Professional Documents
Culture Documents
1 s2.0 S0950061822003099 Main
1 s2.0 S0950061822003099 Main
1 s2.0 S0950061822003099 Main
A R T I C L E I N F O A B S T R A C T
Keywords: Recently with the increase in traffic loading, the traditional materials used for road construction deteriorate at a
Palm oil clinker fine faster rate due to repetitive traffic loading which greatly necessitates bitumen modification to improve its
Asphalt mixtures quality. Amid an ever-increasing waste generation and disposal crisis, researchers came up with multiple ideas,
Stiffness modulus
however, the implementation was halted due to different practitioners’ policies. Palm oil clinker (POC) waste is a
Rutting
prevalent waste dumped around the oil palm mill that pollutes the environment. To harness sustainability, this
Prediction
Response surface methodology study utilizes varying dosages of POC fine (POCF) at 2%, 4%, 6%, and 8% to produce the POCF modified bitumen
Artificial neural network (POCF-MB). Also, the conventional and microstructure properties were evaluated. The objective of this study is
to utilize response surface methodology (RSM) and artificial neural networks (ANN) to optimize and predict the
stiffness modulus and rutting characteristic of asphalt mixtures prepared with POCF modified bitumen (POCF-
MB). The conventional test results revealed that the incorporation of POCF improves the plain bitumen prop
erties with enhanced stiffness and temperature susceptibility. Microstructural analysis highlighted that a new
functional group Si-OH was formed because of the crystalline structure of Si-O that indicates bitumen properties
enhancement with POCF inclusion. Two input and output variables were considered which are POCF dosage, test
temperature, and stiffness modulus and rutting depth respectively. Results showed that all mixtures containing
POCF-MB show better performance than the control mixture. Though, 6% POCF dosage shows improved per
formance compared to other mixtures increasing stiffness by 33.33% and 57.42% respectively at 25 ◦ C and 40 ◦ C,
while rutting at 45 ◦ C shows increased resistance by 25.91%. For both approaches, there was a high degree of
agreement between the model-predicted values and actual. For the model statistical performance index, the RSM
indicates that R2 for stiffness and rutting response were (99.700 and 99.668), RMSE (266.091 and 0.597), and
MRE (68.793 and 3.841) respectively. The ANN R2 for stiffness and rutting response were (99.972 and 99.880),
RMSE (61.605 and 0.280), and MRE (12.093 and 2.044) respectively. The ANN use 70% data for training, 15%
data for testing, and 15% data for validation processes. The ANN model outperforms the RSM model for the
prediction of POCF-MB asphalt mixtures’ stiffness modulus and rutting properties.
* Corresponding author at: Department of Civil and Environmental Engineering, Universiti Teknologi PETRONAS, Bandar Seri Iskandar, Perak 32610, Malaysia.
E-mail address: nura_19001733@utp.edu.my (N.S.A. Yaro).
https://doi.org/10.1016/j.conbuildmat.2022.126618
Received 7 September 2021; Received in revised form 18 January 2022; Accepted 22 January 2022
Available online 4 February 2022
0950-0618/© 2022 Elsevier Ltd. All rights reserved.
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
2
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 1 stiffness modulus and rutting depth of dense graded asphalt mixtures
Physicochemical properties of bitumen, aggregate, and POCF. (AC). Also, a comparison of two optimization approaches (RSM and
Properties Specification Units Results Standard ANN) was carried out to optimize and predict the best responses for
limits POCF-modified asphalt mixtures performance using the experimental
Bitumen runs. For the study, the input variables considered are POCF dosage and
Specific gravity ASTMD70-18 – 1.03 1.0–1.06 testing temperatures while the resilient modulus and rutting depth as
Penetration ASTMD5-13 dmm 65 60–70 response (output) variables. To achieve the aim of the study, laboratory
Softening point ASTMD36-12 ◦
C 50.8 47–52 experimental data were utilized to develop the CCD-based RSM and
Ductility at 25 ◦ C ASTM D113-07 cm 124
ANN-based models. Based on the optimized RSM models which pro
˃100
Mass loss ASTM D2872 % 0.02 –
Coarse aggregate vided good predictability and adequate accuracy, the ANN model was
Abrasion loss ASTM C 131 % 24.82 developed using the RSM dataset. Depending on statistical metrics such
Aggregate crushing BS 812–110 % 19.92 <30 as root mean square error (RMSE), coefficient of determination (R2),
value
mean absolute error (MAE), the performance and effectiveness of the
Bulk specific gravity ASTM C 127 g/ 2.71
cm3 developed models will be assessed and compared.
Absorption ASTM C 127 % 0.42 <2
Flakiness BS 812 Part 105.1 % 8.5 <20 2. Materials and experimental details
Elongation BS 812 Part 105.2 % 13.25 <20
Fine aggregate
Bulk specific gravity ASTM C 128 g/ 2.67
The materials utilized for the study were all obtained from indige
cm3 nous companies and the source was kept constant to avoid inconsistency
Absorption ASTM C 127 % 1.21 <2 during the study and their properties are listed in the subsection below.
Flat and elongated tests ASTM D 4791 % 17.4
Palm oil clinkers fine
2.1. Bitumen and aggregate
Colour Blackish grey
Specific gravity ASTM C188-17 g/m3 2.58 –
Specific surface area ASTM C1274 − m2/g 1.0843 – For this study, a penetration grade of 60/70 bitumen obtained from
12 the PETRONAS Malaysian refinery was utilized. The bitumen physical
Loss on ignition (LOI) ASTM C311-17 % 6.95 – properties are shown in Table 1. Aggregate utilized are crushed granite
Moisture content ASTM D2216-71 % 0.96
aggregates obtained from an indigenous quarry in Ipoh, Perak Malaysia.
The Malaysian JKR/SPJ/2008 Public Work Department (PWD) standard
imitate the biological neurological systems of the human brain [30]. criteria for road works were used for the selection and assessing the
ANN has several essential features with the use of networked computer characteristics of the aggregate [36]. Table 1 provides the physical
devices to aid in the resolution of problems that are too complex [27]. It properties that fulfill the criteria. The mineral filler utilized in the study
is based on these that researchers in material science and pavement was granitic dust passing through an aperture of 0.075 mm.
engineering have taken an interest in utilizing ANN because of its
nonlinear nonparametric modeling, which gives a high degree of pre 2.2. Pocf
cision to tackle complicated issues involving nonlinear interactions [7].
Recently, RSM and ANN modeling techniques have been increasingly The POC was obtained from an indigenous palm oil waste processing
popular among researchers as a means of evaluating and solving issues company in Malaysia. The untreated POC was sorted, and other dirt and
realistically. In the field of the pavement industry, RSM and ANN unwanted materials were removed then the POC was heated overnight
techniques have recently been utilized to forecast or simulate the per at 100 ◦ C, to make it dry. Subsequently, using the Los Angeles abrasion
formance of asphalt concrete materials [7,31–35]. And the outcomes of machine (LAAV), the dry POC was grounded at 150 revolutions per
the successful development of the model from previous researchers have minute (RPM) for 8 h to make it to powder form. The grounded POC was
proven that model prediction utilizing the modeling techniques is of then filtered using a 0.075 mm BS sieve and the filtrate was referred to as
importance. POCF. Table 1 also shows the POCF’s physicochemical properties.
Because the distinctive feature of utilizing these methods is that the
utilization and comparison of the two optimization strategies to forecast 2.3. Characterization
the optimal -operating variables for asphalt mixtures is with lower test
run numbers, which might save time and minimize experimental testing For this study, microstructural analyses were conducted. The X-ray
[7,31–35]. Because, when it comes to pavement structure design and fluorescence (XRF) analyzer (the Bruker XRF model S8 Tiger) was used
analysis, the material’s performances properties are of critical signifi to determine the POCF chemical composition under helium gas condi
cance. However, obtaining this parameter for asphalt mixtures in a tions. The POCF and POCF-MB crystallinity, and mineralogy phases
laboratory is extremely time-consuming and less efficient. Because the were determined by qualitative X-ray diffraction (XRD) using Bruker’s
traditional optimization techniques such as ’single-factor-at-a-time’ are AXS D8 advance. For the analysis of the XRD peaks analysis, the High
tedious, time-consuming, and may fail to provide optimal results. As a score plus analytical software was utilized. FTIR spectroscopy was used
result of the foregoing facts, it is indicated that a knowledge-based to examine the chemical properties of bitumen and POCF. By utilizing
predictive modeling system could provide a transparent and system the Perkin Elmer FT–IR spectrometer (Perkin Elmer spectrum 65, MA,
atic analysis for modeling the effect of modifiers on mixtures’ mechan USA) the chemical properties of POCF were evaluated and it was utilized
ical properties. Thus, the use of statistical methods such as RSM and to characterize the structural properties of base and POCF-modified
ANN to optimize and predict the mechanical performance of an asphalt bitumen.
concrete mixture to generate a model is encouraging and pivotal. Thus,
in this study, RSM and ANN modeling tools were used to generate and 2.4. Preparation and characterization of modified bitumen using
compare the performance of the developed predictive models for the conventional methods
stiffness modulus and rutting depth of POC fine (POCF) modified asphalt
mixtures. For the preparation of the POCF-modified bitumen (POCF-MB) a
This study aims to utilize waste POCF as a bitumen modifier and high shear multi-mix laboratory mixer was utilized. Initially, penetra
investigate the effects of the POCF dosage and temperature on the tion grade 60/70 plain (unmodified) bitumen was preheated in the oven
at 130 ◦ C to make it sufficiently viscous for modification and mixing.
3
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
100
increases the frequency and axial force. Marshall samples with a diam
eter of about 101.6 mm and a depth of 65.1 mm were produced and the
90
dimension at 6 different points was then entered into a computer system
80
Lower Limit linked to the UTM machine. Samples were preconditioned in the UTM
70
Upper Limit machine at various temperatures around 3 h before the test. On the
60 vertical diameter of the specimens, 12 mm wide loading strips were used
% Passing
Sample
50 to apply a transverse compressive force. After an applied force was
40 supplied to the test specimen in pulses, the resulting total recoverable
30 diametric strain was measured along 0 and 90◦ axes and displayed on
the monitor screen. For the type and temperature of each mixture,
20
triplicate samples were analyzed, and the averaged value was reported.
10
0 2.5.3. Wheel tracking
0.01 0.1 1 10 100
Sieve Size(mm) Rutting one of the most prevalent pavement problems is which re
duces road user comfort and increases maintenance expenses. Rutting
Fig. 2. Aggregate gradation for AC14 as stipulated by the PWD [36]. simulator testing utilizing the wheel tracking system is one of the most
utilized tests for the evaluation of compacted asphalt mixtures slabs or
Then, known quantities of base bitumen were measured and put into a mixes containing modified bitumen because it provides better practical
stainless vessel and the acquired weight of the various dosage of POCF emulation [39]. Using a wheel tracking test, asphalt mixture failure
by weight of the plain bitumen measured was then added to the bitumen susceptibility may be evaluated for inadequate bitumen coating aggre
in small quantities to avoid agglomeration during mixing. The POCF and gate and weak adhesion between aggregates and bitumen [25]. This test
plain bitumen were blended using a mechanical stirrer with a stirring was done according to BS 598–110 [40]. Wessex wheel tracking
speed of 1000 rpm for 52 min at 140 ◦ C to obtain a homogeneous equipment was utilized to evaluate for resistance to rutting of the
bitumen blend as determined by [19]. asphalt mixtures. For the wheel track test, a slab-shaped specimen with a
Conventional tests were employed and performed as criteria for measurement of 300 mm × 300 mm × 50 mm was produced by mixing
quantifying the bitumen blend. In accordance with standards, the con roughly 10 kg of aggregates (coarse, fine, and mineral filler) at respec
ventional properties of the base and POCF-MB blends were compared. tive OBC of POCF-MB. The samples are mixed at 150 ± 10 ◦ C and a
penetration values and softening point were determined as specified by kneading compactor was used to compact the samples to obtain 7 ± 2%
The ASTM D5 and ASTM D36 respectively. The bitumen ductility was air void content. For this study, three different temperatures 45 ◦ C,
determined in accordance with ASTM D113. The Penetration Index (PI) 55 ◦ C, and 60 ◦ C were considered for the test. Before the testing, the
of plain/POCF-MB blends revealed bitumen’s temperature sensitivity slabs were preconditioned at various testing temperatures for six hours
based on the Shell bitumen handbook. To assess the high-temperature before the test began. The test was done under normal experiment cir
stability of the plain bitumen and POCF-MB under storage circum cumstances (BS 598–110) with wheel loads of 520 ± 5 N, which move
stances, a storage stability test was also performed. The test was per back and forth in repetitive motion at 42 passes or 21 cycles per minute
formed based on the tube test (PN-EN 13399). for 45 min or until a 15 mm deformation occurs in the samples,
whichever transpires first [25]. During the experiment, the rut depth
2.5. Experimental details and sample preparation was measured every 5 min. For the different types of the mixture, three
samples were produced and analyzed, and the mean value is recorded.
2.5.1. Volumetric and Marshall properties
The Marshall mix approach was applied and evaluated for com 2.6. Mathematical modeling approach
pacted samples in accordance with ASTM D6927 [37] and it was used to
optimize the asphalt mixtures. Asphalt mixture having nominal 2.6.1. Response surface methodology for the experimental based design
maximum aggregate size 14 mm (AC14) dense graded was utilized for The RSM is a technique of mathematical and statistical approaches
the study. The AC14 aggregate particle size distribution is shown in used for creating, enhancing, and optimizing problems when an output
Fig. 2. Using a Marshall compactor, for each mixture type, 0% (control), factor is impacted by several input factors [29]. The central composite
2%, 4%, 6% and 8% POCF-MB, three samples were generated for each design, through the RSM which is a fractional factorial design approach,
mixture type of the five percentages of bitumen (4.0%, 4.5%, 5.0%, gives the synergetic functional connection between the input and output
5.5%, and 6.0% by weight of mixture) to determine the optimal bitumen factors [27]. Stat-Ease Inc.’s Design Expert (V11.1.0) software was uti
content (OBC). Using the Marshall compactor, 75 blows/face were used lized in this study for the design of the experiment, statistical analysis,
to compact the loose asphalt mixture. Prior to the performance testing, modeling, and optimization. This software provides strong tools for
the asphalt mixes were confirmed for volumetric and Marshall charac combining dependent and independent variables in a process, mixed
teristics at OBC, and all modified mixtures values are within the required combination. It also aids in the visualization of data that has a greater
range for AC14 stipulated by the Malaysian Public Work Department impact on other variables, as well as in the exact modeling of the results
(PWD) [36]. obtained. From works of literature, CCD has shown to be an effective
method to show the interconnection between the input and output
2.5.2. Stiffness modulus variables within the laboratory range [7,19,41,42]. Thus, for this study,
The stiffness modulus is a critical factor in examining the roles of the optimal CCD technique was utilized for the design of the experiment
asphalt mixtures layers. It is the elastic modulus utilized in the layered and to describe the variables for analysis, because of its accurate pre
elastic theory and it may be estimated based on the recoverable strain diction, time-saving, and minimizing the number of experimental runs
under dynamic loading. It’s a non-destructive test that is performed [43,44].
based on BS EN 12697–26 (E) [38]. For this study, Universal Testing For both responses (stiffness modulus and rutting depth), the inde
Machine (UTM-30) with the number Model 79 − PV70B12/I2 was used pendent variables’ levels and ranges were chosen based on selected
to measure the indirect tension stiffness moduli of the asphaltic concrete previous works of literature and standard specification [7,19,36,38,40].
sample. The test was performed at varying temperatures of 10 ◦ C, 25 ◦ C, The independent variables considered are the POCF dosage which was
and 40 ◦ C using the UTM. To simulate different road conditions, the varied in five-factor levels (0%, 2%, 4%, 6%, and 8%) and the test
UTM-30 software imitates the changes in asphalt mixtures conditions temperature was varied in three-factor levels (10 ◦ C, 25 ◦ C, and 40 ◦ C)
4
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 2 the complicated non-linear connection between input and output vari
Design of experiment matrix for stiffness modulus and rutting responses. ables that can understand and forecast the trend of experimental data
Exp. Stiffness modulus Rutting depth when trained [46]. ANNs are capable of predicting, clustering, opti
run Autonomous Responses Autonomous Response mizing, and allocating the influence of specific optimization parameters
(Mix) variables variables on the behavior of the desired output, among other functions [45].
A: POCF B: Stiffness A: POCF B: Rut depth
When it comes to improving ANN model estimation precision, many
dosage Temp. modulus dosage Temp. (mm)
(%) (◦ C) (MPa) (%) (◦ C) techniques have been proposed. However, the backpropagation algo
rithm is one of the most effective [7,47]. In the ANNs, besides the input
1 4 25 1612 6 60 3.492
2 4 25 1594 8 55 3.631
layer, the network consists of one or more hidden layers and the output
3 2 10 2302 0 60 4.392 layer. A single artificial neuron gets information from other neurons,
4 4 25 1628 2 55 3.641 analyses it, and then sends it on to other neurons [7]. For the study,
5 8 10 2294 2 55 3.569 John’s Macintosh Project (JMP) Pro statistical software was employed
6 4 25 1598 0 55 4.094
via the feed-forward with one hidden layer, the backpropagation
7 0 25 1076 0 45 3.892
8 8 25 1298 2 45 3.412 learning method was utilized. Hidden layers and output layers were
9 4 10 2494 8 45 3.437 activated using a logarithmic sigmoid transfer function. To avoid over
10 6 10 2895 2 60 4.014 training and over-parameterization, the data was separated into
11 0 10 2127 4 45 3.216 training, cross-validation, and testing sets. For the ANN analysis, the
12 0 40 419 6 60 3.621
same 16 mix runs of the dataset generated from RSM were utilized for
13 8 25 1313 2 55 3.542
14 6 40 978 2 55 3.598 model training, validation, and testing also the ANN uses 70% data for
15 4 40 752 6 45 3.091 training, 15% data for validation, and 15% data for testing processes. To
16 6 40 984 8 55 3.649 ensure accuracy and minimize network overfitting, this ratio was
determined by experimentation. While the network was being trained by
modifying the weights, datasets validation was utilized to evaluate the
for stiffness modulus response, and rutting response having test tem
network’s performance and get the best fitting in terms of coefficient of
perature of the input variables varied in three variable levels (45 ◦ C,
determination (R2) and root mean square error (RMSE) values, and the
55 ◦ C, and 60 ◦ C). After assigning low and high levels from the CCD-RSM
best signal was utilized for the prediction.
DoE, a total of sixteen (16) mixes were generated by the software as
shown in Table 2. The central point was replicated four times by the
RSM software to enhance the reliability of the experiment, analysis, and 2.7. Model performance and comparison criteria
error assessment.
The stiffness modulus and rut depth values obtained from the com The effectiveness and significance of the RSM and ANN models
puter software attached to the test machines were entered into the RSM developed were assessed based on three statistical performance metrics:
DoE for analysis. After running the CCD analysis using an appropriate coefficient of determination (R2), roots mean square error (RMSE) and
polynomial model with terms as indicated in Eq. (1), the optimum in mean relative error (MRE), which were obtained by using Eqs. (2) to (4),
dependent factors for the predicted were attained. respectively.
∑
n n ∑
∑ n (∑ )
(1)
n 2
Y = β0 + βk Xk + βkJ X k X J + ε i=1 (Ei − Pi )
k=1 k=1 j≥k
2
R = 1− ∑n 2
(2)
i=1 (Pi )
Here Y is the predicted response (stiffness and rut depth), β0 is the √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
constant response value of the design center, Xk and Xj are independent 1∑ n
RMSE = (Ei − Pi )2 (3)
variables (POCF dosage and testing temperature, βk is the linear coeffi n i=1
cient, βkj is the interactive coefficient, where K is the linear coefficient,
j is the quadratic coefficient, n is the number of factors and ε represent 1∑ n ⃒⃒
⃒
⃒
⃒100 Ei − Pi ⃒ (4)
error or noise detected in the model. MRE(%) =
n i=1 ⃒ Ei ⃒
To test the model’s performance in terms of the responses inputted,
an ANOVA was performed on the data. With a confidence level of 95%, Here, Ei signifies the experimental data, Pi is the predicted data value
other ANOVA criteria such as lack of fit, F-value, and p-values were by the RSM and ANN, and n is the entire count of data or sample size.
employed to assess the accuracy, suitability, and significance of
regression models. Using design experts, 2D and 3D model graphs were 3. Results and discussions
generated to show the synergetic influence of two input variables on the
selected model. To validate RSM findings, verification was carried out to 3.1. POCF characterizations.
validate and evaluate the statistically derived response outputs. An
additional experiment using estimated values from numerical RSM POCF chemical compositions are shown in Table 3. The table shows
optimization was performed and the verification was analyzed. The that POCF is mostly composed of silica, which accounts for more than
optimum conditions were established through numerical optimization, half of its composition. Other minor chemicals found in the POCF are
which can be used to validate the RSM design results. aluminum oxide, potassium oxide, calcium oxide, and iron oxide. The
POCF could be classified as a class C pozzolanic material based on the
2.6.2. Artificial neural network collected data since the total of SiO2, Al2O3, and Fe2O3 is greater than
ANN is an intelligent modeling tool that imitates the way the bio the minimum criterion of 50%. POCF’s pozzolanic behavior is further
logical nervous system (brain) processes data [45,46]. ANN can describe defined by the presence of high silica and the pozzolanic activity of the
Table 3
POCF XRF analysis data and composition.
Oxides SiO2 Fe2O3 CaO Al2O3 K2 O MgO P2O5 SO3
Abundance (%) 53.70 3.87 17.00 1.46 13.90 2.37 5.29 0.92
5
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 5. XRD diffractogram profiles for plain bitumen and POCF-MB at vary
ing dosages.
amorphous silica. The amorphous silica’s reactivity is owing to its Fig. 6. FTIR spectral analysis for POCF, Plain bitumen, and POCF-MB at
varying dosage.
thermodynamically unstable silica networks [19].
Fig. 3 shows the X-ray diffraction pattern of POCF. A quartz-
dominated crystallization was detected in the POCF’s XRD spectrum. 3.2. Modified bitumen microstructural characterizations
As demonstrated between 20◦ and 35◦ , the POCF phase is mostly crys
talline with traces of amorphosity halo of silica oxide and potassium 3.2.1. Xrd
alumina silicate (2-theta). A-quartz was found to be the main phase, XRD was used to determine plain bitumen and POCF-MB structural
with the highest peaks being (α-SiO2). In addition, the XRD analysis properties, as illustrated in Fig. 5. It was observed that the base bitumen
reveals that the reference number of the POCF is 01–075-8320, with was entirely amorphous (non-crystalline), with no identifiable peaks.
peaks corresponding to 20.82◦ , 26.62◦ , and 39.43◦ , and comparisons However, with the incorporation of POCF, the base bitumen was
were made with (Joint Committee on Powder Diffraction Standards transformed from amorphous to semi-crystalline which indicated the
(JCPDS) entry card number 00–46-1045 for silicon dioxide/quartz [34]. effect of POCF modification on the base bitumen structural character
Fig. 4 shows the results of the POCF FTIR analysis. The FTIR data also istic. The peak of semi-crystalline may be observed at an angle around
identifies the quartz and cristobalite minerals found in POCF. In FTIR 21.8◦ . There was significant molecular silica - bitumen interactions that
spectra, the peak of organic carbon was detected at about 3000 cm− 1. cause the peak diffraction to shift. In other words, the asphaltene mol
Quartz is one of the most important minerals in POCF, and it is always ecules are not separated from the silica and bitumen by crystallization.
present. Crystalline phases such as quartz materials are present in the These results demonstrate that there is intercalation and peeling due to
samples, as evidenced by the existence of two bands at 997 and 1033 the modified bitumen peaks that fall between the carbon and silica peak
cm− 1 that is caused by vibrations in the stretching and bending bands of areas. The findings are in line with the results reported in previous
(Si-O) in SiO4. Si-O-Si vibration bending was responsible for peaks be studies [48,49] where bitumen was modified with a silicious material.
tween 480 and 554 cm− 1.
6
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 7. Combined penetration, softening point, and penetration index of POCF-MB at varying dosages.
7
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 6
Result of volumetric and Marshall properties of POCF-MB asphalt mixtures at respective OBC.
Parameters PWDLimits Mixture type Remark
0% POCF 2% POCF 4% POCF 6% POCF 8% POCF
8
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
4.5 The 6% POCF-MB mixtures (1614 MPa) showed the highest stiffness
4.0 modulus and the control asphalt mixture have the least (1076 MPa) at
25 ◦ C. Also, at 40 ◦ C, the stiffness modulus of the control, 2%, 4%, 6%,
Rut depth (mm)
3.5
and 8% POCF-MB asphalt mixtures were 419 MPa, 686 MPa, 752 MPa,
3.0
984 MPa, and 690 MPa, respectively. The decline in stiffness modulus of
2.5 the 8% POCF-MB mixtures could have been hampered by the high
2.0 dosage of POCF in the base bitumen. Generally, a large dosage of the
1.5 Control
modifier reduces bitumen mastic viscosity, which would lead to a lower
2%POCF-MB stiffness modulus. A similar trend was observed by [50,51]. After
1.0 4%POCF-MB loading at medium and high temperatures, the 6% POCF-MB asphalt
0.5 6%POCF-MB
8%POCF-MB mixture showed greater recoverability than the other mixes. The
0.0 modified bitumen’s elasticity determines the asphalt mixture’s capacity
0 5 10 15 20 25 30 35 40 45 50 to return to its original state. It was observed that the asphalt mixes with
Cycle time (mins)
6% POCF-MB have enhanced stiffness and elasticity.
Fig. 9. POCF-MB asphalt mixtures rutting depth at 45 ◦ C.
3.5. POCF-MB asphalt mixtures rutting test
assisted in the adequate coating of aggregate. While for the MQ values,
as the POCF content increases up to 6% it was also observed to increase Fig. 9 shows the average rut depth versus rut cycle time for both
beyond which it experiences a decline at a higher POCF dosage of 8%. control and POCF-MB asphalt mixtures at various dosages. All the POCF-
MB asphalt mixtures show a lower rut depth than the control mixture,
3.4. POCF-MB asphalt mixture stiffness modulus indicating they are more resistant to rutting. There was a significant
improvement in rutting performance with 4% POCF-MB and 6% POCF-
The stiffness modulus of both the control and different dosages of MB asphalt mixtures. This is because the asphalt mixtures showed sub
POCF-MB asphalt mixtures at 25 ◦ C and 40 ◦ C are depicted in Fig. 8. For stantial improvement in viscoelasticity and stiffness, resulting in the
each mix type, the three samples’ values were recorded at both 0 and improved aggregate coating leading to a well-connected aggregate
90◦ , and the average values were reported. Both test temperatures bitumen matrix with greater resistance to rutting [25,51]. Also, the
showed similar patterns, as it was observed that the stiffness modulus of improved densification of modified mixtures could be attributed for the
the mixtures was considerably improved due to modification. All POCF- reduced rut depth.
MB asphalt mixtures had a higher stiffness modulus value than the Among the POCF-MB mixtures, 8% POCF-MB displays the most
control mixture. The incorporation of POCF as a bitumen modifier might rutting depth. When a high dosage of POCF is incorporated into bitumen
result in an increased stiffness which would probably be due to effective it produces a very viscous blend that is unable to adequately coat the
interaction between the modified bitumen and the other asphalt mixture aggregates, the clustered phase-separated bitumen significantly raises
constituents’ materials [11,25]. In addition, the stiffness improvement the stiffness of the mixture, causing embrittlement. It can also be noted
may be ascribed to the POCF-MB improved bonding and adhesiveness, that 2% of POCF-MB asphalt mixtures also have deep rut depth among
which makes the modified bitumen and aggregate particles have a the modified mixture, which may be ascribed to the POCF weaker
stronger interaction. Another reason can also be attributed to greater impact on the bitumen at the lower dosage, which results in inadequate
POCF-bitumen mastic bonding as observed when POFA was observed to stiffness of the bitumen. The 6% POCF dosage can be deemed as optimal
improve the asphalt mixtures stiffness [11] as it improves the inter for improving the performance of the asphalt mixture against rutting
locking of the aggregates making it have a higher stiffness modulus. compared to other modified mixtures. Because it provides enough
Table 7
ANOVA by RSM for the optimization of stiffness modulus and rut depth for POCF-MB asphalt mixtures.
Source Sum of squares dof Mean square F-value p-value Observation
df: degree of freedom, F-values: Fisher-statistical test values, p-values: probability values, LoF: lack of fit.
9
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 8 effect between the POCF and temperatures so they cannot be eliminated
Regression analysis values for the responses. in the model. The variables AB might not be statically significant, but
Statistical Response’s Comment they are experimental influencing the response output of the model.
analysis value Thus, it is encouraged to retain the AB variable that has a strong theo
Stiffness Rut depth retical justification even if is statistically insignificant. Also, the
modulus (MPa) (mm)
quadratic terms A2 and B2 can force to be removed from the equation
Standard 115.21 0.0716 The models are significant for because it can cause some issues with the analysis which can make the
deviation navigation around the design model non-orthogonal and that can cause some problems during
Mean 1579.00 3.64 space
correlation 0.9805 0.9677
modeling which can cause the model to become ’non-hierarchical.
coefficient Moreover, the F-test for lack of fit (LoF) was used to assess the model’s
(R2) adequacy. LoF is a measure of the variance of the data in the model. The
Adj.R2 0.9707 0.9515 probability value of the LoF was 0.00001, suggesting that the model LoF
Pred.R2 0.9515 0.9249 was significant. Although the LoF was significant, the predicted and
Adequate. 31.1084 28.8542 adjusted coefficient of determination were in strong agreement, and the
precision suggested model may be utilized to explore the design space to discover
Coefficient of 7.30 1.97
an optimal situation [7,33]. There is further evidence that stiffness
variance
modulus reaction was influenced more by a higher test temperature (A)
with an F-value of 459.69 than POCF dosage (B) with an F-value of
density to support loads at high temperatures and displays maximum 10.50. Eq.5 present a statistical model equation that has been developed
stability to resist rutting. This could also be attributed to the significant to predict the stiffness modulus for the POCF-MB asphalt Mixture
interfacial forces between POCF and the bitumen blend, as the modified
Stiffnessmodulus= +1587.67+145.74A− 833.33B+2.28AB− 380.13A2
bitumen bonded strongly and adhered strongly. Since the bitumen and (5)
aggregates were better bonded, this will enhance the performance of the +171.47B2
asphalt mixture. This finding is in line with a study utilizing coconut
Here A and B represent the POCF dosage and test temperature.
shell ash as a modifier [50].
3.6.2. POCF-MB asphalt mixtures rutting depth response
3.6. RSM statistical modeling
One of the most prevalent pavement problems is rutting, which re
duces road user comfort and increases maintenance expenses. As a result
3.6.1. POCF-MB asphalt mixtures stiffness modulus response
of traffic loads, flexible pavements respond in a variety of ways to in
The stiffness modulus test can be used to evaluate the performance of
ternal pressures that develop. Rutting is a term used to describe a
asphalt mixes. The greater the stiffness modulus, the better the perfor
pavement’s permanent deformation. Pavements are more prone to
mance of the asphalt mixture. However, the mixture’s elasticity de
rutting in hotter climes and hotter seasons than in cooler ones. In
creases with increasing stiffness modulus. In addition, the stiffness
addition, areas with a higher concentration of traffic are more prone to
modulus is considered as the major input in the mechanistic-empirical
rutting. Simulator testing utilizing a wheel tracking device is utilized for
pavement design (MEPD) technique. The stiffness modulus is used to
the tests [7]. The depth of rutting in pavements is mostly determined by
quantify horizontal and vertical strains in asphalt layers and subgrades
the stiffness of the asphalt mixture. For the rutting response surface
which may be used to measure the stiffness of asphalt mixture under
quadratic model, the ANOVA analysis is shown in Table 8.
various environmental circumstances [7]. Based on the CCD-RSM, this
The rutting response of the POCF-MB asphalt mixture had a signifi
study investigated the influences of variables (POCF dosage and testing
cant p-value < 0.0001 and a higher F-value of 59.91. The p-value < 0.05
temperature) and optimize and predicted the stiffness modulus of POCF-
for the model showed that there was only a 0.01 percent chance that an
MB asphalt mixtures. After the RSM software regression analysis, a fitted
F-value of this size would come from noise. Analyze the model’s
quadratic polynomial model to predict the stiffness of the POCF-MB
importance. A confidence interval of 95 percent was chosen to evaluate
asphalt mixture was generated. Based on the highest order polynomial
the model’s significance. All models were determined based on the
in which extra terms are significant and are not aliased by the software,
highest level at which the extra variables were required and were non-
the model was chosen for this study. For the response surface quadratic
aliased by the RSM program. probability > F-values were<0.05 for
stiffness modulus and rutting model, the ANOVA results are shown in
the model variables A, B, A2, and B2 suggesting that these terms were
Table 7.
critical to the modified asphalt mixture’s rutting value. Whereas the
The POCF-MB asphalt mixtures stiffness modulus response exhibited
other term AB was irrelevant since its p-value was higher than 0.05.
a significant p-value of<0.0001 and a higher F-value of 100.48. For the
Generally, we eliminate variables from a model that is not significant
model, the p-values were < 0.05, which meant that there was only a 0.01
only if they are not part of a term that is significant. However, since A2
percent chance that an F-value of this intensity would occur due to
and B2 are significant. So, we can’t remove A and B, because they are
random noise. To evaluate the model’s significance the 95% confidence
used in the quadratic terms A2 and B2. Because if we force to remove it.
interval was used. All models were determined based on the highest
This can cause some issues with the analysis. It can make the model non-
level at which the extra variables were required and were non aliased by
orthogonal and that can cause some problems during modeling which
the RSM program prob > F values of<0.05 were found for the model
can cause the model to become ’non-hierarchical. Also, LoF was per
variables A, B, A2, and B2, indicating that the terms were relevant to the
formed to identify with the model, it was shown that there was a cor
stiffness modulus of the modified asphalt mixture., whereas the other
relation between the residual errors and the pure errors in the duplicated
term AB was irrelevant since its p-value was higher than 0.05. This
design point, with p-values of 0.1499. Further, the analysis shows that
means all the variables were important for the stiffness modulus of the
both the test temperature (A) with an F-value of 93.38 and POCF dosage
POCF-MB asphalt mixture, whereas the other variable AB was negligible
(B) with an F-value of 98.72 had a significant influence on the rutting
because its p-value was>0.05. The non-significant p-value suggested
response for this study. Eq.6 shows the statistical model developed to
that the model parameter AB did not substantially impact POCF-MB
predict the rutting model equation for the POCF-MB asphalt mixture.
asphalt mixtures stiffness modulus. From the mathematical models
suggested by the RSM software. Even though AB is an insignificant term, Rutdepthresponse = +3.25− 0.2531A+0.2339B− 0.0253AB+0.5194A2
but they are kept in the model to maintain the hierarchy of the model (6)
+0.1362B2
terms obtained by ANOVA. Furthermore, the term AB is the interaction
10
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig.10. Diagnostic plots for (a and b) stiffness modulus and (c and d) rutting model adequacy.
Where A and B represent the POCF dosage and test temperature. suitability and sufficiency of variance analysis. The smaller the standard
deviation of the produced model is with its mean, the lower its variance
3.6.3. RSM ANOVA, model selection and verification to the test data which means experimentation results will produce less
The model RSM ANOVA is shown in Table 8, the adj R2 values for uncertainty for the model. Based on the foregoing analysis, it can be
stiffness modulus and rutting response are 0.9707 and 0.9515 respec concluded that the generated model is desired, suitable, and effective for
tively. Likewise, the pred.R2 values are 0.9515 and 0.9249 respectively. modeling, optimization, and prediction of POCF-MB asphalt mixtures
The analysis showed that the two are in good agreement indicating the stiffness modulus and rutting response.
models’ validity since the disparity is < 0.2 which shows excellent
model predictability. The high R2 values, approaching unity suggested 3.6.4. Predicted vs actual diagnostic plots for the responses
that the quadratic models had been suitably adjusted to the actual data. Fig. 10a shows the probability plot of the stiffness modulus. From the
Adequate precision (AP) was used to evaluate the models’ signal-to- figure, a close alignment between the residual points and a straight line
noise ratio of the model. AP values > 4 are preferable, as they ensure was observed which reveals the effectiveness of the regression models
that all models created can be used to explore the design region. The AP since the data fell within the equality line. Fig. 10b shows the predicted
values for stiffness modulus and rutting response are 31.1084 and and the actual laboratory findings, which were used to evaluate the
28.8542 respectively, which shows that the model is appropriate, suf model’s fitness and correctness. According to the symmetrical distri
ficient, and effective for traversing the design space. bution of data points around the 45◦ line, stiffness modulus was properly
Furthermore, the standard deviation (SD) for the model was predicted by the model since it matched the real laboratory data nicely.
considerably smaller than the obtained mean values, indicating the Fig. 10c shows the normal percentage for the rutting response model
11
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 11. 2-D and 3-D RSM surface showing the stimulatory influence of POCF dose and temperature on stiffness modulus.
Fig. 12. 2-D and 3-D RSM surface showing the stimulatory influence of POCF dose and temperature on rutting response.
satisfactorily using the output model developed. The plots show a pos
Table 9
itive correlation between predicted and observed results because all
Stiffness modulus response optimization constraints.
points are within the equality line [19].
Factors/Responses Lower and upper Goal Weight Importance
limits
3.6.5. Influence of synergistic variables on stiffness modulus response
POCF dosage (%) 0–8 In range 1 3 The 2D and 3D response plot contours were generated from the CCD-
Test temperature 10–40 In range 1 3
RSM to demonstrate the interaction between factors and responses and
(◦ C)
Stiffness modulus 419–2795 Maximize 1 3 to explain the influence of input variables on the output response. For
(MPa) each model, the factors (POCF dosage and test temperature) were drawn
on the X and Y-axes, while the response (stiffness modulus) was drawn
on the Z-axis. Fig. 11 (a and b) shows the 2D and 3D response plot
vs externally studentized residuals. It should also be observed that the contour of the model and it was observed to have a great synergetic
normal plot of residuals was suitable, as most of the response nodes were effect between the input variables. The model contour lines were found
along the straight line. Fig. 10d shows the predicted vs observed rutting to be semi-elliptical in form, showing a significant interaction between
response plot for the generated model, it was observed that the rutting the POCF dosage and test temperature. The contour plot’s colour in
response of the POCF-MB asphalt mixture was calculated accurately and dicates response values, with blue to green to reddish indicating a lower
12
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 13. (a) Numerical ramp for stiffness modulus response optimization (b) desirability 3D contour.
Fig. 14. (a) Numerical ramps for rutting response optimization (b) desirability 3D contour.
13
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
14
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 17. ANN stiffness modulus response parity graph of actual and predicted values for (a) training (b) validation (c) testing and (d) prediction optimization profiler.
2693 MPa. By calculating ARPE between the experimental value and limits, all parameters were set to exploit all possible solutions and
optimum predicted was determined to be 3.23%. The absolute percent optimize the outcomes. For the input factors and responses, the opti
discrepancy between the actual and predicted value of the stiffness mization criteria are presented in Table 10 and the numerical solution
modulus response was shown to be very minimal, with the allowable ramp for the response rutting is presented in Fig. 14a.
range of < 5 %. This shows that the response can be effective and As illustrated in Fig. 14b the optimal solution showed high desir
accurately predict using the developed model. ability of 0.977. Based on the RSM analysis ramps for the POCF-MB
For the rutting response, the optimum solution was determined using asphalt mixtures the optimum rutting can be obtained by utilizing
a multi-objective optimization technique based on RSM. POCF-MB 4.89% POCF dosage at a testing temperature of 46.2 ◦ C to obtain a
asphalt mixtures were optimized based on optimal rutting perfor minimum rutting depth of 3.12 mm. To confirm the reliability of the
mance, with the primary aim being to determine the optimal combina optimization findings acquired by the RSM model, an extra run of ex
tion of POCF dosage and test temperature to use. Based on the study periments tests was conducted by utilizing the optimized limits of input
15
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Fig. 18. ANN rutting response parity graph of actual and predicted values for (a) training, (b) validation, (c) testing, and (d) prediction optimization profiler.
factors. Three samples were used, and their average value was evaluated efficiency and low errors are determining the best input combination,
and reported. The rutting depth response was examined, and the find network architecture, and model ambiguity [45]. It was therefore
ings were compared to the predicted. By evaluating the ARPE between necessary to use the RSM data to model and optimize the ANN design
the experimental and predicted values as shown in Eq. (7) above. Using matrix to identify the optimal parameter combination. To improve the
the optimum globalized POCF dosage and test temperature, the exper ANN prediction model, the collected data from the RSM-CCD optimi
imental rutting was evaluated to be 3.24 mm. Based on computed ARPE zation were utilized for the training, testing, and validation. In this
it was calculated to be 3.70% still less than the allowable limit of < 5%, study, the performance properties of POCF-MB asphalt mixtures were
the experimental result shows excellent correlation with the RSM pre determined using the neural network. Test temperature and POCF
dicted value. dosage were chosen as input factors. The output variables stiffness
modulus and rutting were determined independently because of the
wide variety of test conditions. To explicate complex system modeling
3.7. Artificial neural network modeling
and identification, the feedforward back-propagation network type is
commonly employed [7,46]. As a result, the number of neurons in the
The most essential phases in building an ANN network with high
16
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 13
Comparison of RSM and ANN prediction performances for stiffness modulus response.
Exp. run(mix) Actual stiffness (MPa) RSM ANN
Predictedstiffness(MPa) Absolute percentage error (%) Predicted stiffness(MPa) Absolute percentage error (%)
Table 14
Comparison of RSM and ANN prediction performances for rutting response.
Exp. run Actualrut depth (MPa) RSM ANN
Predictedstiffness(MPa) Absolute percentage error (%) Predicted stiffness(MPa) Absolute percentage error (%)
Fig. 19. Comparison between RSM and ANN models predictions values with the actual mean values for (a) stiffness modulus (b) rutting response.
17
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
Table 15 forecasting the POCF-MB asphalt mixtures stiffness modulus and rutting
Assessment of model statistical performance index metric of RSM and ANN. depth. While Fig. 18d shows the prediction optimization profiler for the
Responses Statistical performance index metrics. rutting response.
R2 MRE RMSE
RSM ANN RSM ANN RSM ANN
3.8. Comparative evaluation of RSM and ANN models performances
Stiffness modulus 99.700 99.972 68.793 12.093 266.091 61.605
(MPa)
Rut depth (mm) 99.668 99.880 3.841 2.044 0.597 0.280 Recently, the modeling and process optimization approach based on
CCD-RSM and ANN is very effective and popular among various re
searchers [7,27,45,46]. In this study, CCD-RSM and ANN techniques
hidden layer of the ANN was varied to get the optimal network design. were used to forecast the stiffness modulus and rutting depth of POCF-
From the statistical analysis, the highest R2 values for the stiffness MB asphalt mixtures. CCD-RSM and ANN models were compared to
modulus and rutting response are 0.9999 and 0.9933 respectively. While determine the accuracy of the predictions of the statistical models, by
the least RMSE value for stiffness modulus and rutting response are assessing the relationship between the actual and predicted values and
1.38915 and 0.0232 respectively. For both responses, three structural evaluating their respective ARPE. Table 13 and Table 14 show the actual
layers with 2–9-1 (stiffness modulus) and 2–6-1(rutting) topology were and projected outcomes and their respective ARPE for both stiffness
observed to be the best as shown in Table 11 which also shows the modulus and rutting response.
respective statistical values of the various architectural neurons utilized. Also, data from the CCD-RSM and ANN models were compared to
For this study, as shown in Figs. 15 and 16 the first layer has two mean actual data in Fig. 19. to determine the correctness of the models.
input neurons (POCF dosage and test temperature), The second layer Plots were made between each value and the number of runs, the model
indicated the hidden neurons with nine (stiffness modulus) and six predicted values that were closely related to the observed values. In
(rutting) neurons, and the third layer is the predicted response for POCF- comparison to the RSM model, ANN model predictions have a better
MB asphalt mixtures performance properties with respectively two match to the actual data. Thus, the ANN model is better capable of
neurons (stiffness modulus and rutting). generalizing data than the RSM model.
In this study, the ANN optimizer tools were utilized for the optimi From the study, ANN models were observed to be slightly flawless in
zation of the various responses. The major aims to maintain good stiff predicting the responses. The ANN was able to accurately predict the
ness modulus while simultaneously reducing the rut depth within the experimental outcomes and validate them for each data combination
given range of POCF dosage and test temperature. Tables 12 show the utilized for the stiffness modulus and rutting response. Furthermore, the
desired goal parameters for the optimization setup. CCD-RSM and ANN models were compared based on the R2, MRE, and
Fig. 17 (a to c) shows the graph for all input and output variables for RMSE to see if the models were adequate. The high correlation between
the stiffness modulus that compares the test data to the ANN data for these numbers confirms that the developed models are suitable for
predictions of the model. There is a strong correlation between the ANN replicating actual results. Generally, a model with R2 values closer to
model’s predictions and the actual results. The stiffness modulus had R2 unity (1) suggests that it can predict more effectively, better, and more
values of 0.9999, 0.9761, and 0.9962, training, validation, and testing accurately the predicted responses. Table 15 demonstrate both CCD-
and Fig. 17d shows the ANN profiler for the model optimization. RSM and ANN models are statistically significant and have high per
Also, the rutting depth response scatter plots are shown in Fig. 18 (a formance predicted values compared to experimental findings.
to c) a strong relationship between the predicted and actual values with The ANN model R2 values are higher compared to RSM. The ANN
R2 were 0.9933, 0.9664, and 0.9970 for training, validation, and testing has R2 (0.9997 and 0.9988) and MRE (12.093 and 2.044) whereas RSM
were observed respectively. According to the scatter plot, ANN shows has R2 (0.9970 and 0.9967) and MRE (68.793 and 3.841) for stiffness
high-accuracy results when modeling the variable. Almost all values and rutting response respectively. Also, a comparison was made in terms
were dispersed along the 45◦ line, which is an indication of excellent of errors (RMSE) from each model. It was observed that ANN prediction
compatibility between the ANN projected values and the test results. has shown less divergence (lower RMSE) from the actual responses than
Also, the R2 between the test and ANN-predicted outputs showed that RSM. ANN’s superiority in prediction is due to its ability to imitate
the employed ANN model trained with the test data was effective in human intelligence by learning patterns in datasets [7,45,46] and it can
Table 16
ANN and RSM model validation data set for stiffness modulus.
Run POCF dosage (%) Temperature(◦ C) Stiffness modulus (MPa)
Experimental Predictive model values
ANN APE RSM APE
Table 17
ANN and RSM model validation data set for rutting depth.
Run POCF dosage (%) Temperature(◦ C) Rutting depth (mm)
Experimental Predictive model values
ANN APE RSM APE
18
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
generalize the nonlinear correlation between actual and predicted temperature variations have a greater influence on the POCF-MB
outcomes. asphalt mixtures stiffness. Both models accurately forecast stiffness
modulus and rutting of POCF-MB asphalt mixtures. However, the
3.9. Model prediction validation predictive models’ statistical comparison shows that the ANN models
are more effective with the highest R2 and minimal prediction errors
The predictive models’ performance was assessed using statistical (RMSE and MRE) compared to the RSM model.
measures such as R2, MRE, and RMSE. Table 15 shows that the ANN 5. The study shows that the problem of POC landfilling can be allevi
model suited the experimental result better than RSM, as evidenced by ated by using it as a bitumen modifier if the mixture design and POCF
the estimated values of R2, MRE, and RMSE. Thus, a validation test was dosage are properly controlled, it can improve asphalt mixture me
conducted to reverify and validate the predictive model if they are ac chanical properties. Also, the combined RSM-ANN model shows
curate. To assess the performance of both models, we evaluated both more efficiency, accuracy, and reduced trial batches time as well as
models with a completely new series of laboratory tests to determine destructive experiments that result in extra wastage of materials.
how effective they were. As a result, five (5) new combinations different
from the data sets that were utilized in model development was used
with three specimens for each of the new combination were prepared, CRediT authorship contribution statement
and the average value was observed and recorded, and the predicted
stiffness modulus and rutting depth of the POCF-MB asphalt mixture was Nura Shehu Aliyu Yaro: Conceptualization, Data curation, Formal
determined. Table 16 shows the experimental and predictive values of analysis, Investigation, Methodology, Software, Writing – original draft.
stiffness modulus, while Table 17 shows the experimental and predicted Muslich Hartadi Sutanto: Project administration, Resources, Supervi
values for the unknown (new) data set for rutting depth. For the stiffness sion. Noor Zainab Habib: Project administration, Resources, Supervi
modulus, the R2 for the new data set for the ANN and RSM models is sion. Madzlan Napiah: Project administration, Resources, Supervision.
0.998 and 0.993, respectively, while the average error percentage is Aliyu Usman: Software, Validation, Visualization. Ashiru Muham
1.01 and 2.31. While for the rutting model, the R2 for the new data set mad: Writing – review & editing.
for the ANN and RSM models is 0.997 and 0.995, respectively, while the
average error percentage is 1.74 and 3.65. The absolute percent
Declaration of Competing Interest
discrepancy between the experimental and predicted value of the stiff
ness modulus and rutting response was observed to be minimal, within
The authors declare that they have no known competing financial
the allowable range of < 5%. This shows and verifies that the developed
interests or personal relationships that could have appeared to influence
model is effective and accurate for the prediction of POCF-MB asphalt
the work reported in this paper.
mixture stiffness modulus and rutting properties.
In comparison to the RSM model, ANN has a greater potential for
Acknowledgments
applicability. The capacity to simulate a wide collection of non-linear
polynomials, as opposed to RSM’s ability to capture just quadratic ap
The authors gratefully acknowledge University Teknologi Petronas
proximations, is the rationale for its superior performance.
(UTP) which provided financial and laboratory facilities to perform this
study.
4. Conclusion
References
In this paper, the viability of utilizing POCF as a bitumen modifier in
densely graded asphalt mixture (AC14) was assessed and the effects of [1] J. Gao, H. Wang, Z. You, M.R. Mohd Hasan, Y. Lei, M. Irfan, Rheological behavior
the POCF dosage and test temperature on the POCF-MB asphalt mixtures and sensitivity of wood-derived bio-oil modified asphalt binders, Appl. Sci. 8 (6)
performance properties (stiffness modulus and rutting) were analyzed (2018) 919, https://doi.org/10.3390/app8060919.
[2] N. S. A. Yaro, M. Napiah, M. H. Sutanto, A. Usman, A. D. Rafindadi,S. M. Saeed, S.
using CCD-RSM and ANN models. The following are the major conclu Abdulrahman, Evaluation of the Impact of Short-Term Aging on Volumetric and
sions derived from the experimental and computational modeling work: Marshall Properties of Palm Oil Clinker Fine Modified Asphalt Concrete (POCF-
MAC) j. Phy.: Conf. Ser., 1793 (1) (2021), p. 012054, https://doi.org/10.1088/
1742-6596/1793/1/012054.
1. From the POCF-MB samples microstructural analysis, the FTIR
[3] A. Behnood Application of rejuvenators to improve the rheological and mechanical
spectra reveal asymmetric stretching vibrations of Si-O-Si caused by properties of asphalt binders and mixtures: a review 2019/09/10/ J. Cleaner
POCF inclusion. Also, the XRD results showed to improve the inter Prod., 231 (2019), pp. 171-182, https://doi.org/10.1016/j.jclepro.2019.05.209.
action between the POCF and the bitumen mix, with silica traces, and [4] A. Behnood, M. Modiri Gharehveran, Morphology, rheology and physical
properties of polymer-modified asphalt binders, Eur. Polym. J. 112 (2019)
a distinct peak appearing. The modification stiffens the bitumen 766–791, https://doi.org/10.1016/j.eurpolymj.2018.10.049.
blend by reducing penetration and ductility whereas softening point [5] M.T. Rahman, A. Mohajerani, F. Giustozzi, Recycling of waste materials for asphalt
was increased. concrete and bitumen: a review, Materials 13 (7) (2020) 1495, https://doi.org/
10.3390/ma13071495.
2. The utilization of POCF as a bitumen modifier into asphalt mixtures [6] Y. Yang, Y. Zhang, E. Omairey, J. Cai, F. Gu, A.V. Bridgwater, Intermediate
causes improvement in the VFB which describes good binding and pyrolysis of organic fraction of municipal solid waste and rheological study of the
better aggregate bonding that gives better stability. It was further pyrolysis oil for potential use as bio-bitumen, J. Cleaner Prod. 187 (2018) 390–399,
https://doi.org/10.1016/j.jclepro.2018.03.205.
observed that POCF-MB increase mixtures bulk specific gravity [7] A. Usman, M.H. Sutanto, M. Napiah, S.E. Zoorob, N.S.A. Yaro, M.I. Khan,
which causes a decrease in air voids in the POCF-MB mixtures. Comparison of Performance Properties and Prediction of Regular and Gamma-
3. The POCF-MB asphalt mixtures show improved stiffness and rutting Irradiated Granular Waste Polyethylene Terephthalate Modified Asphalt Mixtures,
Polymers 13 (16) (2021) 2610, https://doi.org/10.3390/polym13162610.
resistance compared to control mixtures. The considerable [8] T.L. Tinga, R.P. Jayaa, N.A. Hassana, H. Yaacoba, D.S. Jayantib, M.A.M. Ariffinc,
improvement in stiffness and rutting resistance of the modified mixes A review of chemical and physical properties of coconut shell in asphalt mixture,
could be attributed to greater POCF-bitumen mastic bonding which J. Tekno 78 (4) (2016) 15. https://doi.org/10.11113/jt.v78.8002.
[9] M. Arabani, S.a. Tahami, Assessment of mechanical properties of rice husk ash
improves mixtures performance. It was observed that 6% POCF leads
modified asphalt mixture Constr, Build. Mater. 15 (149) (2017) 350–358.
to optimum mechanical improvement of the POCF-MB asphalt [10] Z. Han, A. Sha, Z. Tong, Z. Liu, J. Gao, X. Zou, D. Yuan, Study on the optimum rice
mixtures. husk ash content added in asphalt binder and its modification with bio-oil, Constr.
4. From the optimization and prediction, the CCD-RSM and ANN have Build. Mater. 30 (147) (2017) 776–789, https://doi.org/10.1016/j.
conbuildmat.2017.05.004.
been shown to be useful tools. The stiffness modulus and rutting [11] K.R. Usman, M.R. Hainin, M.K.I. Mohd Satar, M.N. Mohd Warid, S.N.
properties were influenced by both input factors. However, the N. Kamarudin, S. Abdulrahman, Palm oil fuel ash application in cold mix dense-
19
N.S.A. Yaro et al. Construction and Building Materials 324 (2022) 126618
graded bituminous mixture, Constr. Build. Mater. 287 (2021) 123033, https://doi. [32] A. Milad, S.A. Majeed, N.I.M. Yusoff, Comparative study of utilising neural network
org/10.1016/j.conbuildmat.2021.123033. and response surface methodology for flexible pavement maintenance treatments,
[12] D. Movilla-Quesada, O. Muñoz, A. Raposeiras, D. Castro-Fresno, Thermal Civil Eng. J. (Iran) 6 (2020) 1895–1905. https://doi.org/10.28991/cej-2020-0309
suspectability analysis of the reuse of fly ash from cellulose industry as 1590.
contribution filler in bituminous mixtures, Constr. Build. Mater 160 (2018) [33] F.H. Khairuddin, M.Y. Alamawi, N.I.Md. Yusoff, K. Badri, H. Ceylan, S.N. Mohd
268–277, https://doi.org/10.1016/j.conbuildmat.2017.11.046. Tawil, Physicochemical and thermal analyses of polyurethane modified bitumen
[13] S.A. Tahami, M. Arabani, A.F. Mirhosseini, Usage of two biomass ashes as filler in incorporated with cecabase and rediset: optimisation using response surface
hot mix asphalt, Constr. Build. Mater. 170 (2018) 547–556. methodology, Fuel, 254 (2019), Article 115662, https://doi.org/10.1016/j.
[14] T. Ahmad, M. Rafatullah, A. Ghazali, O. Sulaiman, R. Hashim, Oil palm biomass- fuel.2019.115662.
based adsorbents for the removal of water pollutants: a review, J. Environ. Sci. [34] N. Baldo, E. Manthos, M. Miani, Stiffness modulus and marshall parameters of hot
Health Part C 29 (3) (2011) 177–222, https://doi.org/10.1080/ mix asphalts: laboratory data modeling by artificial neural networks characterized
10590501.2011.601847. by cross-validation, Appl. Sci. 9 (17) (2019) 3502, https://doi.org/10.3390/
[15] P.Y. Gan, Z.D. Li, Econometric study on Malaysia ׳s palm oil position in the world app9173502.
market to 2035, Renew. Sustain. Energy Reviews 39 (2014) 740–747. [35] G. Bosurgi, O. Pellegrino, G. Sollazzo, Optimizing artificial neural networks for the
[16] A.H. Jagaba, S.R.M. Kutty, G. Hayder, A. Noor, MFU bin Md Hafiz, N.S.A. Yaro, et evaluation of asphalt pavement structural performance, Baltic J. Road. Bridge.
al, Palm Oil Clinker as a Waste by-Product: Utilization and Circular Economy Eng., 14 (1) (2019), pp. 58-79, .https://doi.org/10.7250/bjrbe.2019-14.433.
Potential,Elaeis quineensis, IntechOpen (2021), pp. 1-24, 10.5772/ [36] Malaysia, Standard Specification for Road Works, Section 4, Flexible Pavement,
intechopen.97312. Jabatan, Kerja Raya(JKR) Malaysia, Kuala Lumpur, 2008.
[17] M.N. Borhan, A. Ismail, R.A. Rahmat, Evaluation of palm oil fuel ash (POFA) on [37] ASTM, D6927-15 Standard Test Method for Marshall Stability and Flow of Asphalt
asphalt mixtures, Austr. J. Basic. App. Sci. 4 (10) (2010) 5456–5463. Mixtures, ASTM Int, West Conshohocken, PA, USA (2015).
[18] H.M. Hamada, B.S. Thomas, F.M. Yahaya, K. Muthusamy, J. Yang, J.A. Abdalla, R. [38] B. Standard, Bituminous Mixtures-Test Methods for Hot Mix Asphalt, Part 24
A. Hawileh, Sustainable use of palm oil fuel ash as a supplementary cementitious (2012) 12697–12724.
material: Acomprehensive review, J. Build. Eng. 40 (2021) 102286, https://doi. [39] E. Ahmadinia, M. Zargar, M.R. Karim, M. Abdelaziz, E. Ahmadinia, Performance
org/10.1016/j.jobe.2021.102286. evaluation of utilization of waste Polyethylene Terephthalate (PET) in stone mastic
[19] N.S.A. Yaro, M.B. Napiah, M.H. Sutanto, A. Usman, S.M. Saeed, Modeling and asphalt, Constr. Build. Mater. 36 (2012) 984–989, https://doi.org/10.1016/j.
optimization of mixing parameters using response surface methodology and conbuildmat.2012.06.015.
characterization of palm oil clinker fine modified bitumen, Constr. Build. Mater. [40] BS 598-110, Methods of Test for the Determination of Wheel-Tracking Rate and
298 (2021) 123849, https://doi.org/10.1016/j.conbuildmat.2021.123849. Depth, in Sampling and Examination of Bituminous Mixtures for Roads and Other
[20] O. Siew Feen, R. Noor Mohamed, A. Mohamed, N.H.A. Khalid, effects of coarse Paved Areas; British Standard: London, UK, 1998.
palm oil clinker on properties of self-compacting lightweight concrete, J. Tekn. 79 [41] B.S. Mohammed, S. Haruna, M.M.B.A. Wahab, M.S. Liew, Optimization and
(6) (2017). https://doi.org/10.11113/jt.v79.10593. characterization of cast in-situ alkali-activated pastes by response surface
[21] N. Ranjbar, M. Mehrali, U.J. Alengaram, H.S.C. Metselaar, M. Jumaat, Compressive methodology, Constr. Build. Mater. 225 (2019) 776–787, https://doi.org/
strength and microstructural analysis of fly ash/palm oil fuel ash based geopolymer 10.1016/j.conbuildmat.2019.07.267.
mortar under elevated temperatures, Constr. Build. Mater. 65 (2014) 114–121, [42] N. Bala, I. Kamaruddin, M. Napiah, M.H. Sutanto, Polymer nanocomposite-
https://doi.org/10.1016/j.conbuildmat.2014.04.064. modified asphalt: Characterisation and optimisation using response surface
[22] N.J. Van Eck, L. Waltman, Manual for VOSviewer Version 1.6. 8, CWTS Meaningful methodology, Arab. J. Sci. Eng. 44 (5) (2019) 4233–4243, https://doi.org/
Metrics. Universiteit Leiden (2018) https://www.vosviewer.com/documentation/ 10.1007/s13369-018-3377-x.
Manual_VOSviewer_1.6.13.pdf (accessed September 1, 2021). [43] M.O. Hamzah, L. Gungat, B. Golchin, Estimation of optimum binder content of
[23] N.J. Van Eck, L. Waltman, Software survey: VOSviewer, a computer program for recycled asphalt incorporating a wax warm additive using response surface
bibliometric mapping, scientometrics 84 (2) (2010) 523–538. method, Int. J. Pavem. Eng. 18 (8) (2017) 682–692, https://doi.org/10.1080/
[24] K. Muthusamy., A. M. A. Budiea, S.M.S. Mohsin, N.S.M. Zam, and N. E. A. Nadzri 10298436.2015.1121779.
Properties of fly ash cement brick containing palm oil clinker as fine aggregate [44] W. Rafiq, M. Napiah, N.Z. Habib, M. H. Sutanto, W. S. Alaloul, M. I. Khan, M. A.
replacement. Mater. Today: Proc., 46(2021)1652-1656. Musarat, Abdul M.Memon,Modeling and design optimization of reclaimed asphalt
[25] A. Mohammed Babalghaith, S. Koting, N.H. Ramli Sulong, M.R. Karim, pavement containing crude palm oil using response surface methodology, Constr.
B. Mohammed AlMashjary, Performance evaluation of stone mastic asphalt (SMA) Build. Mater., 291 (2021), https://doi.org/10.1016/j.conbuildmat.2021.123288.
mixtures with palm oil clinker (POC) as fine aggregate replacement, Constr. Build. [45] H.A. Isiyaka, K. Jumbri, N.S. Sambudi, Z.U. Zango, N.A. Fathihah Abdullah,
Mater 262 (2020) 120546, https://doi.org/10.1016/j.conbuildmat.2020.120546. B. Saad, A. Mustapha, Adsorption of dicamba and MCPA onto MIL-53(Al)
[26] M.R. Karim, H. Hashim, H. Abdul Razak, S. Yusoff, Yusoff Characterization of palm metal–organic framework: response surface methodology and artificial neural
oil clinker powder for utilization in cement-based applications Constr, Build. Mater network model studies, RSC Adv. 10 (70) (2020) 43213–43224.
135 (2017) 21–29, https://doi.org/10.1016/j.conbuildmat.2016.12.158. [46] Z.U. Zango, A. Ramli, K. Jumbri, N.S. Sambudi, H.A. Isiyaka, N.H.H. Abu Bakar,
[27] S. Ray, M. Haque, T. Ahmed, T.T. Nahin, Comparison of artificial neural network B. Saad, Optimization studies and artificial neural network modeling for pyrene
(ANN) and response surface methodology (RSM) in predicting the compressive and adsorption onto UiO-66(Zr) and NH2-UiO-66(Zr) metal organic frameworks,
splitting tensile strength of concrete prepared with glass waste and tin (Sn) can Polyhedron 192 (2020) 114857, https://doi.org/10.1016/j.poly.2020.114857.
fiber, J. King. Saud. Univ - Eng Sci (2021), https://doi.org/10.1016/j. [47] S. Uslu, M.B. Celik, Performance and exhaust emission prediction of a SI engine
jksues.2021.03.006. fueled with I-amyl alcohol-gasoline blends: an ANN coupled RSM based
[28] G. Shafabakhsh, O.J. Ani, M. Talebsafa, Artificial neural network modeling (ANN) optimization, Fuel 265 (2020) 116922, https://doi.org/10.1016/j.
for predicting rutting performance of nano-modified hot-mix asphalt mixtures fuel.2019.116922.
containing steel slag aggregates, Constr. Build. Mater. 85 (2015) 136–143, https:// [48] S. Sembiring, A. Riyanto, R. Situmeang, Z. Sembiring, Bituminous composite
doi.org/10.1016/j.conbuildmat.2015.03.060. comprising amorphous silica from rice husks, Ceramics-Silikáty 63 (3) (2019).
[29] A. Usman, M. Hartadi Sutanto, M. Bin Napiah, N. Shehu Aliyu Yaro, Response [49] A.A. Hussein, R.P. Jaya, N.A. Hassan, H. Yaacob, G.F. Huseien, M.H.W. Ibrahim
Surface Methodology Optimization in Asphalt Mixtures: A Review, In Response Performance of nanoceramic powder on the chemical and physical properties of
Surface Methodology in Engineering Science; IntechOpen: London, in: bitumen Constr. Build. Mater., 156 (2017), pp. 496-505, https://doi.org/10.1016/
P. Kayaroganam (Ed.), Response Surface Methodology in Engineering Science, j.conbuildmat.2017.09.014.
IntechOpen, 2021, https://doi.org/10.5772/intechopen.95994. [50] S.N.A. Jeffry, R.P. Jaya, N. Abdul Hassan, H. Yaacob, M.K.I.M. Satar, Mechanical
[30] Q.-F. Liu, M.F. Iqbal, J. Yang, X.-Y. Lu, P. Zhang, M. Rauf, Prediction of chloride performance of asphalt mixture containing nano-charcoal coconut shell ash,
diffusivity in concrete using artificial neural network: modelling and performance Constr. Build. Mater. 173 (2018) 40–48, https://doi.org/10.1016/j.
evaluation, Constr. Build. Mater. 268 (2021) 121082, https://doi.org/10.1016/j. conbuildmat.2018.04.024.
conbuildmat.2020.121082. [51] N. Bala, M. Napiah, I. Kamaruddin, Effect of nanosilica particles on polypropylene
[31] Z. Zhao, J. Wang, X. Hou, Q. Xiang, F. Xiao, Viscosity Prediction of Rubberized polymer modified asphalt mixture performance, Case Stud, Constr. Mater. 8 (2018)
Asphalt-Rejuvenated Recycled Asphalt Pavement Binders Using Artificial Neural 447–454, https://doi.org/10.1016/j.cscm.2018.03.011.
Network Approach,“, J. Mater. Civil Eng. 33 (5) (2021) 04021071, https://doi.org/
10.1061/(ASCE)MT.1943-5533.0003679.
20