3DP-2021-0304-ver9-Collado_1P.

3d 04/24/22 12:46am Page 1

3DP-2021-0304-ver9-Collado_1P
Type: research-article
3D PRINTING AND ADDITIVE MANUFACTURING
Volume 00, Number 00, 2022 Open camera or QR reader and
ª Mary Ann Liebert, Inc. scan code to access this article
DOI: 10.1089/3dp.2021.0304 and other resources online.

ORIGINAL ARTICLE

Studying the Effect of Short Carbon Fiber

on Fused Filament Fabrication Parts
Roughness via Machine Learning
AU2 c AU1 c Alberto Garcı́a-Collado,1 Pablo Eduardo Romero-Carrillo,1 R. Dorado-Vicente,1 and Munish Kumar Gupta2

Abstract

Along with the characteristic staircase effect, short carbon fibers, added to reinforce Fused Filament Fabrication
parts, can significantly worsen the resulting surface finishing. Concerning this topic, the present work intends to
improve the existing knowledge by analyzing 2400 measurements of arithmetic mean roughness Ra corre-
sponding to different combinations of six process parameters: the content by weight of short carbon fibers in
polyethylene terephthalate glycol (PETG) filaments f, layer height h, surface build angle h, number of walls w,
printing speed s, and extruder diameter d. The collected measurements were represented by dispersion and main
effect plots. These representations indicate that the most critical parameters are h, f, and h. Besides, up to a
carbon fiber content of 12%, roughness is mainly affected by the staircase effect. Hence, it would be likely to
obtain reinforced parts with similar roughness to unreinforced ones. Different machine learning methods were
also tested to extract more information. The prediction model of Ra using the Random Forest algorithm showed
a correlation coefficient equal to 0.94 and a mean absolute error equal to 2.026 lm. In contrast, the J48
algorithm identified a combination of parameters (h = 0.1 mm, d = 0.6 mm, and s = 30 mm/s) that, independent of
the build angle, provides a Ra < 25 lm when using a 20% carbon fiber PETG filament. An example part was
printed and measured to check the models. As a result, the J48 algorithm correctly classified surfaces with low
roughness (Ra < 25 lm), and the Random Forest algorithm predicted the Ra value with an average relative error
of less than 8%.

Keywords: FFF, short carbon fiber, machine learning, mean roughness, random forest, decision tree

Introduction fabrication or FFF, also called fused deposition modeling.

This technology is based on the fusion and extrusion of a
Additive manufacturing (AM), developed to reduce the
filament made from a thermoplastic material. b AU0
tooling investment required to obtain parts with complex
geometries, is currently a driven technology of industry 4.01
Reinforcement filaments and surface finishing in FFF
and can provide from small to large production runs. AM
helps to produce parts, layer by layer, usually in a single One of the main handicaps of the FFF technique is the
machine (3D printer), from a previously created digital surface quality of the printed parts.3 Current studies highlight
model. Through this approach, AM can contribute to flexible two reasons for the lack of smooth finishing: (1) the effect of
and rapid manufacturing by reducing tooling, operations, and tessellation of the original CAD model4; and (2) the staircase
the number of parts and providing products with complex effect.5 There are several ways to improve the surface quality
shapes. Nevertheless, AM technologies still have some of FFF printed parts: postprocessing methods such as me-
challenges, mainly in the slicing, shape optimization, or pre- chanical polishing (for example, tumbling or sandblasting),
and postprocessing stages.2 Different AM technologies are machining, vapor smoothing, metalizing, and coating,6 and
available, although the most widespread is fused filament the optimization of printing process parameters.7 An

1
Department of Mechanical and Mining Engineering, University of Jaén, EPS de Jaén, Jaén, Spain.
2
Faculty of Mechanical Engineering, Opole University of Technology, Opole, Poland.

1
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 2

2 GARCÍA-COLLADO ET AL.

increasing number of studies focus on selecting the critical Table 1. Mechanical Properties of Printing
FFF process parameters and discussing their effects on the Filaments Used
outputs, such as mechanical properties of the printed parts,
quality, and surface roughness.8–13 Thus, process parameter Mechanical properties PETG PETG 12 PETG 20
optimization seems to be an adequate strategy for improving Carbon fiber by weight (%) 0 11.6 20.5
the surface quality of FFF parts. Tensile modulus (MPa) 3510 4540 5230
One of the FFF advantages is the extensive catalog of Tensile strength (MPa) 53 52.9 56
materials available in the market. The most used polymers Flexural modulus (MPa) 2404 2648 5740
are14: polylactic acid (PLA), acrylonitrile butadiene styrene Flexural strength (MPa) 98.4 80.4 80
(ABS), polyether ether ketone (PEEK), and polyethylene
PETG, polyethylene terephthalate glycol.
terephthalate glycol (PETG). Regarding the interaction of
process parameters and filament material, Boschetoet al.15 or
Vahabli and Rahmati16 studied the relationship between layer In the authors’ opinion, the current lack of knowledge re-
height, build angle, and surface roughness for common ma- garding the effect of fiber content in surface roughness de-
terials (without reinforcement) such as ABS, ABS Plus, serves further study. Moreover, it is interesting to test several
Polycarbonate, ULTEM 9085. successful ML solutions, different from the overused NN
The above paragraph shows studies about the influence of approach. To face the above issue, this work analyzes the
filament material on surface quality. Nevertheless, few works arithmetic mean roughness of 48 PETG samples, each one
analyze the relationship between surface roughness and the with 10 surface build angles, printed using different combi-
use of filaments reinforced with short carbon fiber. The ad- nations of fiber content, layer height, extruder diameter,
dition of short carbon fibers to a polymer matrix improves the number of walls, and printing speed.
mechanical properties of printed parts.17–20 It also increases Among the ML algorithms tested in this study, decision tree
the tensile modulus up to 2.2 times, using fibers with an and Random Forest help to determine the main printing pa-
average length between 60 and 250 lm in traction speci- rameters and capture the relationships among inputs and out-
mens.21 However, the surface quality of printed parts re- puts. They also provide the optimal parameter combination for
inforced with carbon fiber is not well understood. printing parts with carbon fiber reinforcement. While Experi-
mental Procedure Section describes the samples printed on
Surface roughness classification and prediction carbon fiber PETG and the tested factors, ML Modeling Sec-
through machine learning tion analyzes ML solutions to identify the most significant
parameters responsible for the surface finishing and explain
In recent years, machine learning (ML) has become a tool how to obtain roughness predictions. Results and Discussion
to reduce process variability and increase product quality. Section shows the experimental measurements using plots and
The ML capabilities to face AM challenges include design presents and discusses the ML results that point out that carbon
decision support, process and yield optimization, in situ fiber worsens the surface finishing. Moreover, an example of a
monitoring, and printing process control.22 In addition, it is part printed using a specific combination of parameters is used
worth mentioning the use of ML techniques to predict the to validate the proposed predictions. Finally, the main con-
properties of printed parts.23 ML techniques can be super- clusions are drawn in Conclusions section. b AU0
vised or unsupervised.24 In supervised learning, each instance
or vector of input data is labeled to an output value. In un-
Materials and Methods b AU3
supervised learning, each instance is not associated with
output. Supervised learning techniques can work with con- Experimental procedure
tinuous output parameters (surface roughness, tensile
strength, printing time, cost) or with categories (good part or Materials and printing process. To assess the influence of
bad part). Thus, supervised learning techniques are divided short carbon fibers in PETG, three filaments, with differences
into regression or classification techniques. One of the most of around 10% in fiber content, were purchased from dif-
widely used ML algorithms is neural networks (NNs). A NN ferent manufacturers: (1) PETG polymer filament, (2) PETG
has a high evaluation capability to represent complex and 12 polymer filament reinforced with around 12% short car-
highly nonlinear relationships between input and output bon fibers. (3) PETG 20 polymer filament reinforced with
features. There are different NN types, although one of the around 20% short carbon fibers. Filament manufacturers
most widely used is the multilayer perceptron (MLP).25
Several research works use the MLP algorithm to generate
models to predict tensile strength, compressive strength, and
AU0 c dynamic modulus of elasticity in printed parts using FFF.26
NNs are complex and do not have graphical output.
Therefore, some authors propose the use of other ML algo-
rithms that are easier to explain, such as decision trees27 or
the support vector machine (SVM).28 Decision trees are su-
pervised algorithms capable of classifying instances (rough,
not very rough) based on the values acquired by the most
influential variables.29 The SVM (a supervised algorithm)
allows grouping instances into two sets (good, bad) using an FIG. 1. Samples printed to different process parameter
intermediate hyperplane.30 combinations.
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 3

EFFECT OF CARBON FIBER ON FFF PARTS’ ROUGHNESS 3

Table 2. Design of Experiments: Factors and Levels come or response considered to train the ML methods was the
average Ra of the five replications for each angle.
Factor Symbol Unit Levels
Layer height h mm 0.1, 0.2 ML modeling
Surface build angle h  0–90 This section describes the proposed ML algorithms to
Material f Carbon PETG,
identify the most significant parameters and provide an ad-
fiber % PETG 12,
(by weight) PETG 20 equate process parameter configuration. Attached to the ML
No. of outer walls w — 1, 3 study, this work includes a visual analysis of the measured
Nozzle diameter d mm 0.6, 0.8 data and an approximation model to estimate Ra from the
Printing speed s mm/s 30, 60 most significant printing factors.

Data visualization and approximation model. Data visu-

alization can help identify the most significant factors and
provide the carbon content percentage by weight. The ma-
their robustness: the observed measurement dispersion. Main
terial has a diameter of 2.85 mm with variability of –0.03 mm
effect plots can provide, for each parameter, the system
T1 c in diameter. Table 1 summarizes the filaments’ mechanical
means or dispersion values at its different levels.35 Besides
properties. Regarding the default printer parameters, all
identifying the main factors, it is possible to determine the
samples were printed with a 30% infill density, a grid infill
most adequate and robust configuration to get low roughness
pattern, and extrusion and bed temperatures of 235C and
and better mechanical properties. The idea is to identify the
70C, respectively.
best fiber content in the main effect plots regarding the
Many printing parameters can influence surface rough-
roughness. Finally, for each layer height tested, it is possible
ness.31,32 However, to perform a practical study (cost and
to build an m · n degree polynomial Bézier patch to estimate
time limited), this work only considers the following pa-
Ra as a function of two scaled experimental parameters. The
rameters: f the fiber content by weight, h the layer height, h
Bézier representation is used for modeling curves and sur-
the surface build angle, w the number of outer walls, d the
faces in CAD programs36:
nozzle diameter, and s the printing speed.
m n
Ra ¼ +i ¼ 0 +j ¼ 0 bij Bm n
i ðuÞBj ðvÞ , 0  u, v  1: (1)
Design of experiments to build ML models. A 3D printer
Ultimaker S5 with Ultimaker CURA 4.3 software was used where bij are the surface control points, and Bm n
i ðuÞ, Bj ðvÞ are
F1 c to print samples. Figure 1 shows the printing samples to the Bernstein polynomials.
measure surface roughness. Based on the work of Kim and The above surface can be fitted to experimental data using
Oh,33 sample shapes have dimensions of 50 · 30 mm, a height least squares. The surface’s degree is determined among the
of 27 mm, and 10 different build angles from 0 to 90. The lowest values that provide a correlation r >0.9. Although this
extrusion length is developed in the X-Y plane, and the build model establishes a nonlinear relationship among the inde-
direction is developed in Z. pendent and the dependent variable, a linear system is solved
T2 c Table 2 shows the full factorial design with the factors and to obtain the control points. Because of that, this polynomial
levels tested. The 48 printed samples help to achieve all com- regression can be considered a particular case of multiple
binations. Each sample has 10 build angles (from 0C to 90C), linear regression.
and in each angle, the arithmetic mean roughness Ra was mea-
sured five times (2400 measurements). Ra is the arithmetic mean Machine learning. Two types of algorithms have been
of the absolute values of the profile height deviations from the used in this work: classification algorithms and regression
mean line, recorded within the evaluation length.34 A commer- algorithms. Classification algorithms generate models capa-
cial Mitutoyo Surftest SJ-210 (Mitutoyo, Kawasaki, Japan) ble of predicting the class or category to which an instance
profilometer with a sampling length of 2.5 mm and measuring belongs from its data or input variables. In contrast, regres-
the speed of 0.5 mm/s provided the Ra measurements. The out- sion algorithms generate models capable of predicting the

Table 3. Brief Description of Classification and Regression Machine Learning Algorithms

(Elaborated from WEKA Documentation)
Classification Regression Classification and regression
Naive Bayes. Probabilistic classifier Linear Regression MLP. Multilayer Perceptron NN
Logistic. Multinomial logistic regression SMOreg. Support vector IBK. Instance-based learner with parameter k
model with a ridge estimator machine for regression K*. Instance-based classifier.
SMO. Sequential Minimal Optimization M5P. Decision tree with Random forest. Ensemble learning method
LibVSM. SMO algorithm for kernelized linear regressions that uses multitude of decision trees.
support vector machines at the leaves.
J48 (C4.5). Decision tree algorithm
LMT. Logistic Model Tree
MLP, multilayer perceptron; NN, neural network; WEKA, Waikato Environment for Knowledge Analysis.
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 4

4 GARCÍA-COLLADO ET AL.

FIG. 2. Samples’ surfaces obtained using a confocal microscope for different f, h, and h.

numerical values of the output variable from the data corre- ables are ignored, the generated models are of higher quality
sponding to the input variables. For classification methods, 24. In this work, the Ranker algorithm has been used in the
the goal (based on ISO 1302) is to distinguish two groups classification and regression problem.
depending on the roughness of the samples: class 1, for values
below 25 lm; class 2, for values above 25 lm. In this type of
problem, the following metrics provide the performance of the Results and Discussion
model: percentage of correctly classified instances; percentage The visual observation of the samples’ surfaces helps un-
of incorrectly classified instances; kappa statistic. This last derstand the effect of process parameters on the roughness.
parameter measures the concordance between predicted and Note the surface shape of 12 different samples portrayed
observed categorizations of a dataset. Values around unity in Figure 2 and obtained using a confocal laser microscope b F2
indicate a stronger concordance. For the regression methods, LEICA TCS SP5II. As expected due to the staircase effect,
the performance of the model is measured using the following it is possible to consider parallel rows of peaks and valleys in
indicators: r the correlation coefficient; MAE the mean abso- all samples. That supports the idea that a conventional pro-
AU0 c lute error; RMSE the root mean squared error. filometer is enough to characterize the samples’ surface fin-
A set of well-known ML approaches were trained and ishing. In contrast, the surface presents a pattern highly
tested using the experimental data to determine the best dependent on f, h, and h. The staircase pattern depends on h
T3 c classification and regression algorithms. Table 3 shows the and h, while f influences the final height from peak to valley.
ML classification and regression algorithms assessed in this Figure 2 reveals a high dependency of the surface quality
work. The Waikato Environment for Knowledge Analysis with carbon fiber content, which adds a salt-and-pepper noise
(WEKA) tool24 was used to implement the ML algorithms on the staircase pattern. Other authors have noted that using
and extract knowledge from the experimental dataset. filaments reinforced with fibers worsens the surface rough-
WEKA is one of the most widely used ML softwares, and it is ness. For example, Garcı́a et al.42 observed higher Ra values
a frequently applied tool in the manufacturing sector.29,37–41 on specimens printed using PLA with Graphene nanoplate-
WEKA also incorporates attribute selection algorithms, lets than on specimens printed using PLA, and Wang et al.20
which allow us to know which variables have the greatest reported the same phenomenon for carbon fiber reinforced
AU4 c influence on the output variable. Usually, when these vari- PEEK filaments. b AU0

FIG. 3. Photographs of samples’ surfaces obtained using a confocal microscope. PETG (a); PETG 12 (b); PETG 20 (c).
PETG, polyethylene terephthalate glycol.
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 5

EFFECT OF CARBON FIBER ON FFF PARTS’ ROUGHNESS 5

FIG. 4. Ra measurements through different h, for different f and h.

The above phenomenon appears during the extrusion and Regarding the experimental measurements, Figure 4 b F4
deposition processes. The extrusion through the nozzle shows the arithmetic mean roughness Ra on each build
aligns the fiber in the flow direction,43 and the high shear angle of the printed samples taking into account the per-
rates at the end of the nozzle promote compression among centage of fiber and layer height. From these results, it
fibers. Hence, the subsequent expansion out of the nozzle follows that:
stimulates carbon fibers on the sample’s surface, increasing
the roughness.  The fiber content increases Ra values and their disper-
F3 c Figure 3 shows the increment in surface noise due to the sion. The values measured on parts manufactured using
addition of carbon fiber in the PETG. The absence of carbon PETG and PETG 12 showed a lower dispersion than
fiber (Fig. 3a) generates a surface without noise, where the those obtained for PETG 20. While for PETG and
staircase pattern is the main effect. In contrast, Figure 3b and PETG 12, the surface quality mainly depends on the
c shows that the carbon fibers on the surface generate local staircase pattern, for PETG 20, the irregularities due to
irregularities in the staircase pattern that worsen the surface the fibers are the dominant effect. This fact explains
finishing. observed variations and can be seen in Figure 3, where

FIG. 5. Main effect plots of studied factors on Ra.

3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 6

6 GARCÍA-COLLADO ET AL.

Table 4. Analysis of Variance Table for the Experimental Measurements

Parameter Degree of freedom Sum of squares Mean square F-ratio p
f 2 7312.79 3656.40 142.31 1.09 · 10-48
d 1 134.15 134.15 5.22 0.023
w 1 95.47 95.47 3.72 0.054
v 1 183.08 183.08 7.13 0.008
h 1 3725.76 3725.76 145.01 3.62 · 10-29
h 9 23862.50 2651.38 103.20 7.73 · 10-104
Error 453 11638.60 25.6922
Total 468 46952.30

the staircase pattern, although blurred, can be appreci- Main effect and regression model
ated on the PETG 12 sample but not on the PETG 20 Representing the experimental data by main effect plots
one. helps compare the process parameters’ effect on Ra. Figure 5, b F5
 Regarding the angle, the roughness values seem to made with the scientific software Mathematica, collects the
draw a convex curve with minimum values at 0 and main effects: Ra mean values with error bars (symmetric to two
90 build angles. The results for PETG and PETG 12 times the standard error of the mean) for the factors and levels
filaments (Figs. 3 and 4) agree with the claim of Bo- detailed in Table 2. Factors that show a larger difference be-
schetto et al.15 or Vahabli and Rahmati,16 who showed tween the maximum and minimum values of mean Ra, such as
how the staircase effect causes the roughness to reach a f, h, and h, have a higher effect on Ra than factors that show a
maximum when parts are printed at angles of around slight difference, such as the number of walls and the extruder
20. In contrast, it was not easy to identify that effect speed and diameter. For the case of f, the roughness shows a
for the PETG 20 filament. constant value and low dispersion up to PETG 12, and h
 PETG and PETG 12 lead to similar Ra distributions. Thus presents the highest dispersion values. Due to its high disper-
parts based on PETG 12 could have similar roughness but sion and dependency on the part geometry, h is not a suitable
better mechanical properties than PETG ones. parameter to control the roughness during the process. How-
 Although samples obtained from PETG 20 filaments ever, the other factors show lower dispersion than h and can be
have, in general, higher Ra, the presented measurements controlled to provide robust responses from the printing sys-
prove that some specific 3D printer settings led to tem. According to the main factor f and considering the cov-
PETG 20 samples with a Ra similar to that obtained erage factor of 2 used for the error bars, to obtain a Ra < 25 lm
with PETG filaments. with a reinforced filament, PETG 12 is the best option since it
 The maximum roughness obtained for specimens with will provide 18.5 lm <Ra < 21.0 lm with 95% confidence. b AU0
PETG 20 is similar in both cases represented in Fig- While main effect plots allow comparing the importance
ure 4 (h = 0.1 mm and h = 0.2 mm) and is close to among factors, the analysis of variance (ANOVA) helps
45 lm. Specimens with PETG and PETG 12 reach quantify the likely importance of each factor. It is possible to
maximum values of 25 lm (when using h = 0.1 mm) compute the ANOVA table (Table 4) assuming responses b T4
and 35 lm (when using h = 0.2 mm). with equal variances, independent errors, and a normal

FIG. 6. Degree 2 · 4 polynomial Ra prediction model for different layer heights: h = 0.1 mm (a) and h = 0.2 mm (b).
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 7

EFFECT OF CARBON FIBER ON FFF PARTS’ ROUGHNESS 7

Table 5. Classification Algorithms’ distribution. This statistical analysis shows that h, f, and h
Performance Metrics p-values provide the strongest evidence for rejecting the null
hypothesis (probabilities of F-ratio *1 lower than 0.001).
Correctly Incorrectly Therefore, they are considered the more significant parame-
classified classified Kappa
Algorithm instances instances statistic ters. The lower the p-value of a factor, the more likely it
influences Ra. According to this criterion, the factors studied
NaiveBayes 80.25 19.75 0.4919 can be ordered as follows: h, f, h, v, d, and w.
LibVSM 62.63 37.37 0.114 For each h, and the most important factors: f and h, it is
Logistic 75.79 24.21 0.4219 possible to fit the Ra measurements by a polynomial surface
MLP (NN) 91.51 8.49 0.8057 that provides roughness estimations. The approximation sur-
SMO 73.46 25.54 0.3982 faces were made in the Mathematica software. Figure 6 shows b F6
IBk 93.20 6.80 0.8424 the resulting degree 2 · 4 Bézier patches. For both approxi-
KStar 94.26 5.74 0.8668
mations: MAE £2.34 lm, RMSE £3.75 lm, and r > 0.90.
Random Forest 95.11 4.89 0.8873
J48 (C4.5) 94.69 5.31 0.8771
LMT 94.69 5.31 0.8784 ML approach
ML classification and regression methods can help better
understand the studied factors’ effect on Ra. The proposed
classification and regression ML models are based on five

FIG. 7. J48 (C4.5) decision tree for five parameters: h, f, d, s, and h.

3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 8

8 GARCÍA-COLLADO ET AL.

Table 6. Regression Algorithms’ 95.11% of the instances (in 10-fold cross-validation). These
Performance Metrics results are better than those obtained by the algorithm in
previous research works for materials without reinforce-
Algorithm r MAE/lm RMSE/lm ment.39 This work uses the J48’s results due to its capability
Linear regression 0.449 7.213 9.000 to portray the information with a graphic decision tree.
MLP 0.809 4.568 6.122 Figure 7 shows the decision tree generated by the J48 al- b F7
SMOreg 0.375 6.895 9.433 gorithm (C4.5), as can be seen:
IBk 0.886 2.557 4.798 
kStar 0.876 4.054 5.720 Layer height, fiber content, and angle are the most
M5P 0.924 2.722 3.990 important parameters as they are nodes of the main tree
Random forest 0.941 2.026 3.409 branches.
 Parts printed with PETG 12 and h = 0.1 mm present a
MAE, mean absolute error; RMSE, root mean squared error. surface roughness lower than 25 lm, independent of the
value of the build angle. This conclusion substantiates
parameters: m, h, h, d, and s. The Ranker attribute selection the observations provided by the main effect plots.
 Parts printed with PETG 20 present a surface roughness
algorithm indicates that the variable w has the least influence
on surface roughness. lower than 25 lm independent of the angle to be printed
if the following printing parameters are used:
Classification algorithms. First, it was necessary to con- h = 0.1 mm, d = 0.6 mm, and s = 30 mm/s.
 The last observation provides new information that was
vert the numerical output variable into a categorical one:
roughness values below 25 lm were classified in class 1; not attained with the data visualization and the main
roughness values above 25 lm were assigned to class 2. The effect plots. Hence, the authors consider it to be one of
classifier efficiency of different ML algorithms, measured the main contributions of the work.
using the Kappa statistic and the number of correctly and
T5 c incorrectly classified data, is shown in Table 5. The most Regression algorithms. It is also possible to define a
successful algorithms are the Random Forest (kappa = prediction model based on ML algorithms. Table 6 summa- b T6
AU5 c 0.8873), the J48 (kappa = 0.8784), and LMT (kappa = 0.8784) rizes the regression’s MAE, RMSE, and the correlation co-
algorithms. The Random Forest can correctly classify efficients for the several ML algorithms tested. The best

FIG. 8. Random Forest predictions versus actual Ra measures.

3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 9

EFFECT OF CARBON FIBER ON FFF PARTS’ ROUGHNESS 9

regression algorithm is Random Forest (r = 0.94). The Ran- and 20, irregularities are noted on the staircase pattern.
dom Forest algorithm is one of the preferred ensemble clas- During extrusion, carbon fibers are compressed, and the
sification algorithms in manufacturing studies. It has been subsequent expansion out of the nozzle could explain
previously used in the literature in the field of AM. For ex- the presence of fibers on the samples’ surfaces and,
ample, Wu et al.44 used it to predict surface roughness on therefore, the observed irregularities.
printed parts from data measured using different sensors (table  A preliminary analysis of the experimental measures
temperature, extruder temperature, table vibration, extruder indicates that PETG 12 samples have a surface
F8 c vibration, melt pool temperature). Figure 8shows a plot of roughness similar to PETG ones. The maximum Ra for
actual versus predicted values. As shown in the figure, the PETG 12 is 25 lm for an h = 0.1 mm and 35 lm for an
model generated by the algorithm achieves a good fit for the h = 0.2 mm. Ra(h) curve has a shape caused by the
values corresponding to filament with PETG and PETG 12. In staircase effect similar to a normal distribution skewed
the case of the PETG 20 filament, the modeling is acceptable in to the right, already identified previously by other au-
the range of 0–25 lm. Above that value, the model finds it thors. The parts printed on PETG 20 present higher
AU0 c more difficult to predict the roughness to be obtained. surface roughness values, reaching maximum values
close to 50 lm. In contrast, the values measured in this
A real example case have larger variability and do not seem to follow
the Ra(h) shape mentioned above, most likely due to the
Printing a part allows checking the goodness of the pro- short fibers increasing the noise on the staircase pattern.
posed classification and regression models. The piece printed  Simple observation of main effect plots and the AN-
F9 c was the fixture shown in Figure 9with a configuration that, OVA test support that f, h, and h are the most critical
according to the ML approach used, provides class 1 results, parameters. ML methods also substantiated this claim.
Ra < 25 lm: PETG20, h = 0.1 mm, d = 0.6 mm, and s = 30 mm/ Furthermore, a polynomial approximations Ra(f, h) for
s. Ultimaker Cura default definition is maintained for other different h provide acceptable roughness predictions:
printer parameters. MAE £2.34 lm, RMSE £3.75 lm, and r > 0.90.
Figure 9 portrays the prediction model estimations (Bézier Taking advantage of ML to deal with large datasets and
patches and ML regressions) and the actual measurements for many independent variables, we studied the capabilities of
different parts’ build angles. ML regression identified the some of these algorithms to classify and model our experi-
expected class 1 roughness (Ra < 25 lm for the defined pa- mental dataset:
rameter combination). In contrast, the polynomial regression
failed to predict the class 1 roughness. Finally, we consider that  In the case of the classification problem, the best al-
the residuals between experimental values and estimations are gorithms are Random forest and J48. Since J48 can
acceptable due to the mean relative error lower than 8%. provide a decision tree graphic and has similar effi-
ciency to Random Forest, its classification capabilities
allow identifying those parameter combinations that
Conclusions
lead to Ra < 25 lm. It is worth mentioning that the
The visual and ML analysis conducted in this work shows that decision tree provided a combination of parameters
fiber content has an important influence on the Ra measurements (h = 0.1 mm, d = 0.6 mm, and s = 30 mm/s) for 20% fiber
on PETG FFF printed samples. Samples were printed and content filaments, with a high probability to generate Ra
measured to demonstrate the previous claim: 48 parts, with 10 < 25 lm, independent of the part build angle. This fact
different build angles (480 measurements in total). The variables was not observed from the mean effect plots, where
studied were: carbon fiber percentage by weight, nozzle diam- PETG 20 is related with an average Ra > 25 lm.
eter, layer height, build angle, printing speed, and the number of  For the regression problem, the model that achieves the
perimeters. These are the most relevant results obtained: best predictions is the Random Forest with MAE =
 Pictures of a confocal laser microscope show that the 2.026 lm, RMSE = 3.41 lm, and r = 0.94.
 Finally, the J48 classification algorithm correctly b AU6
staircase pattern mainly depends on layer height and
build angle. For those samples printed with PETG 12 identified the class 1 roughness (Ra < 25 lm) in the

FIG. 9. Actual sample: measures and estimations. Printed conditions: PETG 20, d = 0.6 mm, h = 0.1 mm, s = 30 mm/s.
 3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 10

10 GARCÍA-COLLADO ET AL.

example part used with validation purposes. In contrast, 16. Vahabli E, Rahmati S. Improvement of FDM parts’ surface
the Random Forest regression provided a good pre- quality using optimized neural networks—Medical case
diction with a mean relative error lower than 8%. studies. Rapid Prototyp J 2017;23:825–842.
17. Bhandari S, Lopez-Anido RA, Gardner DJ. Enhancing the
Author Disclosure Statement interlayer tensile strength of 3D printed short carbon fiber
reinforced PETG and PLA composites via annealing. Addit
No competing financial interests exist. Manuf 2019;30:100922.
18. Shanmugam V, Rajendran D, Babu K, et al. The mechan-
Funding Information ical testing and performance analysis of polymer-fibre
Grant PID2019-104586RB-I00 funded by MCIN/AEI/ composites prepared through the additive manufacturing.
10.13039/501100011033 and the Andalusian Economy, Polym Test 2021;93:106925.
Knowledge, Enterprise, and University Council, under grant 19. Tekinalp HL, Kunc V, Velez-Garcia GM, et al. Higly
HICOOL, reference 1263034, cofinanced by the European oriented carbon fiber-polymer composites via additive
manufacturing. Compos Sci Technol 2014;105:6.
Regional Development Fund, supported this work.
20. Wang P, Zou B, Ding S, et al. Preparation of short CF/GF
reinforced PEEK composite filaments and their compre-
References
hensive properties evaluation for FDM-3D printing. Com-
AU7 c 1. Sartal A, Bellas R, Mejias A, et al. The sustainable pos Part B Eng 2020;198:108175.
manufacturing concept, evolution and opportunities within 21. Ferreira R, Amatte I, Dutra T, et al. Experimental charac-
Industry 4.0: A literature review. Adv Mech Eng 2020;12. terization and micrography of 3D printed PLA and PLA
2. Oropallo W, Piegl LA. Ten challenges in 3D printing. Eng reinforced with short carbon fibers. Compos Part B Eng
Comput 2016;32:135–148. 2017;124:88–100.
AU8 c 3. Pérez M, Garcı́a-Collado A, Carou D, et al. On Surface 22. Jiang J, Xiong Y, Zhang Z, et al. Machine learning inte-
Quality of Engineered Parts Manufactured by Additive grated design for additive manufacturing. J Intell Manuf
Manufacturing and Postfinishing by Machining. 1 ed. 2020;33:1073–1086.
Manufacturing HiA, editor, 2021. 23. Meng L, McWilliams B, Jarosinski W, et al. Machine Learning
AU9 c 4. Nasr ESA, Al-Ahmari A, Moiduddin K. CAD issues in in Additive Manufacturing: A Review. Jom 2020;72:2363–2377.
additive manufacturing. In: Comprehensive Materials Pro- 24. Witten IH, Frank E, Hall MA, et al. Data Mining: Practical
cessing. Oxford: Elsevier, 2014; pp. 375–399. Machine Learning Tools and Techniques. 4th ed. Morgan
5. Pandey P, Reddy N, Dhande S. Slicing procedures in layered Kaufmann, 2017; CP1–CP621 p.
manufacturing: A review. Rapid Prototyp J 2003;9:274–288. 25. Qi X, Chen G, Li Y, et al. Applying neural-network-based
6. Diegel O, Nordin A, Motte D. A Practical Guide to Design machine learning to additive manufacturing: Current ap-
for Additive Manufacturing. Springer, 2019. plications, challenges, and future perspectives. Engineering
7. Perez M, Medina-Sanchez G, Garcia-Collado A, et al. 2019;5:721–729.
Surface quality enhancement of fused deposition modeling 26. Wang C, Tan X, Tor S, et al. Machine learning in additive
(FDM) printed samples based on the selection of critical manufacturing: State-of-the-art and perspectives. Addit
printing parameters. Materials 2018;11:1382. Manuf 2020;36:101538.
8. Adel M, Abdelaal O, Gad A, et al. Polishing of fused depo- 27. Barrios J, Romero P. Decision tree methods for predicting
sition modeling products by hot air jet: Evaluation of surface surface roughness in fused deposition modeling parts.
roughness. J Mater Process Technol 2018;251:73–82. Materials 2019;12:2574.
9. Geng P, Zhao J, Wu W, et al. Effects of extrusion speed 28. Delli U, Chang S. Automated process monitoring in 3D
and printing speed on the 3D printing stability of extruded printing using supervised machine learning. Procedia
PEEK filament. J Manuf Process 2019;37:266–273. Manuf 2018;26:5.
10. Jin Y, Wan Y, Zhang B, et al. Modeling of the chemical 29. Molero E, Fernandez J, Rodriguez-Alabanda O, et al. Use
finishing process for polylactic acid parts in fused deposi- of data mining techniques for the prediction of surface
tion modeling and investigation of its tensile properties. J roughness of printed parts in polylactic acid (PLA) by fused
Mater Process Technol 2017;240:233–239. deposition modeling (FDM): A practical application in
AU10 c 11. Kam M, Ipekci A, Sengul O. Investigation of the effect of frame glasses manufacturing. Polymers 2020;12:840.
FDM process parameters on mechanical properties of 3D 30. Aoyagi K, Wang H, Sudo H, et al. Simple method to
printed PA12 samples using Taguchi method. J Thermo- construct process maps for additive manufacturing using a
plast Compos Mater 2021. support vector machine. Addit Manuf 2019;27:353–362.
12. Plaza E, Lopez P, Torija M, et al. Analysis of PLA geo- 31. Bikas H, Stavropoulos P, Chryssolouris G. Additive
metric properties processed by FFF additive manufacturing: manufacturing methods and modelling approaches: A crit-
effects of process parameters and plate-extruder precision ical review. Int J Adv Manuf Technol 2016;83:389–405.
motion. Polymers 2019;11:1581. 32. Vahabli E, Rahmati S. Application of an RBF neural network
13. Popescu D, Zapciu A, Amza C, et al. FDM process pa- for FDM parts’ surface roughness prediction for enhancing
rameters influence over the mechanical properties of poly- surface quality. Int J Precis Eng Manuf 2016;17:1589–1603.
mer specimens: A review. Polym Test 2018;69:157–166. 33. Kim G, Oh Y. A benchmark study on rapid prototyping
14. Ngo T, Kashani A, Imbalzano G, et al. Additive manufacturing processes and machines: Quantitative comparisons of me-
(3D printing): A review of materials, methods, applications and chanical properties, accuracy, roughness, speed, and ma-
challenges. Composites Part B Eng 2018;143:172–196. terial cost. Proc Inst Mech Eng Part B J Eng Manuf
15. Boschetto A, Giordano V, Veniali F. Surface roughness 2008;222:201–215.
prediction in fused deposition modelling by neural net- 34. Standard AN. Surface Texture (Surface Roughness, Wa- b AU11
works. Int J Adv Manuf Technol 2013;67:2727–2742. viness and Lay). 2009.
 3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 11

EFFECT OF CARBON FIBER ON FFF PARTS’ ROUGHNESS 11

AU12 c 35. Jiju A. Design of Experiments for Engineers and Scientists. 42. Garcı́a E, Nunez PJ, Chacon JM, et al. Comparative study of
2nd ed. 2014. geometric properties of unreinforced PLA and PLA-Graphene
36. Farin G. Curves and Surfaces for CAGD. 5th ed. 2002. composite materials applied to additive manufacturing using
37. Bustillo A, López de Lacalle LN, Fernández-Valdivielso A, FFF technology. Polym Test 2020;91:16.
et al. Data-Mining Modeling for the prediction of wear on 43. Lewicki J, Rodriguez J, Zhu C, et al. 3D-printing of meso-
forming-taps in the threating of steel components. J Com- structurally ordered carbon fiber/polymer composites with
put Des Eng 2016;3:11. unprecedented orthotropic physical properties. Sci Rep
38. Bustillo A, Urbikain G, Perez J, et al. Smart optimization of 2017;7:43401.
a friction-drilling process based on boosting ensembles. 44. Wu D, Wei Y, Terpenny J. Predictive modelling of surface
J Manuf Syst 2018;48:108–121. roughness in fused deposition modelling using data fusion.
39. Cerro A, Romero P, Yigit O, et al. Use of machine learning Int J Product Res 2019;57:3992–4006.
algorithms for surface roughness prediction of printed parts
in polyvinyl butyral via fused deposition modeling. Int Address correspondence to:
J Adv Manuf Technol 2021;115:2465–2475. R. Dorado-Vicente b AU13
40. Maudes J, Bustillo A, Guerra A, et al. Random Forest Department of Mechanical and Mining Engineering
ensemble prediction of stent dimensions in microfabrica- University of Jaén
tion processes. Int J Adv Manuf Technol 2017;91:879– EPS de Jaén
893. Campus Las Lagunillas
41. Rodriguez J, Quintana G, Bustillo A, et al. A decision- Jaén 23071
making tool based on decision trees for roughness predic- Spain
tion in face milling. Int J Comput Integr Manuf 2017;30:
943–957. E-mail: [email protected]
3DP-2021-0304-ver9-Collado_1P.3d 04/24/22 12:46am Page 12

AUTHOR QUERY FOR 3DP-2021-0304-VER9-COLLADO_1P

AU0: The Publisher requests for readability that no paragraph exceeds 15 typeset lines. This paragraph contains 16 lines or
more. Please divide where needed.
AU1: Please identify (highlight or circle) all authors’ surnames for accurate indexing citations.
AU2: Please confirm if the initial ‘‘R’’ of author Dorado-Vicente is correct for publication.
AU3: Include IRB approval or waiver statement in the Materials and Methods section. The Clinical Trial Registration
number, if applicable, should be included at the end of the abstract.
AU4: Please check if the usage of ‘‘24’’ in the sentence, ‘‘Usually, when these variables are ignored, the generated models
are of higher quality 24,’’ is correct.
AU5: Please expand ‘‘LMT.’’
AU6: Please check if edit made to the sentence, ‘‘Finally, the J48 classification algorithm correctly identified the class 1
roughness (Ra < 25 lm) in the example part used with validation purposes,’’ is ok.
AU7: In Ref. 1, please mention the page range.
AU8: In Refs. 3, 6, and 24, please mention the location of the publisher.
AU9: In Ref. 4, please mention the editor names.
AU10: In Ref. 11, please mention the volume number and page range. If these are unavailable, please supply the article’s
full DOI number.
AU11: In Ref. 34, please provide the other publication details or URL.
AU12: In Refs. 35 and 36, please mention the publisher and location.
AU13: Please confirm if the initial ‘‘R’’ of corresponding author ‘‘Dorado-Vicente’’ is correct for publication.