitoring of 3D cation of gears.
printing to eliminate printing time and waste. Those research Furthermore, by applying a similar range for each parameter,
reports have provided valuable results in terms of static force a small number of experiments could potentially provide insights
analysis and monitoring of the 3D printing; however, dynamic into the complex combination of each test data category, roughly
analysis of 3D printed parts require further investigations. Li reducing a total of 45 × 45 × 40 × 60 = 5,940,000 possible com-
et al. [19] has introduced a method using support vector ma- binations of 3D printing settings to a sample set of 50 chosen by
chine to predict dynamic contact characteristics for helical the Sobol sequence. Gears were produced using an Ultimaker 3
gears. Sun et al. [20] used neural networks to optimise and extended fused deposition modelling (FDM) system. Gears were
predict a gear hobbing process to improve the efficiency and printed on a tufnol print surface due to superior adhesion between
reduce the cost. Sun et al. [21] used artificial neural networks the nylon and tufnol, eliminating the peel off during the 3D
and support vector machines with genetic algorithms to mon- printing process. A total of 100 gears were printed (50 matched
itor the faults in gears. To find the correlation of 3D printing pairs) with an average printing time of around 6 h per gear
process parameters and dynamic performance of polymer (depending on the setting of the parameters). After printing, gears
gears would be a significant benefit to researchers both in were mounted on a gear wear test rig subject to the gear perfor-
the fields of 3D printing and gear manufacture to increase mance life cycle with 10 Nm torque. The tests included the
the efficiency of the 3D printing process and quality of the recording of the wear occurring at the gear tooth and showed
resultant 3D printed spur gears. the different stages of gear operation until gear failure. The time
Performing multi-parameter regression has many chal- from commencing the test run on the test rig until the gear failure
lenges, for example missing data and data noise, as well as was considered as the fatigue time. After acquiring the test data
high dimensionality which impacts the ability to identify the for the printed gears, the 3D printing parameters were used as
relations between parameters [22]. Through ordinary mathe- input, and the corresponding gear life cycle data from test rig was
matical solutions, it is computationally complex to solve used as an output to create a neural network model of correlations
multi-target modelling, and targets often may not corre- between input and output. The Gaussian process was also
late. However, by using some baseline methods, such as employed to perform multi-parameter regression to find out the
Gaussian processes, neural networks or support vector approximate likelihood of output accuracy. By using the model
machines, the complexity of the problem can be much generated by the artificial neural network (ANN) and GP, a sub-
reduced [23]. sequent sensitivity analysis was carried out to investigate the
Int J Adv Manuf Technol (2020) 107:4481–4492 4483

Fig. 1 a 3D printed nylon gear Module 2mm

cross section in simplified 3D. b
Specifications of gears Tooth number 30
Pressure angle 20°
Face width 15mm
Nominal backlash 0.18mm
Tooth thickness 3.14mm
Contact rao 1.65
(a) (b)

relations of each multi-parameter. The process was shown in Fig. 4th row X(4,:) of X to get the infill values
1. Table 1 shows the result of the tests via the different Sobol
sequence settings. Inputð:; 1Þ ¼ Temperature0

Create a matrix called ‘Input’ and make the first column as

2.1 Sobol sequence the 50 temperature values by typing the above (you
need to transpose the vector of temperature values by
The Sobol sequence is a method to sample data in a quasi-
using a prime, i.e. ′).
random sequence, in which data was selected in a uniformly
random form. The Sobol sequence was first introduced by Inputð:; 2Þ ¼ Speed0
Russian mathematician I.M Sobol [24]. According to
Savine, Sobol’s sequence could provide better evenness and Second column of input is the speed value.
higher speed to fill the space within a hyper cube [25]. Sobol’s
Inputð:; 3Þ ¼ BedT0
sequence had over past 20 years of improvement of the algo-
rithm to capable applied to high dimension. Hence, Sobol’s Third column of input is the bed temperature values.
sequence became a best practice in different applications.
Sobol’s sequence was generated with Sobol’s generator fitted Inputð:; 4Þ ¼ sInfill0
in MATLAB; experiment data of each parameter was gener-
Fourth column of input is the infill values.
ated based on the algorithm of Sobol’s sequence. This code
below creates 50 vectors (4 components in each vector), ac-
cording to a 4-variate uniform distribution implemented
2.2 Test rig
(approximately) using the Sobol sequence. Each component
The test rig was designed to test the gear wear whilst the gears
in each vector is a number between 0 and 1. The command
were meshed and running. Details of the test rig employed
above produces a matrix ‘X’ that lines up each of the vectors
have been discussed in previous paper [1, 2]. 3D printed gears
as a column in the matrix X. There are therefore 50 columns
can be tested in much the same way as metal gears, using a
and 4 rows.
back to back test configuration where the gears are loaded by
Temperature ¼ 225 þ X ð1; :Þ*50 winding in the torque to a prescribed level [2]. A torque of
10 Nm was added to the gears, with each test gear operated
This takes the first component of each of the 50 vectors and until failure. A motor was used to drive gears with an exter-
rescales it to get a temperature input value (between 225 and nally applied torque, with the reaction force between gear
275 K). Basically, use the first row X(1,:) of X. teeth were equivalent to the bearing block and loading arm
Speed ¼ 20 þ X ð2; :Þ*50 (Fig. 2). This loading method permits large amounts of wear
without significantly affecting the applied torque (Fig. 3).
Same as above but use the 2nd row X(2,:) of X to get the
printing speed values (between 20 and 70 rpm). 2.3 Artificial neural networks
BedT ¼ 10 þ X ð3; :Þ*50 Artificial neural networks simulate the physiological structure
and mechanism of the human brains in order to solve complex
3rd row X(3,:) of X to get the bed temperature values
problems. It is a machine learning process which is distinctly
Infill ¼ 20 þ X ð4; :Þ*50 different from common methods such as signal reasoning and
4484 Int J Adv Manuf Technol (2020) 107:4481–4492

Table 1 Input parameters generated by the Sobol sequence and output from test rig

Testing number Printing temp (°C) Printing speed (m/s) Bed temperature (°C) Infill percentage (%) Test result, gear fatigue time (hours)

1 230 25 30 20 0.04
2 253 50 50 50 20
3 264 38 60 35 11.11
4 241 63 40 65 30
5 247 44 55 28 1.94
6 269 69 35 58 24.69
7 258 31 45 43 9.32
8 236 56 65 73 21.03
9 238 41 43 61 15.57
10 261 66 63 31 10.1
11 272 28 53 76 30.18
12 250 53 33 46 20.6
13 244 34 68 54 10.12
14 267 59 48 24 6.66
15 255 47 38 69 12.9
16 233 72 58 39 0.36
17 234 48 64 44 12.77
18 257 73 44 74 36.8
19 268 36 34 29 1.65
20 245 61 54 59 16.66
21 251 30 49 37 2.88
22 274 55 69 67 20.16
23 262 42 59 22 2.67
24 240 67 39 52 10.32
25 237 33 51 71 12.24
26 260 58 31 41 1.96
27 271 45 41 56 7.28
28 248 70 61 26 0.06
29 243 39 36 78 21.24
30 265 64 56 48 27.78
31 254 27 66 63 25.71
32 231 52 46 33 0.39
33 232 38 54 55 25.2
34 255 63 34 25 11.38
35 266 26 44 70 8.4
36 243 51 64 40 1.76
37 249 32 39 62 5.16
38 271 57 59 32 4.17
39 260 45 69 77 34.49
40 238 70 49 47 15.67
41 241 29 62 28 0.07
42 263 54 42 58 14.79
43 274 41 32 43 3.06
44 252 66 52 73 30.45
45 246 48 47 21 0.04
46 269 73 67 51 12.77
47 257 35 57 36 16.38
48 235 60 37 66 32.77
49 234 37 41 79 25
Int J Adv Manuf Technol (2020) 107:4481–4492 4485

Table 1 (continued)

Testing number Printing temp (°C) Printing speed (m/s) Bed temperature (°C) Infill percentage (%) Test result, gear fatigue time (hours)

50 256 62 61 49 25.41

logical thinking approaches [26]. ANN is an appropriate over when the gap between the input value and output value
method for solving incomplete associative memory, defect is smaller than the expected value.
characteristic pattern recognition and automatic learning Figure 4 shows the structure of the ANN model. The ANN
[27]. There are three main reasons that ANN was selected model in this paper was carried out based on MATLAB neural
for this research; first of all, the calculation speed of the network toolbox. Moreover, there is a loop fitted in the model
ANN is significantly computationally cheaper than other aimed to select optimised hidden number of neural from 1 to
methods [28]. Secondly, ANN has strong fault-tolerant 20. Result shows 5 hidden sizes providing less error. The
ability to minimise the uncertainty during the experi- ANN model in this paper is composed of 4 input layer nodes,
ments. Thirdly, ANN is adept in addressing problems 5 hidden layer nodes and 1 output layer nodes. The initial
with multi-parameter regression, which is hard to solve parameters of ANN, such as the connection weights between
with purely numerical methods [29]. Back-propagation input layer, hidden layer and output layer, and threshold value
(BP) training algorithm is the most frequently used of hidden layer and output layer have large influence on the
ANN training method [30]. predictive performance. Due to the small number of training
data, best validation performance could occur at epoch 1 as
2.3.1 Back-propagation networks shown in Fig. 5.

The principles of the back-propagation networks The detailed 2.4 Genetic algorithm
stages of BP training method are the following: (1) The sam-
ple data for training are input to the network. (2) Data moves For the traditional ANN predictive models, without combin-
forward from input stage to each hidden layer until the output ing optimization algorithms, the initial parameters are deter-
stage, then the output data is generated. (3) The difference mined randomly, which is inefficient or prone to converging
between input data and output data is compared, and if the to local optima, slow convergence speed, overtraining, sub-
differences are larger than expected, they will be transferred jectivity in the determining of model parameters and often
back to the hidden layer. (4) The weight of each neuron is pose a convergence problem [31]. The optimised algorithm
adjusted based on the deviation via the steepest descent meth- GA is able to optimise the initial parameters of machine learn-
od that means calculating the minimum value (maximum val- ing models to increase the estimating accuracy and accelerate
ue) of the loss function along the gradient descent (ascent) the convergence speed of the ANN models [32, 33].
direction and the deviation transited to the input layer. (5) Genetic algorithm (GA) is a parallel random search optimisa-
The value proceeds forward again and after repeated iteration; tion algorithm to simulate the genetic mechanism of natural GA,
the error constantly diminishes. (6) The training process is and biological evolution GA can conduct efficient heuristic

Fig. 2 Schematic of 3D printing optimising process

4486 Int J Adv Manuf Technol (2020) 107:4481–4492

Fig. 3 Schematic of wear test rig 1. Driver gear 8. Pulley

2. Driven gear 9. Motor
3. Pivot block 10. Motor
assembly. controller
4. Driven sha 11. Weight
5. Universal 12. LVDT
couplings.
6. Driving sha 13. Centre
spacer
7. Conical clutch 14. Pivot

search and parallel computing [34]. It introduces the biological the new generation is related to the fitness value of the indi-
evolutionary principle of ‘survival of the fittest’ in the coded vidual. The better the individual fitness value, the higher the
tandem population formed by optimisation parameters and probability of being selected [38].
chooses individuals according to the fitness function of the indi-
viduals and the operations of selection, cross and mutation to 2.4.2 Cross operation
make the individuals with high fitness value be retained; the
individuals with low fitness be eliminated [35]. The new gener- The cross operation refers to the selection of two individ-
ation would inherit the information of the previous generation uals from the old generation to produce new individuals by
and be superior to the previous generation. This iteration is re- randomly exchanging and combination of the chromosomal
peated until the predetermined expired criterion is met [36]. locations of the two old individuals [39].
The basic operations of the GA are divided into:
2.4.3 Mutation operation
2.4.1 Select operation
The mutation operation refers to the selection of an individual
The selection operation refers to the selection of individuals from the old generation and choosing a point in the chromosome
from the old generation to the new generation [37]. The prob- of the individual to mutate to produce a new individual. The
ability that the individual is selected from the old generation to basic process of GA is shown in Fig. 6.

Fig. 4 Schematic structure of

ANN model
Int J Adv Manuf Technol (2020) 107:4481–4492 4487

Best Validaon Performance is 0.043 at Epoch 1 Error Histogram with 20 Bins

10
Training
7
Validaon
Test
Mean Squared Error (mse) 6 Zero error

5
10

Instances
4
Train
Validaon 3
Test
Best
10 Goal
2

-0.02
0.04
0.11
0.18
0.25
0.32
0.39
0.46

0.60
-0.72
-0.65
-0.58
-0.51
-0.44

-0.30
-0.23
-0.16
-0.09

0.53
-0.37
10 0 1 3 4 5 7
2 6
7 Epoch Errors=Targets-Outputs
(a) (b)

(c)
Fig. 5 Performance validation of ANN

The detailed method of applying GA in improving the

performance of ANN is following: the GA is used to opti-
mise the initial parameters of ANN. Each particle in GA
contains all information of the initial parameters of the
ANN model. According to the fitness function of the indi-
viduals and the operations of selection, cross and mutation
to make the individuals with high fitness value be retained,
the individuals with low fitness are eliminated. This itera-
tion is repeated until the predetermined expired criterion is
met. The initial parameters of the particle with the highest
fitness are assigned to the ANN model. The objective func-
tion (fitness function) is the R-squared. The crossover co-
efficient of the GA algorithm is 0.2, the mutation coeffi-
cient is 0.2, the size of population is 100 and the maximum
iteration number is 100. Fig. 6 Schematic of GA process
4488 Int J Adv Manuf Technol (2020) 107:4481–4492

(a) (b)

(c)
Fig. 7 Performance result fitted with test data

Fig. 8 Optimisation process of

GA on weight ratio of ANN
Int J Adv Manuf Technol (2020) 107:4481–4492 4489

Fig. 9 Performance of different

models compared with test result

2.5 Leave-one-out cross validation algorithm is shown in Eq. 1; the results of the sensitivity anal-
ysis by using Garson’s algorithm is shown in Fig. 9.
Leave-one-out cross validation is a method which evaluates  
L   N  
the performance of a machine learning algorithm, which in ∑ W ij W jk = ∑ W rj 
this case is an ANN. As a technique, it can increase the pre- j¼1 r¼1
Rij ¼   ð1Þ
N L   N  
diction accuracy by increasing the training data point to and
∑ ∑ W ij W jk = ∑ W rj 
decrease the test data point to 1. Hence, leave-one-out cross i¼1 j¼1 r¼1
validation could eliminate the randomness of dividing in-
stances into for training and testing. By changing the ratio of where Rijis the relative importance of input parameters,
training and testing of AAN, it could maximised the training Wij,Wjkare the connection weights of the input layer hidden
algorithm to provide a better understanding of model and layer and the hidden-output layer, i = 1,2….N, k = 1,2….M
clearer pattern of the Sobol sequence [40]. Due to the small (N and M are the numbers of the input parameters and output
amount of data, it is workable to maximise the number of the parameters).
training data.

2.6 Garson’s algorithm 3 Result and discussion

Based on the established machine learning models, the sensi- Figure 7 shows the performance of each model fitting with
tivity analysis of the input parameters is conducted by original test data. Figure 7a shows the linear fitting between
adopting Garson’s algorithm. In 1991, Garson proposed the ANN model and test data given a Pearson product-
Garson’s algorithm [41, 42], later modified by Goh, for deter- moment correlation coefficient of 0.85326 and R square of
mining the relative importance of the input parameters to the 0.728. It shows high correlation related to the original test data
output parameter [39, 41, 43], the equation of Garson’s [44, 45]. However, to optimise the performance, it is possible

Fig. 10 Sensitivity response

contributes to result
4490 Int J Adv Manuf Technol (2020) 107:4481–4492

Fig. 11 3D printing process

optimisation by ANN simulation

to increase the accuracy of the prediction model. Hence, the In the GA optimisation process, 200 iterations were selected
GA-based ANN (Fig. 7b) has been applied to the model which due to the decrease in computational time and convergence
yields closer agreements between the measured and predicted towards an optimised solution. Each iteration has a population
values of gear fatigue time. R2 increases from 0.728 to 0.8 of 50. The plot on the solid line represents the average error
when the GA is applied; moreover, Pearson’s r increased by corresponding to the test data, and the dotted line represents
nearly 5%. This result could be explained by the fact that the the best fit corresponding to the test data from the wear test rig.
proposed ANN-based predictive model accuracy in this case It is shown that average error was decreased from around 23 to
was increased with GA optimisation. Furthermore, the initial 10%; moreover, best fit was improved from 10 to less than
target was to achieve an R-squared value greater than 0.9; 5%, respectively. Hence, it can be shown that applying GA
hence, the GA-based ANN can provide a relatively satisfac- can increase the efficiency and accuracy of the ANN regres-
tory result. However, optimisation and prediction accuracy sion model (Fig. 8).
can be further increased by applying leave-one cross Figure 9 shows the comparison of the prediction of each
validation. model and test performed by wear test rig. Compared with
Figure 4c shows the model applied with both GA and evaluation methods such as Pearson’s r and R-squared, it is
leave-one cross validation, Pearson’s r and R2 dramatically shown that leave-one cross validation applied to the GA-based
increase from 0.83 to 0.97 and 0.728 to 0.956, respectively. ANN model provides better accuracy compared with a con-
Hence, a model with applied leave-one cross validation was ventional ANN model and the GA-based ANN model. Hence,
selected as the final model to carry out further analysis. as a result, leave-one cross validation applied to the GA-based
Figure 5 shows the result of optimisation performed by ANN model can provide an efficiently and relatively accurate
GA, which is used to optimise the ratio between ω and δ in model to carry out the prediction of performance of 3D printed
order to improve the accuracy of the ANN. The solid plot nylon spur gears with different manufacturing parameters.
represents the average error corresponding to the real test data. The model reveals (Fig. 10) that printing temperature con-
tributes to the performance of a printed gear by around 22% in
terms of weighting. Printing speed has around a 23% influ-
ence on the performance. Bed temperature contributes an
8.6% influence on the final result, showing a reduced impor-
tance compared with the rest of the parameters. Hence, by
using Garson’s algorithm, it is possible to identify the most
influential parameter regarding gear performance is infill per-
centage. Conceptually, this result makes sense as it is possible
that by increasing infill percentage, the rigidity of gear under
loads is increased.
In order to explore the power of the model in predicting
optimal gear performance and outputting the 3D printer pa-
rameters required, a simulation was carried out. Figure 8
Fig. 12 5 tests using optimisation setting for 3D printed gears shows the simulation of 14,256 combinations of different
Int J Adv Manuf Technol (2020) 107:4481–4492 4491

parameters. In this simulation, printing temperature is in- 4 Conclusion

creased from 230 to 275 °C by 5 °C each step. Hence, there
are 10 data points created for printing temperature instead of In this paper, a set of experimental data has been designed by the
50. Printing speed was increased from 20 to 75 mm/s by Sobol sequence, providing relatively higher tolerance and cover-
5 mm/s each step. Hence, there are 12 data points generated. ing a much larger range of input data with minimal test data being
Bed temperature is increased from 30 to 70 °C with 5 °C each required. Four 3D printing parameters were selected via specific
time, with 9 data points required for the analysis. Infill per- requirements of polymer gears which require rigidity and light
centage was increased from 20 to 80%, with 12 data points. As weight. A prediction model of 3D printed gears has been carried
mentioned earlier, there are more than 5 million combinations out with three models including an ANN model, a GA-based
that could be used in generating test input data; howev- ANN model and a leave-one cross validation–applied GA-
er, errors in the 3D printing process and errors in the based ANN model. The results show that all models provide
test rig could counter the tolerance of the setting, hence relatively accurate predictions and provide satisfactory fitting to
a gap between parameters by factor of 5 could provide the test data. A leave-one cross validation–applied model pro-
relatively accurate results. Simulation was carried out by vides the strongest correlation with test results, with Pearson’s r
leave-one cross validation–applied GA-based ANN mod- equal to 0.97 and R2 equal to 0.956, respectively. Moreover, by
el. Simulation number 7532 showed 52.07 h of potential simulating an experiment, the printing parameters have been
gear performance with 3D printer settings of a printing optimised to increase the performance of the 3D printed polymer
temperature of 250 °C, printing speed of 70 mm/s, bed gears. The results suggest an optimised setting of the 3D printer
temperature of 25 °C and infill percentage of 80% as as follows: printing temperature is equal to 250 °C, a printing
shown in Fig. 11. speed of 70 mm/s, a bed temperature of 25 °C and the infill
Validation of this model result was performed by producing percentage is 80%. The operational time of the resultant 3D
a 3D printed gear using the same settings suggested by the printed polymer gear was increased more than 3 times compared
ANN optimisation. A total of 5 pairs of gears were printed and with the gears produced using the default print settings.
tested on the wear test rig, with the results shown in Fig. 12. Sensitivity analysis performed by Garson’s algorithm indicated
The results showed that the 5 tests yielded an average perfor- that infill percentage has the most influence on the performance
mance of 51.46 h, which was very close to the ANN simula- of a 3D printed gear, and bed temperature has the least influence
tion value of 52.07 h. Hence, optimisation simulation could be on the test result.
considered as a valid simulation.
Previous paper has carried out the analysis of 5 different
nylon 3D printing materials. There are 5 different materials 5 Limitations and future scope
that have been printed and tested including Nylon 618, Nylon
645, alloy 910 onyx and Markforged nylon filaments. Nylon Due to the unique characteristic of the ANN process, true
618 3D printed gear provided best wear performance among 5 correlation between each parameter was not fully studied.
different 3D printing filament materials. Research shows that Moreover, more data points added to the model could increase
the different mechanical performance between nylon fila- the accuracy of the simulation. There are several possible di-
ments was caused by differences in crystallinity and unique- rections based on this research. Firstly, to carry out the study
ness of the FDM process. Another outstanding behaviour of of the polymer molecular structure to explain the influence of
Nylon 618 was shown in SEM (scanning electron microsco- different parameter settings. Secondly, to investigate several
py); result shows dramatically different wear behaviour for the other 3D printed materials in order to understand the correla-
3D printed gears when compared with the literature reports of tion between different materials and creating model to predict
injection-moulded gears. Furthermore, Nylon 618 material the performance of gears produced by using different mate-
has outstanding thermal performance of gears during wear rials and elicit the required print parameters.
tests and together with SEM, which was used to analyse gear
failure mechanisms. The performance results showed that Open Access This article is licensed under a Creative Commons
gears 3D printed using Nylon 618 actually performed better Attribution 4.0 International License, which permits use, sharing, adap-
than injection-moulded Nylon 66 gears when low to medium tation, distribution and reproduction in any medium or format, as long as
you give appropriate credit to the original author(s) and the source, pro-
torque was applied. Associating with the result of optimal 3D
vide a link to the Creative Commons licence, and indicate if changes were
printer setting, it is believed that by improving the manufactur- made. The images or other third party material in this article are included
ing procedure, the performance of Nylon 618 was further in the article's Creative Commons licence, unless indicated otherwise in a
enhanced. By applying machine learning method to credit line to the material. If material is not included in the article's
Creative Commons licence and your intended use is not permitted by
manufacturing 3D printed gears could dramatically increase
statutory regulation or exceeds the permitted use, you will need to obtain
the mechanical behaviour of 3D printed part in a highly dy- permission directly from the copyright holder. To view a copy of this
namic criterion. licence, visit http://creativecommons.org/licenses/by/4.0/.
4492 Int J Adv Manuf Technol (2020) 107:4481–4492

References 22. Hadavandi E, Shahrabi J, Hayashi Y (2016) SPMoE: a novel

subspace-projected mixture of experts model for multi-target re-
gression problems. Soft Comput 20(5):2047–2065
1. Zhang Y, Purssell C, Mao K, Leigh S (2020) A physical investigation
23. Melki G, Cano A, Kecman V, Ventura S (2017) Multi-target support
of wear and thermal characteristics of 3D printed nylon spur gears.
vector regression via correlation regressor chains. Inf Sci 415:53–69
Tribol Int 141:105953
24. Sobol’ IM (1967) On the distribution of points in a cube and the
2. Mao K, Li W, Hooke C, Walton D (2009) Friction and wear behav-
approximate evaluation of integrals. Zh Vychisl Mat Mat Fiz 7(4):
iour of acetal and nylon gears. Wear 267(1–4):639–645
784–802
3. Hu Z, Mao K (2017) An investigation of misalignment effects on the
25. Savine A (2018) Modern computational finance: AAD and parallel
performance of acetal gears. Tribol Int 116:394–402
simulations. John Wiley & Sons
4. Hooke C, Mao K, Walton D, Breeds A, Kukureka S (1993) 26. Wang S-C (2003) Artificial neural network. Interdisciplinary com-
Measurement and prediction of the surface temperature in polymer puting in java programming. Springer, Boston, MA, p 81–100
gears and its relationship to gear wear. J Tribol 115(1):119–124
27. Ye Z, Kim MK (2018) Predicting electricity consumption in a
5. Gauvin R, Patrick G, Henry Y (1984) Maximum surface temperature building using an optimized back-propagation and Levenberg–
of the thermoplastic gear in a non-lubricated plastic/steel gear pair. Marquardt back-propagation neural network: case study of a shop-
Ecole Polytechnique de Montreal. J Manuf Process 20–27 ping mall in China. Sustain Cities Soc 42:176–183
6. Mao K (2007) A numerical method for polymer composite gear flash 28. Vogl TP, Mangis J, Rigler A, Zink W, Alkon D (1988) Accelerating
temperature prediction. Wear 262(11–12):1321–1329 the convergence of the back-propagation method. Biol Cybern
7. Berman B (2012) 3-D printing: the new industrial revolution. Bus 59(4–5):257–263
Horiz 55(2):155–162 29. Kshirsagar AP, Rathod MN (2012) Article: artificial neural net-
8. Wohlers T (2015) Additive manufacturing and 3D printing state of work. IJCA Proceedings on National Conference on Recent
the industry. Wohlers Associates, Fort Collins ISBN 978–0– Trends in Computing 2:12–16
9913332-1-9; 2013 30. Quinlan JR (2014) C4. 5: programs for machine learning. Elsevier
9. Chacón J, Caminero MA, García-Plaza E, Núñez PJ (2017) Additive 31. Mohamad ET, Armaghani DJ, Momeni E, Yazdavar A, Ebrahimi M
manufacturing of PLA structures using fused deposition modelling: (2018) Rock strength estimation: a PSO-based BP approach. Neural
effect of process parameters on mechanical properties and their op- Comput & Applic 30(5):1635–1646
timal selection. Mater Des 124:143–157 32. Dormishi A, Ataei M, Khaloo Kakaie R, Mikaeil R, Shaffiee HS
10. Gomez-Gras G, Jerez-Mesa R, Travieso-Rodriguez JA, Lluma- (2019) Performance evaluation of gang saw using hybrid ANFIS-
Fuentes J (2018) Fatigue performance of fused filament fabrication DE and hybrid ANFIS-PSO algorithms. J Min Reclam Env 10(2):
PLA specimens. Mater Des 140:278–285 543–557
11. Kuznetsov VE, Solonin AN, Tavitov AG, Urzhumtcev OD, 33. Babanouri N, Nasab SK, Sarafrazi S (2013) A hybrid particle
Vakulik AH (2019) Increasing strength of FFF 3D printed parts swarm optimization and multi-layer perceptron algorithm for bivar-
by influencing on temperature-related parameters of the process. iate fractal analysis of rock fractures roughness. Int J Rock Mech
Rapid Prototyp J. https://doi.org/10.1108/RPJ-01-2019-0017 Min Sci 60:66–74
12. Prajapati H, Ravoori D, Woods RL, Jain A (2018) Measurement of aniso- 34. Holland JH (1992) Adaptation in natural and artificial systems: an
tropic thermal conductivity and inter-layer thermal contact resistance in introductory analysis with applications to biology, control, and ar-
polymer fused deposition modeling (FDM). Addit Manuf 21:84–90 tificial intelligence. MIT Press
13. Bellehumeur C, Li L, Sun Q, Gu P (2004) Modeling of bond for- 35. Chipperfield AJ, Fleming P, Pohlheim H (1994) Genetic algorithm
mation between polymer filaments in the fused deposition model- toolbox: for use with MATLAB; user’s guide (version 1.2):
ing process. J Manuf Process 6(2):170–178 University of Sheffield, Department of Automatic Control and
14. Sun Q, Rizvi G, Bellehumeur C, Gu P (2008) Effect of processing Systems Engineering
conditions on the bonding quality of FDM polymer filaments. 36. Simpson AR, Dandy GC, Murphy LJ (1994) Genetic algorithms
Rapid Prototyp J 14(2):72–80 compared to other techniques for pipe optimization. J Water Resour
15. Wang M, Zhu Z, Dong Y, Zhou L (2017) Study of mixed-mode I/II Plan Manag 120(4):423–443
fractures using single cleavage semicircle compression specimens 37. Jadav K, Panchal MP (2012) Optimizing weights of artificial neural
under impacting loads. Eng Fract Mech 177:33–44 networks using genetic algorithms. Int J Adv Res Comput Sci
16. Deng L, Feng B, Zhang Y (2018) An optimization method for Electron Eng 1(10):47–51
multi-objective and multi-factor designing of a ceramic slurry: com- 38. Vose MD (1999) The simple genetic algorithm: foundations and
bining orthogonal experimental design with artificial neural net- theory. MIT Press
works. Ceram Int 44(13):15918–15923 39. Houck CR, Joines J, Kay MG (1995) A genetic algorithm for func-
17. Koeppe A, Padilla CAH, Voshage M, Schleifenbaum JH, Markert B tion optimization: a Matlab implementation. Ncsu-ie tr 95(09):1–10
(2018) Efficient numerical modeling of 3D-printed lattice-cell struc- 40. Wong T-T (2015) Performance evaluation of classification algo-
tures using neural networks. MFGLET 15:147–150 rithms by k-fold and leave-one-out cross validation. Pattern
18. Delli U, Chang S (2018) Automated process monitoring in 3D printing Recogn 48(9):2839–2846
using supervised machine learning. Procedia Manuf 26:865–870 41. Garson GD (1991) Interpreting neural-network connection weights.
19. Li W, Lin W, Yu J (2016) Predicting contact characteristics for AI Expert 6(4):46–51
helical gear using support vector machine. Neurocomputing 174: 42. Momeni E, Nazir R, Armaghani DJ, Maizir H (2014) Prediction of
1156–1161 pile bearing capacity using a hybrid genetic algorithm-based ANN.
Measurement 57:122–131
20. Sun S, Wang S, Wang Y, Lim TC, Yang Y (2018) Prediction and
43. Goh AT (1995) Back-propagation neural networks for modeling
optimization of hobbing gear geometric deviations. Mech Mach
complex systems. Artif Intell Eng 9(3):143–151
Theory 120:288–301
44. Lee Rodgers J, Nicewander WA (1988) Thirteen ways to look at the
21. Samanta B (2004) Gear fault detection using artificial neural net-
correlation coefficient. Am Stat 42(1):59–66
works and support vector machines with genetic algorithms. Mech
45. Stigler SM (1989) Francis Galton’s account of the invention of
Syst Signal Process 18(3):625–644
correlation. Stat Sci 73–79