Catalysis Today 299 (2018) 289–302

Contents lists available at ScienceDirect

Catalysis Today
journal homepage: www.elsevier.com/locate/cattod

Statistical review of dry reforming of methane literature using decision tree MARK
and artificial neural network analysis
⁎
Ayşe Neslihan Şenera, M. Erdem Günayb, Aybüke Lebaa, Ramazan Yıldırıma,
a
Department of Chemical Engineering, Boğaziçi University, 34342, Bebek, Istanbul, Turkey
b
Department of Energy Systems Engineering, Istanbul Bilgi University, 34060, Eyup, Istanbul, Turkey

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

Keywords: The aim of this work was to extract knowledge for dry reforming of methane (DRM) reaction from experimental
Dry reforming of methane data using data mining tools such as decision trees and artificial neural networks. An extensive database con-
Data mining taining 5521 data points depending on 63 catalyst preparation and operational variables was constructed from
Knowledge extraction 101 papers published between 2005 and 2014; the output variables were CH4 conversion, CO2 conversion and
Decision trees
H2/CO ratio of the product stream. Then, the database, as a whole or as subsets for different base metals were
Artificial neural network
analyzed using decision trees (DT) to develop heuristics for high performance and artificial neural networks
(ANN) to determine relative importance of input variables and predict the performance under unstudied con-
ditions; mostly CH4 conversion, which is the most frequently reported output variable, were used in analysis. The
testing accuracy of the decision tree was about 80% leading to four heuristics (i.e. four possible courses of action)
for high CH4 conversion over Ni based catalyst. The first decision point to separate these heuristics is the reaction
temperature as can be expected. This is followed by the other variables such as support type, W/F and reduction
temperature. ANN analysis revealed that operational variables have higher relative importance (55%) compared
to catalyst preparation variables (45%). The most important operational variable was found to be the reaction
temperature while the active metal and the support are the most important catalyst preparation variables. ANN
model was also tested to predict the data, which was not seen by the model before, and the data in 65 papers out
of 101 were predicted within 15% error while 76 papers had the error rate of less than 20%.

1. Introduction been found to be more active and resistant to carbon deposition, their
high cost and low availability are preventing them to become a prac-
Dry reforming of methane (DRM) has been investigated significantly tical alternative to Ni [3]. Significant amount of work has been also
over the last few decades to convert two greenhouse gases (CO2 and reported aiming to improve the performance of Ni catalyst using some
CH4) into synthesis gas with a H2/CO ratio desirable for the synthesis of promoters. The situation is similar in the use of other catalyst materials
long chain hydrocarbons and oxygenated chemicals through Fischer- and preparation and reaction conditions; although some alternatives
Tropsch like process [1]. However, there are some serious problems like (like γ-Al2O3 for support, incipient wetness impregnation as the pre-
coke formation and metal sintering that decrease the efficiency and paration method and fixed bed as the reactor configuration) have re-
long term stability of the catalysts, and therefore prevent the com- ceived more attention, many other alternatives were also tested. As a
mercialization of this process. To overcome these problems, significant result, there are more than 1000 article in Web of Science database on
amount of work has been devoted in recent years by testing various dry reforming of methane. In these work, the effects of very large
combinations of active metal, promoter, support and catalyst prepara- number of catalyst preparation variables (like active metal type and
tion method as well as various configuration of reaction systems and loading, promoter type and loading, support, preparation method,
operational conditions [2]. calcination conditions and reduction conditions) and reaction variables
The most common catalyst material that has been studied for DRM (like temperature, feed compositions, flow rate and catalyst weight)
is supported nickel. However, Ni based catalysts have a tendency to were investigated creating large amount of data that cannot be ana-
form coke on catalyst surface; this can result in deactivation of the lyzed using simple literature search. Data mining tools can be used in
catalyst and/or plugging of tubes inside reactor. Although noble metals order to do this more effectively and extract valuable knowledge from
such as ruthenium and rhodium have been also studied and they have the database constructed from these publications, and this knowledge

⁎
Corresponding author.
E-mail address: [email protected] (R. Yıldırım).

http://dx.doi.org/10.1016/j.cattod.2017.05.012
Received 29 October 2016; Received in revised form 8 March 2017; Accepted 4 May 2017
Available online 29 July 2017
0920-5861/ © 2017 Elsevier B.V. All rights reserved.
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

can be used to improve the future experimental work. found to be influenced by 63 independent variables related to catalyst
Data mining is the process of extracting useful information from the preparation and reaction tests as given in Table 1 with their applic-
large databases using some tools like k-means clustering, decision trees, able ranges.
artificial neural networks, support vector machines and so on [4]. It
aims to identify patterns, trends and correlations that may be hidden 2.2. Computational details
within the network of data points, and that may not be observed with
naked eyes. There are several applications of these tools in the literature MATLAB R2013b was used for computational work. Decision tree
for modeling the experimental data in various fields of science or en- analysis was used to develop heuristics for higher catalytic performance
gineering including dry reforming reactions. For example, Omata et al. while artificial neural networks were used to determine the relative
[5] studied the effect of preparation parameters of Co-MgO catalysts for importance of input variables and to predict the outcome of unstudied
the dry reforming of methane using an experimental design procedure conditions. CH4 conversion for DRM reaction was used as the output
coupled with artificial neural networks; they also implemented a grid variable with 5521 data points. CO2 was not modeled because the
search algorithm to find the conditions for maximum catalytic activity. number of instances is small (3930), and CH4 conversion is already
Similarly, Ayodele and Cheng [6] used artificial neural networks and sufficient to determine performance; however, decision tree analysis
Box Behnken design together to model and optimize the input variables was also developed for H2/CO even though the number of data points
(reaction temperature, reactant feed ratio and CH4 partial pressure) for was also small (2323) because there are no other output variables that
desired CH4 and CO2 conversions and syngas ratio. In another study, may substitute this.
Hossain et al. [7] applied artificial neural networks for the same reac- In decision tree analysis, “classregtree” function of MATLAB in
tion over Ni/CaFe2O4 catalysts to model H2 and CO yields as well as classification mode was used. The CH4 conversion values are classified
CH4 and CO2 conversions. as low (0–50%), medium (51–75%) and high (76–100%); this division is
In the work cited above, the models were constructed using the not only meaningful physically but also suitable to avoid class im-
experimental data generated by the researchers themselves, hence their balance problem because each class had approximately equal number
applicability is limited. However, more general models and heuristics, of data points [4]. For H2/CO ratio, the classes were determined as
which can be valid in wider range, could be developed if these methods 0–0.89, 0.90–1.09 and > 1.1 (or 1.1–3.3). Gini's diversity index was
are applied to a larger database constructed from the entire literature, used to select the optimal split. The optimum decision tree was found
and they can help the researchers to benefit from the large experience by minimizing the error rate of individual leaf nodes (the fraction of
accumulated in the literature over the years. Indeed, we utilized such wrong classifications in that node) and the total error rate of the entire
approach in various reaction systems starting from selective CO oxi- tree (the weighted average error of all the leaves). Large trees may have
dation over Cu-Ce catalysts [8] and noble metal catalysts [9] as well as small training errors; however, they may overlearn, and may give hints
water gas shift reactions [10]. Then, we moved to more complex sys- that are too complex for practical purposes. Hence, the optimal tree size
tems like direct alcohol fuel cells [11] and biodiesel production [12]. was decided by minimizing the error obtained by using the tree to
Recently we successfully implemented the decision tree analysis for classify a dataset not seen before (testing). For this purpose, the total
steam reforming of methane [13], which is quite closer to methane dry data set was randomly divided into two sets, one containing 2/3 of the
reforming process. Similar analysis were also conducted by other re- data points while the other has 1/3. The larger set was used as training
searchers on the data acquired from other sources for steam reforming and constructing the decision tree and the smaller one was used for
[14], photocatalysis over titania [15] and oxidative coupling of me- testing and evaluating its generalization ability. Then, the tree size with
thane [16,17]. an acceptable error rate for training and the smallest error rate for
In this communication, we reported the construction of an extensive testing was chosen as optimum [9,18].
database from past publications in the literature for dry reforming of In the artificial neural network modeling, various network topolo-
methane and analysis of this database using decision tree and artificial gies were investigated first to determine the optimal topology using
neural network methods to develop models and heuristics to improve root mean square error (RMSE) of testing. Levenberg-Marquardt
the future experimental work. (trainlm) was used for training while Bayesian Regularization (trainbr)
algorithm was used for testing together with tangent sigmoid (tansig) as
2. Material and methods transfer function. Average root mean square error (RMSE) for testing
was used as the performance indicator for the neural network by im-
2.1. Database construction plementing k-fold cross validation method. The entire database was
randomly divided into k sets (k was chosen as 4 in this study), k-1 sets
The database was constructed by extracting the experimental data were used to train the network while remaining one set was used for
on dry reforming of methane from published articles in various da- testing the ability of network for unseen data. This procedure was re-
tabases (such as ACS, ScienceDirect and Wiley) between 2005 and peated 4 times and average RMSE was used to select the optimum to-
2014. The annual numbers of papers related to DRM is presented in pology. Each network topology was trained 10 times in order to elim-
Fig. 1 (more than 1000 in total); this figure was drawn using Web of inate the effects of random initialization of the neural network weights.
Science search for papers containing “dry reforming of methane (or Early stopping technique was used in order to prevent the model
CH4)”, “carbon dioxide (or CO2) reforming of methane (or CH4)” or overfitting [19].
“methane (or CH4) dry reforming” in their title directly; this number The change of root mean square error technique was used to find the
is probably much larger since some articles may not use these words relative importance of input variables. The procedure was as follows:
in their titles (value for 2016 covers only the publications up to mid- one of the input variables was removed from the data set, and then
October). 403 of these research publications were reviewed but 101 network was trained with the remaining variables. The RMSE value of
articles containing 5521 data points were used to construct the data the model calculated in the absence of this input variable was compared
set because remaining papers were not suitable to extract data or with the original RMSE calculated in the presence of all input variables.
contained variables that are not repeated in sufficient number of pa- The difference was used as the significance of this input group [19].
pers for a reliable model. The output (or performance) variables were One input neuron was assigned for each individual continuous
chosen as CH4 conversion (5521 data points; all publications reported variable (such as temperature, W/F ratio, feed compositions), one input
this variable), CO2 conversion (3930 data points; some publications neuron was assigned for each option for each categorical variable (for
did not report this variable) and H2/CO ratio (2323 data points; some example one for incipient to wetness impregnation, one for sol-gel and
publications did not report this variable as well). These outputs were so on.) with a value of 1 (if the method was used for that data) and 0 (if

290
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Fig. 1. Number of publications on dry reforming of methane through years.

Table 1
Input variables and their ranges.

Input variable Ranges for continuous variables or identities for discrete variables

catalyst preparation method Wet impregnation (WI), incipient to wetness impregnation (IWI), sequential impregnation (SI), co-impregnation (CI), sol-gel (SG), co-
precipitation (CP)
Metal Au, Ba, Ca, Ce, Co, Cu, Ir, K, La, Li, Mg, Mn, Ni, Pd, Pt, Rh, Ru, Sr, Y, Zr (0-30 wt percent depending on the metal)
support type α- Al2O3, γ-Al2O3, BaO, BaTiO3, CaO, CeO2, H-ZSM-5, ZSM-5, La2O3, MCM-41, Meso-Al2O3, ZrO2, Meso-ZrO2, MgAl2O4, MgO, Nano-MgO,
PrO2, SiO2, TiO2 V2O5, Y2O3, MnO, Si3N4, SBA-15
calcination time (h) maximum 12
calcination temperature (°C) maximum 900
reduction temperature (°C) maximum 850
reduction Time (h) maximum 10
H2% in reduction medium maximum 100
CH4% in reaction stream 10–83.3
CO2% in reaction stream 10–83.3
Ar, He or N2% in reaction stream 0–80
time on stream (TOS) (min) 0–7800
W/F (mgcat.min/ml) 0.15–88.24
reaction temperature (°C) 350–900
pressure (atm) 1 (constant)

it was not used). For active metal, promoter and support, one input for input importance and prediction in the following sections. Our in-
neuron was assigned for each option and the percent used in that data itial work suggested that we should perform decision tree analysis for
point was used as the value (zero if the metal was not used in that the subset containing one type active metal only; otherwise we would
experiment). increase the complexity of the models by not using clear information we
already know (i.e. all the other conditions will be depending on the
2.3. Verification experiments metal type). We will discuss our results for CH4 conversion over Ni
based catalysts in Section 3.2. The results of CH4 conversion on Co and
The verification experiments were chosen in a way that they may be Pt based catalyst as well as H2/CO ratio (for entire set) are presented in
used to test the heuristics developed by decision tree analysis but they an e-component file (Sener_et_al_additional_results.docx) to keep the
are also used to verify the ANN predictions (since any experiment manuscript in reasonable size. We did not perform DT analysis for other
within the range could be used to verify ANN model). Approximate active metals because the number of data points is small. We did not
range/level of input variables were decided with the help of Table 2 model CO2 conversion either because it does not make significant
(for better representation of literature), but the experiments reported in contribution to the information we obtain from CH4 conversion models.
literature were not repeated exactly (there would be no point of doing Similarly, we construct the neural network model for CH4 conversion
that); instead we only restricted ourselves to the conditions imposed by only. However, we used the entire data set this time by using active
the heuristics and determined the other conditions independently metal type as another variable; since the data involving different active
within the ranges reported in literature. The 8%wt.Ni-2%wt.Co/MgO metals may have similar information regarding the other variables, they
and 7%wt.Ni-2%wt.Co/γ-Al2O3 catalysts were prepared using incipient enrich neural network models because the number of data points in-
to wetness impregnation method. Activity tests were performed in a creases.
packed-bed, downward tubular, quartz reactor with a 775 mm length
and 10 mm inner diameter. Product stream was measured by Schi- 3.1. Review of publications from 2005 to 2014
madzu GC 2014 Instrument with a Carboxen1000 column.
Numerous studies have been reported in the literature in order to
3. Results and discussion develop stable catalyst and improve operating conditions for DRM, and
the number of publications has been increasing with time as shown in
We reviewed the literature to make simple analysis and presented in Fig. 1. As it was mentioned above this figure only contains the papers
the first section of Results and discussion. Then we discussed the de- that have dry reforming of methane in its name, hence there may be
cision tree analysis for heuristics and artificial neural network analysis more publications with slightly different naming. Type and loading of

291
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Table 2
Maximum CH4 conversion reported and conditions leading these values in reference publications.

292
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

the metals used, addition of promoter or other metals, preparation lower cost and wide availability [3]. Indeed 4223 of 5521 data points
technique, pretreatment (calcination and reduction) conditions and were obtained over Ni based catalysts. The distribution of data points
choice of support are the most important factors affecting the activity, among the metal types is presented in Fig. 2; the situation should be
selectivity and stability of catalyst. Operating conditions such as reac- similar in the entire literature since we did not intentionally favor or
tion temperature and pressure, reactant composition, gas hourly space disfavor nickel catalysts.
velocity and time on stream are also important because they can affect Although Ni has the low cost and availability advantages, it also has
catalytic performance. The conditions used to obtain the highest CH4 some problems such as catalyst deactivation due to carbon formation
conversion in the papers in the database were summarized in Table 2 to and metal sintering; both reduce activity and selectivity [126]. Noble
give the readers some general insight about the sources of data used in metals such as Ru, Rh and Pt supported catalysts are highly active and
analysis (some representative data from 2015 and 2016 were also stable for DRM with resistance to coke formation [3]; however, they are
added to table). The full database containing all information used is more expensive and rare, hence they are not industrially feasible. For
presented as an e-component file (Sener_et_al_dataset.xlsx). this reasons, many researchers are trying to improve the performance of
nickel catalyst by adding noble metals or other metal oxides as the
3.1.1. Effect of metal types and loading second metals or promoters [127]. For example, the influence of alka-
The widely used metals for DRM are Ni, Co, Pt, Pd, Ru, Rh and Ir. line earth metal oxides (MgO, CaO and BaO) on the activity and sta-
The data set contained twenty types of metals or metal oxides as the bility of 5 wt.%Ni catalyst supported on γ-Al2O3 in DRM were in-
base metals or promoters (like Ba, Ca, Cu, Sr and Y). Nickel was the vestigated in the study of Alipour et al. [1]. 5 wt.%Ni/γ-Al2O3 catalyst
most commonly used base metal in published papers because of its promoted with 3 wt.%MgO had the highest activity in the temperature

293
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Fig. 2. Common metals used (as active metal or promoter) for DRM.

Fig. 3. Common supports used for DRM.

range of 550–700 °C in terms of CH4 and CO2 conversion because MgO the catalytic performances of bimetallic Co-X/ZrO2 catalysts were af-
can improve the Ni dispersion by surface rearrangement, catalyst ac- fected by the type of metals because of their different dispersion
tivity due to formation of MgAl2O4 and catalyst stability by reducing properties. Unpromoted Co/ZrO2 had high initial activity but activity
the carbon formation on the catalyst. However, the incorporation of Ca loss was occurred because of carbon formation during six hour time on
and Ba into Ni/γ-Al2O3 catalyst resulted in a decrease in catalytic ac- stream. 5 wt.%Co–2 wt.%La/ZrO2 catalyst exhibited the highest stabi-
tivity and coke formation [1]. In another work, Özkara-Aydınoğlu and lity with average activity while 5 wt.%Co–2 wt.%Ce/ZrO2 catalyst ex-
Aksoylu [58] studied 2 wt.%La, Ce, Mn, Mg, K promoted 5 wt.%Co/ hibited the highest activity with a limited loss of activity in six hour
ZrO2 catalysts to increase the resistance of Co-based catalysts to coke time on stream experiments [58].
formation and to improve the activity of the catalysts, and found that The metal loading is also important variable to improve the activity
or prevent deactivation. There are numerous published papers that
were analyzed the effect of metal loading and tried to find the optimum
metal composition of the catalyst. For example, Xu et al. [88] tested
various catalysts with different Ni weight percentages (3, 5, 7, 10,
15 wt.%) over mesoporous alumina (MA). They found that, CH4 and
CO2 conversion increased significantly until Ni content reached to 5 wt.
% at various temperatures ranged from 600 °C to 800 °C, and then
further of increase of Ni to 15 wt.% resulted in slight improvement in
activity because 5 wt.%Ni/MA catalyst may already provide enough
active sites for reactants to achieve desired conversion levels [88].

3.1.2. Effect of support

In most of the work on dry reforming, the active metal is loaded
over a support with a relatively high surface area and large and reg-
ularly shaped pores [128]. Type of support and interaction between
Fig. 4. Common catalyst preparation methods used for DRM.
active metal and support affect the catalytic activity in various ways

294
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Fig. 5. Classification tree for CH4 conversion on Ni based catalysts (training accuracy: 81.5%, testing accuracy: 80.1%).

Although the most of the work in the database were performed over
Table 3 a single oxide supports, there are also some work that were performed
Confusion matrix of CH4 conversion (Ni based catalysts) for training and testing. over the mixed oxides. For example, García et al. [56] were studied the
effect of mixed oxide supports (ZrO2-MgO) with different amounts of
Experimental Data Predictions Classification
accuracy, % MgO content ranged from 0 to 5 wt.% on the catalytic performance of
Class # of Data 0−50 51–75 76–100 Ni/ZrO2-MgO catalysts at 600 °C for DRM. The best amount of MgO in
the catalyst in terms of CH4 conversion was found as 0.4 wt.%. The
Training 0−50 865 735 116 14 85.0
methane conversion remained almost constant at 26% during 300 min
51–75 934 124 726 84 77.7
76–100 845 32 118 695 82.2 on stream. The addition of MgO into Ni/ZrO2 inhibited the deactivation
Total 2644 81.5 of catalyst because of the strong interaction of metal with support of the
Testing 0−50 423 362 54 7 85.6 catalyst [56]. Similarly, the effect of Ce/Zr ratios on the catalytic ac-
51–75 482 74 356 52 73.9 tivities of the 7 wt.%Ni/CeZr catalysts at temperatures ranged from 600
76–100 417 15 61 341 81.8
to 800 °C was investigated in the study of Xu et al. [86]. The optimum
Total 1322 80.1
Ce content was found as 50 mol%, and 7 wt.%Ni-Ce50%Zr50% catalyst
exhibited the highest activity at each temperature [86].
like direct participation, metal dispersion and carbon deposition [3].
Data obtained over 24 supports were used in the database with the lead
of γ-Al2O3, which was the choice in about 39% of the data points 3.1.3. Effect of catalyst preparation method
(Fig. 3); the supports like ZrO2, CeO2, MgO, SiO2 and TiO2 were also Impregnation is the most common catalyst preparation techniques
used in decreasing frequency. (as wet or incipient to wetness impregnation); if there are more than
one active metal in the catalyst, both sequential and co-impregnation

Table 4
Heuristics for maximum CH4 conversion for Ni based catalysts [120–125].

295
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Table 5
Verifying heuristics over two catalyst formulationa.

Test # CH4 (%) CO2 (%) Reduc. T (°C) W/F (mgcat min/ml) CH4 Conv%

Reaction Temperature (°C): 650 700 750 800 850 900

1 43 43 800 1.43 38 55 70 84 93 97

H1 is satisfied (CH4 conv is 84 > 75% at 800 °C); H3 is not satisfied (CH4 conv is 70 < 75% at 750 °C)
2 43 43 800 1.43 37 47 61 73 84 93

H1 is not satisfied (CH4 conv is 73 < 75% at 800 °C); H3 is not satisfied (CH4 conv is 61 < 75% at 750 °C)
3 33 33 500 6.67 67 81 90

H1 is satisfied (CH4 conv is 90 > 75% at 800 °C); H2 is not satisfied (CH4 conv is 67 < 75% at 700 °C)
4 33 33 500 16.67 71 84 91

H1 is satisfied (CH4 conv is 91 > 75% at 800 °C); H2 is not satisfied (CH4 conv is 71 < 75% at 700 °C)
5 15 75 500 4.00 86 98 100

H4 is satisfied (CH4 conv is 86 > 75% at 650 °C)

6 33 33 800 4.00 47 63 78 88
H1 is satisfied (CH4 conv is 88 > 75% at 800 °C); H3 is satisfied (CH4 conv is 78 > 75% at 750 °C)

a
Tests 1 is over 8%wt.Ni-2%wt.Co/MgO, test 2–6 is over 7%wt.Ni-2%wt.Co/γ-Al2O3.

methods were utilized. There are also some works performed using the
techniques such as sol-gel and co-precipitation. Fig. 4 shows the dis- Table 6
tribution of the data points among six catalyst preparation techniques Relative importance of input variables.
used in the database.
Input Variable RMSE RMSE Relative Group
Although the simplicity of the preparation method seems to influ-
(found) differencea Importance Importance%
ence the choice of preparation conditions, the effectiveness was also %b
studied by some investigators. For example, the effect of different im-
pregnation strategies over Pt/ZrO2 catalysts promoted with Ce for DRM Metal 8.59 4.31 21 Catalyst
Support 8.07 3.79 18 Preparation
reaction was investigated by Özkara-Aydınoğlu et al. [53]. While the
Reduction 4.92 0.64 3 Variables 45
interaction between Pt and Ce increasing the oxygen storage of the Conditions
catalysts was stronger for co-impregnated Ce-Pt/ZrO2, sequentially Calcination 4.72 0.44 2
impregnated catalysts led to lower interaction between Pt and Ce. Conditions
Therefore co-impregnated catalysts had higher activity and resistance Catalyst 4.52 0.24 1
Preparation
to carbon deposition than the sequentially-impregnated catalysts [53]. Method
Reaction 12.84 8.56 41 Operational
3.1.4. Effect of pretreatment (calcination and reduction) conditions Temperature Variables 55
Feed Compositions 5.61 1.32 6
Calcination and reduction processes are also significant for the
W/F 5.49 1.21 6
performances of catalysts. The calcination variables are usually tem- TOS 4.68 0.40 2
perature and time while the hydrogen content of the feed is also im-
portant in reduction. The calcination temperature as high as 900 °C and a
The difference between RMSE without the variable or variables and RMSE of the
calcination time as high as 12 h appear in the database. There is also original model (4.28).
b
work performed without calcining the catalyst while some investigators Calculated as (RMSE difference for that variables)/(sum of differences).

systematically studied the influence of calcination conditions. For ex-

ample, Zhang et al. [40] investigated the effect of calcination

Fig. 6. Experimental versus predicted CH4 conversion for: (a) training, (b) testing data by the optimal ANN topology.

296
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Table 7 Table 7 (continued)

ANN prediction errors of each publication.
Article # of # of RMSE-Pa RMSE-Eb
Article # of # of RMSE-Pa RMSE-Eb Data Exp.
Data Exp.
Rezaei et al. [41] 75 6 6.5 7.6
Xu et al. [88] 45 10 19.8 6.2 Al-Fatish et al. [52] 40 11 44.6 26.6
Al-Fatesh et al. [102] 165 25 24.5 4.7 Hou et al. [34] 36 20 30.6 28.4
Yang et al. [69] 9 9 12.2 11.7 Habibi et al. [107] 119 8 9.7 5.2
Therdthianwong et al. [45] 12 10 22.9 20.4 Newnham et al. [89] 20 1 32.4 21.1
Zanganeh et al. [97] 40 10 7.0 8.9 Damyanova et al. [80] 18 2 25.7 13.8
Li et al. [112] 5 1 11.6 9.6 García-Diéguez et al. [65] 8 4 6.5 7.7
Zhang et al. [40] 30 3 10.3 6.2 Tankov et al. [104] 100 10 38.0 7.6
Pawelec et al. [37] 24 12 9.2 15.2 Mirzaei et al. [106] 91 12 24.4 18.6
Rezaei et al. [43] 20 4 19.0 12.1 Nematollahi et al. [113] 16 3 8.3 7.4
Meshkani and Rezaei [61] 16 4 22.8 23.3 Fakeeha et al. [115] 28 14 12.4 4.1
Yasyerli et al. [75] 131 13 31.8 4.3 Lee et al. [116] 20 8 17.3 30.0
San-José-Alonso et al. [46] 10 5 14.6 13.3 Shi and Zhang [85] 30 6 43.1 30.4
Xu et al. [67] 16 4 11.7 9.4 Ballarini et al. [90] 12 6 26.6 34.0
Sengupta et al. [117] 24 4 88.7 12.6 Sarusi et al. [73] 70 10 10.2 2.3
Zhang et al. [30] 28 7 5.6 8.2 Shang et al. [76] 52 15 13.3 9.3
Zhang et al. [29] 32 3 12.1 31.3 Nematollahi et al. [77] 25 5 7.6 8.6
Lv et al. [81] 9 1 57.0 11.9 José-Alonso et al. [79] 5 5 14.5 22.0
Pan et al. [42] 24 3 15.5 6.3 Özkara-Aydınoğlu and Aksoylu [58] 78 8 75.1 28.7
Nandini et al. [27] 251 21 13.3 9.5 Reddy et al. [63] 160 5 20.8 37.2
Barroso-Quiroga,Castro-Luna [62] 46 6 32.4 21.9 Damyanova et al. [51] 96 6 3.0 7.1
Kambolis et al. [68] 38 8 21.5 11.2 Özkara-Aydınoğlu et al. [53] 84 12 19.9 21.0
Alipour et al. [110] 104 12 13.6 8.9 Li et al. [44] 12 4 30.4 9.1
García et al. [56] 40 4 9.3 6.1 Wang et al. [36] 3 3 29.4 20.8
Meshkani et al. [103] 106 10 23.4 7.9 Rezaei et al. [31] 165 20 13.7 5.8
Alvar and Rezaei [49] 20 4 11.4 8.4 Yang and Papp [32] 18 2 20.4 21.0
Hadian and Rezaei [92] 25 6 11.7 10.1 Ballarini et al. [20] 40 7 16.4 17.4
Chang et al. [33] 6 6 44.3 5.6 Bouarab et al. [22] 24 4 48.6 22.2
Moniri et al. [59] 36 5 14.8 15.0 Takanabe et al. [25] 94 10 30.1 19.5
Al-Fatesh and Fakeeha 82] 48 13 26.8 16.4
a
García-Diéguez et al. [66] 30 5 8.1 6.5 Testing based on papers.
García-Diéguez et al. [64] 6 3 7.2 10.9 b
Testing based on individual experiment.
Xu et al. [86] 99 22 19.7 13.3
Yao et al. [96] 20 4 37.6 4.4
Zanganeh et al. [105] 9 3 22.6 28.8
Luisetto et al. [87] 53 6 12.6 3.7
temperature on the catalytic performance of Ni/TiO2-SiO2 catalysts that
Zhang et al. [38] 47 10 16.0 8.4
Özkara-Aydınoğlu and Aksoylu [78] 62 7 30.4 25.8
were prepared by impregnation combined with sol gel method, and
Rezaei et al. [50] 10 2 5.0 3.9 found that the catalyst calcined at 700 °C had higher activity and sta-
Al-Fatesh [94] 119 14 14.5 2.8 bility than the catalysts calcined at 550 °C and 850 °C for 5 h [40].
Wang et al. [95] 110 7 7.3 12.3 For reduction, on the other hand, the highest temperature in the
Liu et al. [60] 28 4 20.8 22.8
database is 850° C while the highest time is 10 h in various values of
Tang et al. [114] 70 14 17.2 16.3
Naeem et al. [118] 15 6 25.2 13.8 hydrogen content of the feed (up to 100%). For instance, Takanabe
Pompeo et al. [47] 30 6 17.7 12.4 et al. [25] investigated the influence of reduction temperature on the
Ranjbar and Rezaei [84] 123 13 35,4 4.5 catalytic performance of 10 wt.%Co/TiO2-anatase catalysts that were
Wang et al. [48] 9 1 19.2 33.7 prepared by incipient to wetness impregnation and calcined at 400 °C
Meshkani and Rezaei [71] 147 10 27.9 17.5
Hadian et al. [83] 111 10 4.2 4.8
for four hours. Before the reaction, the catalysts were reduced in situ
Bellido and Assaf [54] 144 4 29.7 40.2 with a H2 flow (25 ml/min) at different temperatures ranged from
Pompeo et al. [21] 9 3 11.2 5.3 700 °C to 950 °C for 1 h. The activities of the catalysts decreased with
Fajardo et al. [26] 14 2 13.4 6.0 increasing reduction temperatures. The catalysts reduced at 700–800 °C
Alipour et al. [1] 128 14 17.4 5.0
exhibited higher CH4 conversion and stability during 25 h than the
Al-Fatesh et al. [70] 76 20 9.1 4.3
Juan–Juan et al. [35] 44 8 28.6 8.0 catalysts reduced at 850–950 °C [25].
Halliche et al. [23] 12 1 44.1 27.0
Fakeeha et al. [93] 24 8 34.1 7.7 3.1.5. Effect of feed compositions
Vafaeian et al. [98] 24 4 36.1 16.2 Some papers used the mixture of CH4 and CO2 while some added
Takanabe et al. [24] 85 7 38.3 15.4
Serrano-Lotina and Daza [101] 20 3 13.6 15.4
inert gases (like Ar, He and N2) as diluent. The ratio of CH4/CO2 has a
Zanganeh et al. [100] 24 2 12.7 16.9 significant effect on catalyst activity and coke deposition, so different
Abdollahifar et al. [111] 6 1 14.8 7.5 values of carbon dioxide to methane ratios have been studied in order
Gonzaléz et al. [99] 174 9 11.3 14.2 to find out the optimum value. For example, the effects of this ratios on
Khajenoori et al. [109] 81 12 6.3 5.3
the 7 wt.%Ni–3 wt.%Co/La2O3-Al2O3 bimetallic catalysts at 800 °C in-
Xu et al. [57] 140 5 25.8 10.1
Qian et al. [108] 54 6 26.7 3.8 vestigated in the study of Xu et al. [67]. It was reported that CO2
Luengnaruemitchai and Kaengsilalai 48 3 44.3 19.3 conversion and H2/CO ratio increased from 60% to 99% and from 0.65
[39] to 1.38, respectively when the CH4/CO2 ratios was increased from 1/3
Jóźwiak et al. [28] 7 7 35.4 7.7 to 3/1. However, CH4 conversion declined from 96% to 58%. They also
Taufiq-Yap et al. [91] 162 12 38.4 6.0
Wu et al. [119] 70 7 16.3 14.7
found that, when CH4/CO2 ratio was higher than or equal to 1 in the
Ocsachoque et al. [74] 9 9 22.2 7.5 feed stream, the carbon deposition was also increased, hence CO2
Bellido et al. [55] 144 4 50.3 36.2 concentration in feed stream could be increased [67]. In another work,
Nagaraja et al. [72] 25 5 32.3 8.2 Yasyerli et al. [75] studied the effects of feed composition on the 1 wt.%
Ru-Ni/MCM-41 catalyst at 600 °C with gas hourly space velocity of

297
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

Fig. 7. Experimental verification of ANN model.

36000 ml/hgcat. A gaseous mixture of CO2/CH4/Ar with different vo- in a node) of 125 with the minimum test error percentages was ob-
lume ratios of 1/0.66/1.33, 1/0.5/1.5 and 1/1/1 was fed into the re- tained with 37 nodes, corresponding to prune level 3, training and test
actor. When the ratio of CH4/CO2 was higher than 1, the coke forma- errors of 18.5% and 19.9%, respectively. The optimal decision tree
tion and reactor blockage was observed [75]. As a result, the CH4/CO2 was shown in Fig. 5.
ratio of one or lower was used in the most of the work as it is also These error rates are equivalent to the accuracy rates of 81.5% and
apparent from Table 2. 80.1% for training and testing, respectively (the decision tree correctly
placed 80.1% of the data points not seen before; only 19.9 percent of
3.1.6. Effect of reaction temperature the new data was misplaced); the results are quite satisfactory. The
The most important variable for dry reforming of methane reaction accuracy distribution among the classes was also presented in confusion
is temperature due to the endothermicity of the reaction. DRM requires matrices, which compares the actual number in each class with the
higher temperatures to achieve higher conversion of CH4 and CO2. decision tree predictions (Table 3). As seen from the table, even the
However, the catalysts also lose their activity due to sintering of metals misplaced instances are not completely misplaced; most of them are in
or supports and carbon deposition. Therefore, to design a thermally neighboring classes (as if they leaked into the neighboring classes). The
stable catalyst having higher activity is the most important issue to lower accuracy in the middle class is an expected result in the decision
carry out this environmentally and industrially advantageous reaction tree (due to the leaks from both sides).
at the big scale [127]. As a result, the effect of reaction temperature was
studied in almost all work as expected, and temperature for the best 3.2.1. Developing heuristics
performance is highly dependent on the catalyst content (Table 2). The data was first divided, at the top of the tree according to the
reaction temperature indicating that this variable was the most sig-
3.2. Decision tree analysis of Ni based catalyst nificant factor determining the CH4 conversion level. Temperatures that
were lower than 625 °C led to lower CH4 conversion while temperatures
The Ni based data including 3966 instances with 52 attributes that were higher than 625 °C resulted in higher CH4 conversion for
were classified according to their CH4 conversion levels as 1288 in- some catalysts and operating conditions. This division was acceptable
stances for 0–50% (low) conversion class, 1416 instances for 51–75% and consistent with the results of the experiments because of the highly
(medium) conversion class, 1262 instances for 76–100% (high) con- endothermic nature of the DRM reaction. 625 °C was the split value
version class. These numbers are quite close to each other, hence there calculated by the learning algorithm of the tree and it is not an exact
should be no class imbalance problem [4]. Two-third (2644 instances) physical limit. The same is valid for the values of all the other variables
were used for constructing the tree and the remaining one-third (1322 so they should be thought as some empirical approximations. The
instances) were used as testing set. The optimal decision tree was second decision point was again reaction temperature. If the reaction
found by testing various split and pruning values, and comparing the temperature was higher than 775 °C, the desired conversion level was
testing error. The tree having splitmin (minimum number of instances possible only if the amount of ZrO2 in the support and the amount of Co

298
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

metal in the catalyst were lower than 0.47 and 6, respectively. In the 3.3. Artificial neural network analyses of entire data base
study of Luisetto et al. [87], they found that bimetallic Ni-Co/CeO2
catalyst was more active and selective than the monometallic Ni/CeO2 The artificial neural network analysis, which was performed on
catalyst because Co was highly effective to prevent carbon deposition entire data set, was used for two purpose in this work; 1) to determine
[87]. However, the amount of Co and Ni in catalyst was important relative importance of input variables (the variables that could be
because excessive amount of them could be resulted in deactivation of manipulated to improve performance), and 2) to predict the outcome of
catalyst. unstudied conditions. The first purpose requires the use of the entire
If the reaction temperature was lower than 775 °C (but still higher dataset, which could be also better for the second purpose even if we
than 625 °C), the desired conversion level was possible with three are interested in a specific metal; the catalyst containing other metals
different ways. In order to reach the desired conversion level, W/F, may improve the accuracy of model due to the common variables like
reduction temperature and some other variables have to be set. The temperature, W/F, feed composition and so on. Hence the ANN analysis
branches leading to high CH4 conversion were presented in Table 4. If explained in this section was performed using the entire dataset, and
the rows were read from left to right, the conditions (as if’s) described CH4 conversion was used as the response variable.
in cells may be treated as heuristics for high CH4 conversion. This
should be especially true for the first tree line considering that they 3.3.1. Constructing ANN model
are verified by a significant number of publications. Several network topologies were tested and compared with their
root mean square error (RMSE) values of training (error in predicting
3.2.2. Verifying the heuristics the data already used for constructing the model) and testing (error in
To verify the heuristics, first we used some experimental results predicting the data not seen before by the model) using 4-fold cross
over 8%wt.Ni 2%wt.Co/MgO, which is a catalyst we studied experi- validation. The training error is continuously decreasing as the size of
mentally in another project and not used in the database; the catalyst the network while the testing error decreases first and then increases
preparation and testing conditions are in accordance with heuristics 1 after the optimum network size as an indicator of overlearning. Hence
and 3. We also prepared a 7%wt·Ni-2%wt·Co/γ-Al2O3 catalyst and the optimum neural network topology was determined by using the
tested at various conditions that we can be used to verify all heur- value of testing RMSE, and a single layer 20 neuron network with
istics. tangent sigmoid function was found to be optimum. The experimental
The results are presented in Table 5; for each test in the table, the versus predicted CH4 conversion plots obtained for this network for
experimental conditions and CH4 conversion are presented with an both training and testing are given in Fig. 6, which indicates con-
explanation about the related heuristics. Normally, test 1 should satisfy siderably successful fittings. The RMSE and corresponding R2 values of
heuristic 1 and 3; indeed it satisfies heuristics 1, which required that training and testing are also quite satisfactory.
75% conversion should be achieved at 800 °C while it fails to conform
Heuristic 3, which state that 750 °C should be sufficient for 75% 3.3.2. Determining relative importance of input variables
(however, 70% CH4 conversion is still close to the expected 75%). It The relative importance of catalyst design and operational variables
should be noted that conditions forming these heuristics should not on CH4 conversion were analyzed using the optimal neural network
have to be exclusive; same conditions may obey more than one heur- topology. The relative importance of input variables was calculated by
istics. We have a similar problem in test 2, which has the same W/F the method of the change of root mean square error that is suitable for
value (1.43 mgcat min/ml) and reduction temperature but different both categorical and continuous attributes. One of the input variables
supports. Test 3, 4 were successful for heuristic 1 while they are slightly (such as metal types or temperature) was removed from the data set and
off from the prediction of Heuristic 2, which is valid in narrow tem- the network was trained with the remaining input variables, and the
perature range of 700–750 °C; the CH4 conversions are lower at 700 °C RMSE difference is used as the indicator of relative importance of the
while it is more than required 75% at 750 °C. Test 5 is for heuristic 4, variable; the larger deviations in RMSE shows higher relative im-
and it is successful. portance since the model was deteriorated more. This procedure was
Considering that the successful tests for heuristics 1 use high W/F repeated for all input variables. As it is shown in Table 6, the catalyst
ratio, and also noting that heuristic 1 and 3 requires high reduction preparation variables were found to have 45% relative importance
temperatures (like 800 °C), we performed another test (test 6) with a while operational variables had 55%. Among the operating variables,
higher W/F ratio (4.0) and this time, both heuristic 1 and 3 are sa- the reaction temperature was found to be the most important variable
tisfied. Then we inspected the database and found out that there is few (41%) affecting the CH4 conversion level. The active metal type (and
work reporting high conversion at low W/F. For example, only the weight percent) was found to be the second most important variable
work by Moniri et al. [59] out of 26 papers, which were used for (21%) followed by support (18%). These results are consistent with the
heuristic 1, was performed with W/F ratios lower than 1.43 (they used nature of DRM reaction. However, this analysis is still empirical; it is
1.0), and produced higher CH4 conversion values than 75%. Nor- valid within the range of variables employed in the literature, and the
mally, we eliminate the less frequently used values of variables be- high relative importance of a variable indicates that changing that
cause we cannot make generalization from few instances. However, variable in the range employed in the literature makes significant difference
the low W/F ratios in dataset were repeated in sufficiently high in the performance; in other words, the best value of that parameter is
numbers to keep in the model but apparently they went to the other not yet established or range is not yet narrowed down. For example, if
branches due to low CH4 conversion. Although, the model could be the best temperature was known and everyone used close values to that,
modified by going back and eliminating these few instances, this the relative importance of temperature would be low (results would not
course of action may decrease the number of data and it may weaken be changed much in that narrow range) even though theoretically this
the other conditions for high performance. Another approach would is the most important variables for DRM reaction. Having relative im-
be to keep the model as it is but aware of the weakness as we have portance in accordance with the theory simply indicates that the op-
done in this work. The weakness of the model in heuristic 2 is prob- timum setting of these variables is not yet established (still there is too
ably due to the fact that it is valid in narrow temperature range much variation in their values through literature)
(700–750 °C); DRM process is quite sensitive to the change in the
temperature, and it is not easy to measure catalyst temperature ac- 3.3.3. Testing the predictive power of ANN model
curately. However, considering that even the unsatisfied results are The 4-fold cross validation method (reserving one fourth of data for
still close to 75% CH4 conversion predicted by the heuristics, the testing and repeating this four times to cover entire range) is sufficient
decision tree model can be treated as successful. to analyze the relative importance of input variables. However, the

299
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

same procedure is not suitable to test the predictive power of model Fig. 7, the ANN prediction for 5 experiments out of 6 is quite successful;
because instances from the same paper (even from the same experi- only the prediction for Test 5 deviates from the experimental results
ment) may be distributed between training and testing set (for example, significantly. The success rate is much higher than in testing results in
some of the data points in a temperature profile may be in training set Section 3.3.3 although we do not have our own data in the database.
while the others in test set), and the predictions for test data may be This may be attributed to the fact that we carried out the experiments to
much higher than the true potential of the model. Hence we employed a verify the heuristics, hence they are in the most repeatedly studied
different procedure to test the predictive power of the model: we ex- region of experimental domains in which the ANN model is also strong.
cluded one paper from the model, trained the model with the other
papers and tested the predictive power of the model on the excluded 4. Conclusions
paper; we repeated this 101 times. As can be seen from Table 7, 17 of
101 papers were predicted with the RMSE values less than 10 (ap- The aim of this work was to extract knowledge for DRM reaction
proximately equivalent to have ± % 10 deviation in CH4 conversion) from experimental data using data mining tools such as decision tree
while 41 papers predicted with RMSE of less than 15; the number of analysis and artificial neural network; the data over Ni-based catalysts,
papers predicted with the RMSE of less than 20 (which can be still which is the most common catalyst for this reaction, were analyzed in
considered as acceptable) is 54. Although the model seems to have details and the following conclusion were drawn
some predictive power, the accuracy level is not sufficient for practical
purposes. • Decision tree analysis of Ni-based data was quite successful with the
Another approach would be to separate the data not based on pa- testing accuracy rate of 80%. The first decision point to separate
pers, but based on experiments (i.e., the data from a paper may be these heuristics was the reaction temperature as can be expected.
distributed between the testing and training set but the data from the This is followed by the other variables such as support type, W/F
same experiment may not be so). We observed in our previous work and reduction temperature. About 20% misplacement rate in DT
that if we divide the data this way, the predictive power of the model analysis is quite reasonable for such large and complex database.
increases significantly. • Decision tree analysis resulted in four heuristics (i.e. four course of
We also tested this approach, and for this, 753 independent ex- action) for high CH4 conversion, and they were either experimen-
periments in 101 papers were identified and the predictive power of the tally verified or generated results that are slightly lower than the
model was tested using similar to the procedure explained above: One predicted CH4 conversion.
experiment from 753 was removed and network was trained with the • ANN analysis of entire dataset revealed that operational variables
remaining experiments; this procedure was repeated for all experi- has higher relative importance (55%) compare to catalyst prepara-
ments, and the average of RMSE obtained from the experiments be- tion variables (45%). The most important operational variable was
longing the same paper is used for the prediction accuracy for that found to be the temperature while the active metal and the support
paper. As expected, the predictive capacity of the model increased are the most important catalyst preparation variables.
significantly: 48 papers had RMSE values less than 10 while 65 out of • ANN model was also tested to predict the data unseen before; the
101 papers were predicted with the RMSE of less than 15. The number model prediction of the test data from the database and the results of
of predictions with the RMSE of less than 20 was 77 out of 101. verification experiments were quite successful. (65 papers out of
To favor this approach, one possible explanation may be the fact 101 were predicted within 15% error while 77 papers had the error
that some unwritten information belonging to one experimental system less than 20%.
or to the laboratory is transferred to the model during training and used
in testing the data from the same group, and increases the predictive As a result, it can be concluded that analysis of past publications
power of the model. Another possible reason is that some papers tested using data mining tools could be beneficial to extract useful knowledge
significantly wide ranges of experimental conditions; excluding them in the form of heuristics and models that may help the future studies.
also weakens the model in some part of experimental domain, this risk
is reduced in the second approach. Acknowledgement
On the other hand, one may also argue about the possible con-
taminations of testing data if it is acquired from the same source with This work was supported by TÜBİTAK through 2211 National
training data. Although this is possible, it is unlikely that we may have Support Program for Master Students.
contamination significant enough to explain this improvement (other-
wise one could not build and test a model using his/her own data). Appendix A. Supplementary data
Another concern would be whether the applicability of the model de-
veloped this way is limited to the group contributed to the database or Supplementary data associated with this article can be found, in the
not (if unwritten information is transferred to the model). This is always online version, at http://dx.doi.org/10.1016/j.cattod.2017.05.012.
possible (model is still empirical). These two concerns are turned out to
be insignificant in the verification experiments discussed in the next References
section (note that we did not make contribution to the database, all the
experiments we performed are completely new to the model). [1] Z. Alipour, M. Rezaei, F. Meshkani, J. Ind. Eng. Chem. 20 (2014) 2858–2863.
[2] M. Usman, W.M.A. Wan Daud, H.F. Abbas, Renew. Sustain. Energy Rev. 45 (2015)
710–744.
3.3.4. Experimentally verifying the ANN model [3] A. Budiman, S.-H. Song, T.-S. Chang, C.-H. Shin, M.-J. Choi, Catal. Surv. Asia 16
Normally the model created using large number of variables and (2012) 183–197.
instances should not be validated using a few experiments. This is the [4] P.-N. Tan, M. Steinbach, V. Kumar, Introduction to Data Mining, first ed., Pearson
Addison Wesley, Boston, 2006.
main reason for the procedure employed in Section 3.3.3 (using all [5] K. Omata, N. Nukui, M. Yamada, Ind. Eng. Chem. Res. 44 (2005) 296–301.
papers or all experiments, and expecting successful results in most of [6] B.V. Ayodele, C.K. Cheng, J. Ind. Eng. Chem. 32 (2015) 246–258.
them) instead of experimental verification because it is hard to perform [7] M.A. Hossain, B.V. Ayodele, C.K. Cheng, M.R. Khan, Int. J. Hydrogen Energy 41
(2016) 11119–11130.
statistically sufficient number of experiment to verify a model created [8] M.E. Günay, R. Yildirim, Ind. Eng. Chem. Res. 50 (2011) 12488–12500.
with such large number of variables and data points. However, we al- [9] M.E. Günay, R. Yildirim, ChemCatChem 5 (2013) 1395–1406.
ready performed verification experiment for the heuristics, and we [10] Ç. Odabaşı, M.E. Günay, R. Yıldırım, Int. J. Hydrogen Energy 39 (2014)
5733–5746.
thought that it would be interesting to see whether the ANN model can
[11] N.A. Tapan, M.E. Günay, R. Yildirim, Chem. Eng. Res. Des. 105 (2016) 162–170.
predict the results of these experiments or not. As it was presented in

300
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

[12] N. Alper Tapan, R. Yıldırım, M. Erdem Günay, Bioprod. Biofuels Biorefin. 10 [67] J. Xu, W. Zhou, Z. Li, J. Wang, J. Ma, Int. J. Hydrogen Energy 35 (2010)
(2016) 422–434. 13013–13020.
[13] M. Baysal, M.E. Günay, R. Yıldırım, Int. J. Hydrogen Energy 42 (2017) 243–254. [68] A. Kambolis, H. Matralis, A. Trovarelli, C. Papadopoulou, Appl. Catal. A: Gen. 377
[14] V. Arcotumapathy, A. Siahvashi, A.A. Adesina, AIChE J. 58 (2012) 2412–2427. (2010) 16–26.
[15] D. Chesterfield, A.A. Adesina, J. Chem. Eng. Jpn. 42 (2009) s185–s198. [69] R. Yang, C. Xing, C. Lv, L. Shi, N. Tsubaki, Appl. Catal. A: Gen. 385 (2010) 92–100.
[16] E.V. Kondratenko, M. Schluter, M. Baerns, D. Linke, M. Holena, Catal. Sci. [70] A.S.A. Al-Fatesh, A.H. Fakeeha, A.E. Abasaeed, Chin. J. Catal. 32 (2011)
Technol. 5 (2015) 1668–1677. 1604–1609.
[17] U. Zavyalova, M. Holena, R. Schlögl, M. Baerns, ChemCatChem 3 (2011) [71] F. Meshkani, M. Rezaei, J. Nat. Gas Chem. 20 (2011) 198–203.
1935–1947. [72] B.M. Nagaraja, D.A. Bulushev, S. Beloshapkin, J.R.H. Ross, Catal. Today 178
[18] D.T. Larose, C.D. Larose, Discovering Knowledge in Data: An Introduction to Data (2011) 132–136.
Mining, second ed., John Wiley & Sons, 2014. [73] I. Sarusi, K. Fodor, K. Baán, A. Oszkó, G. Pótári, A. Erdőhelyi, Catal. Today 171
[19] D.T. Larose, Discovering Knowledge in Data: an Introduction to Data Mining, first (2011) 132–139.
ed., Wiley-Interscience, Hoboken, N.J, 2005. [74] M. Ocsachoque, F. Pompeo, G. Gonzalez, Catal. Today 172 (2011) 226–231.
[20] A.D. Ballarini, S.R. de Miguel, E.L. Jablonski, O.A. Scelza, A.A. Castro, Catal. [75] S. Yasyerli, S. Filizgok, H. Arbag, N. Yasyerli, G. Dogu, Int. J. Hydrogen Energy 36
Today 107–108 (2005) 481–486. (2011) 4863–4874.
[21] F. Pompeo, N.N. Nichio, O.A. Ferretti, D. Resasco, Int. J. Hydrogen Energy 30 [76] R. Shang, X. Guo, S. Mu, Y. Wang, G. Jin, H. Kosslick, A. Schulz, X.-Y. Guo, Int. J.
(2005) 1399–1405. Hydrogen Energy 36 (2011) 4900–4907.
[22] R. Bouarab, O. Cherifi, A. Auroux, Thermochim. Acta 434 (2005) 69–73. [77] B. Nematollahi, M. Rezaei, M. Khajenoori, Int. J. Hydrogen Energy 36 (2011)
[23] D. Halliche, O. Cherifi, A. Auroux, Thermochim. Acta 434 (2005) 125–131. 2969–2978.
[24] K. Takanabe, K. Nagaoka, K. Nariai, K.-i Aika, J. Catal. 232 (2005) 268–275. [78] Ş. Özkara-Aydınoğlu, A.E. Aksoylu, Int. J. Hydrogen Energy 36 (2011)
[25] K. Takanabe, K. Nagaoka, K. Nariai, K.-i Aika, J. Catal. 230 (2005) 75–85. 2950–2959.
[26] H.V. Fajardo, A.O. Martins, R.M. de Almeida, L.K. Noda, L.F.D. Probst, [79] D. San José-Alonso, M.J. Illán-Gómez, M.C. Román-Martínez, Catal. Today 176
N.L.V. Carreño, A. Valentini, Mater. Lett. 59 (2005) 3963–3967. (2011) 187–190.
[27] A. Nandini, K.K. Pant, S.C. Dhingra, Appl. Catal. A: Gen. 290 (2005) 166–174. [80] S. Damyanova, B. Pawelec, K. Arishtirova, J.L.G. Fierro, Int. J. Hydrogen Energy
[28] W.K. Jóźwiak, M. Nowosielska, J. Rynkowski, Appl. Catal. A: Gen. 280 (2005) 36 (2011) 10635–10647.
233–244. [81] X. Lv, J.-F. Chen, Y. Tan, Y. Zhang, Catal. Commun. 20 (2012) 6–11.
[29] W.D. Zhang, B.S. Liu, C. Zhu, Y.L. Tian, Appl. Catal. A: Gen. 292 (2005) 138–143. [82] A.S.A. Al-Fatesh, A.H. Fakeeha, J. Saudi Chem. Soc. 16 (2012) 55–61.
[30] M. Zhang, S. Ji, L. Hu, F. Yin, C. Li, H. Liu, Chin. J. Catal. 27 (2006) 777–781. [83] N. Hadian, M. Rezaei, Z. Mosayebi, F. Meshkani, J. Nat. Gas Chem. 21 (2012)
[31] M. Rezaei, S.M. Alavi, S. Sahebdelfar, Z.-F. Yan, J. Nat. Gas Chem. 15 (2006) 200–206.
327–334. [84] A. Ranjbar, M. Rezaei, J. Nat. Gas Chem. 21 (2012) 178–183.
[32] M. Yang, H. Papp, Catal. Today 115 (2006) 199–204. [85] C. Shi, P. Zhang, Appl. Catal. B: Environ. 115–116 (2012) 190–200.
[33] J.-S. Chang, D.-Y. Hong, X. Li, S.-E. Park, Catal. Today 115 (2006) 186–190. [86] L. Xu, H. Song, L. Chou, Int. J. Hydrogen Energy 37 (2012) 18001–18020.
[34] Z. Hou, P. Chen, H. Fang, X. Zheng, T. Yashima, Int. J. Hydrogen Energy 31 (2006) [87] I. Luisetto, S. Tuti, E. Di Bartolomeo, Int. J. Hydrogen Energy 37 (2012)
555–561. 15992–15999.
[35] J. Juan-Juan, M.C. Román-Martínez, M.J. Illán-Gómez, Appl. Catal. A: Gen. 301 [88] L. Xu, H. Zhao, H. Song, L. Chou, Int. J. Hydrogen Energy 37 (2012) 7497–7511.
(2006) 9–15. [89] J. Newnham, K. Mantri, M.H. Amin, J. Tardio, S.K. Bhargava, Int. J. Hydrogen
[36] R. Wang, X. Liu, Y. Chen, W. Li, H. Xu, Chin. J. Catal. 28 (2007) 865–869. Energy 37 (2012) 1454–1464.
[37] B. Pawelec, S. Damyanova, K. Arishtirova, J.L.G. Fierro, L. Petrov, Appl. Catal. A: [90] A. Ballarini, F. Basile, P. Benito, I. Bersani, G. Fornasari, S. de Miguel,
Gen. 323 (2007) 188–201. S.C.P. Maina, J. Vilella, A. Vaccari, O.A. Scelza, Appl. Catal. A: Gen. 433–434
[38] S. Zhang, J. Wang, H. Liu, X. Wang, Catal. Commun. 9 (2008) 995–1000. (2012) 1–11.
[39] A. Luengnaruemitchai, A. Kaengsilalai, J. Chem. Eng. 144 (2008) 96–102. [91] Y.H. Taufiq-Yap, U. Sudarno, Rashid Z. Zainal, Appl. Catal. A: Gen. 468 (2013)
[40] S. Zhang, J. Wang, X. Wang, J. Nat. Gas Chem. 17 (2008) 179–183. 359–369.
[41] M. Rezaei, S.M. Alavi, S. Sahebdelfar, Z.-F. Yan, J. Nat. Gas Chem. 17 (2008) [92] N. Hadian, M. Rezaei, Fuel 113 (2013) 571–579.
278–282. [93] A.H. Fakeeha, W.U. Khan, A.S. Al-Fatesh, A.E. Abasaeed, Chin. J. Catal. 34 (2013)
[42] Y.-X. Pan, C.-J. Liu, P. Shi, J. Power Sources 176 (2008) 46–53. 764–768.
[43] M. Rezaei, S.M. Alavi, S. Sahebdelfar, P. Bai, X. Liu, Z.-F. Yan, Appl. Catal. B: [94] A. Al-Fatesh, J. King, Saud Univ.—Eng. Sci. 27 (2013) 101–107.
Environ. 77 (2008) 346–354. [95] N. Wang, X. Yu, K. Shen, W. Chu, W. Qian, Int. J. Hydrogen Energy 38 (2013)
[44] X. Li, Q. Hu, Y. Yang, J. Chen, Z. Lai, J. Rare Earths 26 (2008) 864–868. 9718–9731.
[45] S. Therdthianwong, C. Siangchin, A. Therdthianwong, Fuel Process. Technol. 89 [96] L. Yao, J. Zhu, X. Peng, D. Tong, C. Hu, Int. J. Hydrogen Energy 38 (2013)
(2008) 160–168. 7268–7279.
[46] D. San-José-Alonso, J. Juan-Juan, M.J. Illán-Gómez, M.C. Román-Martínez, Appl. [97] R. Zanganeh, M. Rezaei, A. Zamaniyan, Int. J. Hydrogen Energy 38 (2013)
Catal. A: Gen. 371 (2009) 54–59. 3012–3018.
[47] F. Pompeo, D. Gazzoli, N.N. Nichio, Int. J. Hydrogen Energy 34 (2009) [98] Y. Vafaeian, M. Haghighi, S. Aghamohammadi, Energy Convers. Manage. 76
2260–2268. (2013) 1093–1103.
[48] Y.-H. Wang, H.-M. Liu, B.-Q. Xu, J. Mol. Catal. A: Chem. 299 (2009) 44–52. [99] A.R. González, Y.J.O. Asencios, E.M. Assaf, J.M. Assaf, Appl. Surf. Sci. 280 (2013)
[49] E.N. Alvar, M. Rezaei, Scripta Mater. 61 (2009) 212–215. 876–887.
[50] M. Rezaei, S.M. Alavi, S. Sahebdelfar, Z.-F. Yan, Scripta Mater. 61 (2009) [100] R. Zanganeh, M. Rezaei, A. Zamaniyan, H.R. Bozorgzadeh, J. Ind. Eng. Chem. 19
173–176. (2013) 234–239.
[51] S. Damyanova, B. Pawelec, K. Arishtirova, M.V.M. Huerta, J.L.G. Fierro, Appl. [101] A. Serrano-Lotina, L. Daza, Int. J. Hydrogen Energy 39 (2014) 4089–4094.
Catal. B: Environ. 89 (2009) 149–159. [102] A.S. Al-Fatesh, M.A. Naeem, A.H. Fakeeha, A.E. Abasaeed, Chin. J. Chem. Eng. 22
[52] A.S.A. Al–Fatish, A.A. Ibrahim, A.H. Fakeeha, M.A. Soliman, M.R.H. Siddiqui, (2014) 28–37.
A.E. Abasaeed, Appl. Catal. A: Gen. 364 (2009) 150–155. [103] F. Meshkani, M. Rezaei, M. Andache, J. Ind. Eng. Chem. 20 (2014) 1251–1260.
[53] Ş. Özkara-Aydınoğlu, E. Özensoy, A.E. Aksoylu, Int. J. Hydrogen Energy 34 (2009) [104] I. Tankov, K. Arishtirova, J.M.C. Bueno, S. Damyanova, Appl. Catal. A: Gen. 474
9711–9722. (2014) 135–148.
[54] J.D.A. Bellido, E.M. Assaf, Appl. Catal. A: Gen. 352 (2009) 179–187. [105] R. Zanganeh, M. Rezaei, A. Zamaniyan, Adv. Powder Technol. 25 (2014)
[55] J.D.A. Bellido, J.E. De Souza, J.-C. M’Peko, E.M. Assaf, Appl. Catal. A: Gen. 358 1111–1117.
(2009) 215–223. [106] F. Mirzaei, M. Rezaei, F. Meshkani, Z. Fattah, J. Ind. Eng. Chem. 21 (2014)
[56] V. García, J.J. Fernández, W. Ruíz, F. Mondragón, A. Moreno, Catal. Commun. 11 662–667.
(2009) 240–246. [107] N. Habibi, M. Rezaei, N. Majidian, M. Andache, J. Energy Chem. 23 (2014)
[57] J. Xu, W. Zhou, Z. Li, J. Wang, J. Ma, Int. J. Hydrogen Energy 34 (2009) 435–442.
6646–6654. [108] L. Qian, Z. Ma, Y. Ren, H. Shi, B. Yue, S. Feng, J. Shen, S. Xie, Fuel 122 (2014)
[58] Ş. Özkara-Aydınoğlu, A.E. Aksoylu, Catal. Commun. 11 (2010) 1165–1170. 47–53.
[59] A. Moniri, S.M. Alavi, M. Rezaei, J. Nat. Gas Chem. 19 (2010) 638–641. [109] M. Khajenoori, M. Rezaei, F. Meshkani, J. Ind. Eng. Chem. 21 (2015) 717–722.
[60] D. Liu, W.N.E. Cheo, Y.W.Y. Lim, A. Borgna, R. Lau, Y. Yang, Catal. Today 154 [110] Z. Alipour, M. Rezaei, F. Meshkani, Fuel 129 (2014) 197–203.
(2010) 229–236. [111] M. Abdollahifar, M. Haghighi, A.A. Babaluo, J. Ind. Eng. Chem. 20 (2014)
[61] F. Meshkani, M. Rezaei, Int. J. Hydrogen Energy 35 (2010) 10295–10301. 1845–1851.
[62] M.M. Barroso-Quiroga, A.E. Castro-Luna, Int. J. Hydrogen Energy 35 (2010) [112] J.F. Li, C. Xia, C.T. Au, B.S. Liu, Int. J. Hydrogen Energy 39 (2014) 10927–10940.
6052–6056. [113] B. Nematollahi, M. Rezaei, M. Asghari, A. Fazeli, E.N. Lay, F. Nematollahi, Fuel
[63] G.K. Reddy, S. Loridant, A. Takahashi, P. Delichère, B.M. Reddy, Appl. Catal. A: 134 (2014) 565–572.
Gen. 389 (2010) 92–100. [114] M. Tang, L. Xu, M. Fan, Int. J. Hydrogen Energy 39 (2014) 15482–15496.
[64] M. García-Diéguez, I.S. Pieta, M.C. Herrera, M.A. Larrubia, L.J. Alemany, Appl. [115] A.H. Fakeeha, M.A. Naeem, W.U. Khan, A.S. Al-Fatesh, J. Ind. Eng. Chem. 20
Catal. A: Gen. 377 (2010) 191–199. (2014) 549–557.
[65] M. García-Diéguez, E. Finocchio, M.Á. Larrubia, L.J. Alemany, G. Busca, J. Catal. [116] J.-H. Lee, Y.-W. You, H.-C. Ahn, J.-S. Hong, S.-B. Kim, T.-S. Chang, J.-K. Suh, J.
274 (2010) 11–20. Ind. Eng. Chem. 20 (2014) 284–289.
[66] M. García-Diéguez, I.S. Pieta, M.C. Herrera, M.A. Larrubia, L.J. Alemany, J. Catal [117] S. Sengupta, K. Ray, G. Deo, Int. J. Hydrogen Energy 39 (2014) 11462–11472.
270 (2010) 136–145. [118] M.A. Naeem, A.S. Al-Fatesh, A.E. Abasaeed, A.H. Fakeeha, Fuel Process. Technol.

301
A.N. Şener et al. Catalysis Today 299 (2018) 289–302

122 (2014) 141–152. M.A.F. Melo, D.M.A. Melo, Int. J. Hydrogen Energy 41 (2016) 14047–14057.
[119] H. Wu, G. Pantaleo, V. La Parola, A.M. Venezia, X. Collard, C. Aprile, L.F. Liotta, [124] R.K. Singha, A. Yadav, A. Agrawal, A. Shukla, S. Adak, T. Sasaki, R. Bal, Appl.
Appl. Catal. B: Environ. 156–157 (2014) 350–361. Catal. B: Environ. 191 (2016) 165–178.
[120] I. Luisetto, S. Tuti, C. Battocchio, S. Lo Mastro, A. Sodo, Appl. Catal. A: Gen. 500 [125] G. Moradi, F. Khezeli, H. Hemmati, J. Nat. Gas Sci. Eng. 33 (2016) 657–665.
(2015) 12–22. [126] E. Lovell, Y. Jiang, J. Scott, F. Wang, Y. Suhardja, M. Chen, J. Huang, R. Amal,
[121] C. Shi, P. Zhang, Appl. Catal. B: Environ. 170–171 (2015) 43–52. Appl. Catal. A: Gen. 473 (2014) 51–58.
[122] J. Estephane, S. Aouad, S. Hany, B. El Khoury, C. Gennequin, H. El Zakhem, J. El [127] D. Pakhare, J. Spivey, Chem. Soc. Rev. 43 (2014) 7813–7837.
Nakat, A. Aboukaïs, E. Abi Aad, Int. J. Hydrogen Energy 40 (2015) 9201–9208. [128] J.R.H. Ross, Heterogeneous Catalysis: Fundamentals and Applications, Elsevier,
[123] R.L.B.A. Medeiros, H.P. Macedo, V.R.M. Melo, Â.A.S. Oliveira, J.M.F. Barros, Amsterdam, 2012.

302