Journal of CO2 Utilization 61 (2022) 102034

Contents lists available at ScienceDirect

Journal of CO2 Utilization

journal homepage: www.elsevier.com/locate/jcou

Data analysis of CO2 hydrogenation catalysts for hydrocarbon production

Aleksandr Fedorov *, David Linke *
Leibniz-Institut für Katalyse e.V., Albert-Einstein-Str. 29a, 18059 Rostock, Germany

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

Keywords: Literature data of CO2 hydrogenation catalysts for hydrocarbon production via Fischer-Tropsch synthesis (FTS)
CO2 hydrogenation catalysts were extracted from literature, collected into a SQLite database, and analyzed using data science and machine
Data analysis learning techniques. A challenge was to include performance data obtained by different groups which vary in the
Machine learning
applied reaction conditions but vary also in the set of performance indicators that are reported. The thermo­
Anderson-Schulz-Flory distribution
dynamic ratio was used for analyzing reverse water gas shift reaction as the first stage of CO2 hydrogenation. A
tailored Anderson-Schulz-Flory distribution modified by two additional parameters was suggested based on the
data and applied for describing hydrocarbon production. The data analysis indicated the existence of direct CO2
conversion to hydrocarbons via the FT mechanism for some catalysts. It was shown that the chain growth
probability as the key parameter of catalyst selectivity slightly depends on reaction condition (temperature,
pressure, H2:CO2 ratio) and is mainly defined by catalyst composition, method of catalyst preparation, nature of
active sites and pretreatment conditions. With respect to the catalyst composition, it was found that doping
catalysts by alkali metals like K or Na is the most effective measure to improve catalyst performance leading to
increased chain growth probability, increased olefins content and a decrease of undesired CH4 formation.

1. . Introduction (RWGS) (Equation 1) followed by exothermic hydrogenation CO to

hydrocarbons (Equation 2) via FTS in the second stage. The RWGS has a
The generation of carbon dioxide by combustion gas, liquid, and thermodynamic constraint limiting CO2 conversion to a range of
solid fuels is one of the main reasons for global warming resulting from 20–50% for the range of conditions considered for CO2 FTS [16,17].
the greenhouse effect. For reducing and controlling CO2 emissions in the The second stage has a kinetic limitation wherein a proportion of
atmosphere, several strategies are considered [1]. The ones of perspec­ hydrocarbons can be described by the well-known Anderson-Schulz-
tive are clean energy technologies where obtained carbon dioxide is Flory distribution (ASF)* [18–20]:
captured and used for valuable compounds production [2,3] by hydro­
genation with H2. Such methods of CO2 conversion include methanation Wn = n(1 − α)2 αn− 1 ,
reaction, alcohol production [4,5], CO2 Fischer Tropsch synthesis (FTS)
where n – carbon number, Wn – mass fraction of hydrocarbon with n
[5,6]. The latter approach allows one to obtain hydrocarbons in the form
carbon atoms, α – chain growth probability. The physical meaning of the
of lower olefins [7,8] or fuels [9,10]. There are interesting two-steps
α parameter is a ratio of prolongation rate to the sum of rates of pro­
methods to produce high value-added hydrocarbons where obtained
longation and termination [21,22] because FTS is a typical polymeri­
hydrocarbons in the CO2-FTS are converted to olefins, aromatics, or
zation process [23]. In effect, the chain growth probability α defines
iso-paraffins on zeolite catalysts [11–13].
hydrocarbons distribution.
According to the literature [14,15], the CO2-FTS proceeds in the two
It is worth mentioning about the ASF distribution in the logarithmic
stages:
form:
CO2 + H2 ⇄CO + H2 O
Wn = n(lnα)2 αn− 1 ,
CO + 2H2 ⇄ − (CH2 ) − +H2 O that is an approximation to Equation 3 at α ≥ 0.5. Determination of
The first stage is the endothermic reverse water gas shift reaction chain growth probability using both equations from an experimental

* Corresponding author.
E-mail addresses: [email protected] (A. Fedorov), [email protected] (D. Linke).

https://doi.org/10.1016/j.jcou.2022.102034

Available online 17 May 2022

2212-9820/© 2022 Leibniz-Institut für Katalyse e.V., Albert-Einstein-Str. 29a, 18059 Rostock, Germany. Published by Elsevier Ltd. This is an open access article
under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

data give approximately the same results. products, hydrocarbon distribution etc.) was extracted from scientific
Iron is the most popular active component in CO2-FT catalysts articles and converted into an SQL database. The obtained database was
because it can catalyze both reactions 1 and 2. Bulk or supported Fe- used for analyzing and extracting new knowledge about CO2-FTS cata­
based catalysts without dopants produce a lot of methane in the pro­ lysts using data science methods. Particular attention was paid to
cess of CO2-FTS [24]. Because CH4 is one of the undesirable products, describing CO2-FT catalytic data with a minimum number of parame­
there are many publications about using different dopants for reducing ters, which helps to minimize correlation between catalytic performance
the methane production. Primarily, it is noteworthy that alkali metals measures and thereby improves the significance of the analysis.
[25–27] allow one to significantly improve the catalyst performance by
enhancing adsorption of CO2 on the surface of the catalysts [27]. Other 2. Software and hardware specifications
popular dopants are transitional metals like Mn [28], Co [8], Cu [29], Ru
[30,31], Zn [32] that increase the performance and stability of CO2-FT The SQLite database was used for collecting the data [46,47]. Python
catalysts. Some authors report a synergic effect between transitional and programming language (version 3.8) was used for calculations [48].
alkali metals [33]. The catalyst preparation method also strongly in­ SQLAlchemy was used to interface the database from Python [49].
fluences catalyst performance [34]. Scientific libraries SciPy [50], NumPy [51], Pandas [52], Scikit-learn
The CO2 hydrogenation catalysts must satisfy requirements of high [53] were used for the data evaluation; Plotly [54] was used to visu­
catalytic productivity and stability under reaction condition. A high alize the results.
productivity implies high activity for carbon dioxide conversion, high
C2+ hydrocarbons selectivity, high olefins-paraffins ratio, and low 3. Developing a CO2-FTS-tailored data analysis
values of CO and CH4 selectivity. The stability requires sustaining the
catalytic activity under reaction condition as well as mechanical sta­ 3.1. Analysis of RWGS reaction for CO2-FTS
bility. The latter is of special importance when applying a 3-phase slurry
reactor where the catalyst attrition rate has to be kept low [29,35,36]. In this part the methodology of CO2-FT catalyst analysis is described.
The main parameters of the CO2-FTS are temperature, pressure, H2/CO2 Based on the two-stages mechanism of CO2-FTS, Reactions 1 and 2 were
ratio, space velocity, and catalyst reduction pretreatment. considered separately. In this case of chemical reaction with thermo­
It is a difficult task to compare CO2-FT catalysts with respect to their dynamic limitation, the thermodynamic value (equilibrium factor, ratio)
performance data that were obtained in different ways and at different η is usually used for analyzing the closeness to equilibrium [33,55,56].
reaction condition. Moreover, the multiplicity of catalytic features (CO2 Thermodynamic ratio η of the RWGS reaction can be calculated by the
conversion, selectivity etc.) do not facilitate this task. Ideally, the following equation:
comparison of catalyst activity should be made at the same reaction
condition (temperature, partial pressure of reagents etc.). However, PCO PH2 O
η= ,
recalculation and normalization of the catalytic data measured at the PCO2 PH2 K
different condition requires knowledge about the reaction kinetics and where Pi – partial pressure of i-compound in the flow from the
activation energy that are usually unknown. Therefore, improved ap­ reactor, K – equilibrium constant, that can be calculated by the following
proaches to normalization are valuable. They assist the researcher in equation [57]:
understanding dependencies and relationships between catalyst char­ ( )
acteristics and activity. Such analysis benefits from applying data sci­ K = exp −
4577.8
+ 4.33 ,
ence and machine learning methods. These methods can help in T
analyzing, systematizing a large amount of the data, and predicting new
materials under requirement constraints [37]. Besides, machine where T – temperature in Kelvin scale. The partial pressures of CO2, H2,
learning techniques can reveal latent knowledge from the scientific data CO, and H2O were estimated by using simple calculation described in
consisting inside [38]. the Supporting information. The RWGS reaction is at equilibrium for a
In the case catalytic data, Smith et al. [39] used neural network for thermodynamic ratio of η = 1. In the case of η < 1 the reaction rate is
analyzing catalysts for water-gas shift reaction and influence reactions directed towards the formation of CO and H2O (forward reaction).
condition and catalyst composition on catalytic activity. Keisuke Suzuki
et al. [40] analyzed the catalytic data from the literature on oxidative 3.2. Analysis of hydrocarbon distribution
coupling of methane (OCM), water gas shift, and CO oxidation reactions
and created gradient boosting regression models with XGBoost for pre­ The second stage of CO2-FTS (Reaction 2) has a kinetic limitation. In
diction of catalytic performance. Based on the made analysis, the top 20 theory, hydrocarbon selectivity may be described by only one parameter
promising catalyst candidates for the OCM reaction were suggested for α of the ASF distribution. It means direct correlations exist between
the future investigation. It is worth mention earlier works [41,42] about proportions for one hydrocarbon (C1, C2 etc.) or any range (e.g., C2–4,
the analysis of OCM catalyst database that are directed at analysis the C5–15). Fig. 1 shows the dependence between theoretical mass fraction of
literature data. In the case of CO2 hydrogenation catalyst, Qingxin Yang hydrocarbons in different fraction and chain growth probability α. The
et al. [43] used regression trees and ANOVA for statistical analysis of the calculated values of CH4 and the fraction of typical products for the
literature data of CO2 hydrogenation catalysts. It was confirmed that different change growth probabilities of Fig. 1 are compared in Table 1.
doping iron catalyst by alkaline and transition metals increases C2+ One can see that there are restrictions of target fraction because of
selectivity and olefins/paraffins ratio that is completely consistent with the kinetic limitation imposed by the assumed polymerization mecha­
the previous results [8,26,44,45]. Authors of that work limited their nism. Moreover, CH4 proportion strongly depends on chain growth
analysis to some key features like a rate of CO2 hydrogenation, a rate of probability. For example, maximum production of C2-C4 fraction at α
C2+ production, C2+ selectivity, and olefins/paraffins ration but did not = 0.47 corresponds to 28.1% mass fraction of methane. In fact, the
consider hydrocarbon distribution and chain growth probability α. The desire to decrease CH4 selectivity leads to hydrocarbon production with
last one is known to be a key parameter of CO2-FTS. higher values carbon number. The chain growth probability α also de­
Thus, the main aim of this investigation is a detailed analysis of the fines the maximum selectivity of CO2-FTS and thus is the key parameter
literature data of CO2 hydrogenation catalysts. For this aim, the data of the CO2 hydrogenation catalyst.
with all information about catalysts such as composition, preparation The obtained database contains around 500 experimental data points
method, reduction pretreatment condition etc., and the information with catalytic information about CO2-FT catalysts. All of them have the
about catalytic data (reaction condition, conversion and selectivity of information about reaction conditions (temperature, pressure, H2/CO2

2
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

Fig. 1. The dependencies between the chain growth probability and theoretical
mass fraction of hydrocarbons in different range (C2-C4, C5-C15, C8-C16) among
all hydrocarbons formed by FTS calculated by Equation 3.

Table 1
Fig. 2. The circle diagram of the information reported to describe the hydro­
Mass fraction of products at different hydrocarbons range for fixed the chain
carbon distribution in the analyzed papers.
growth probability.
Range CH4, % target fraction, % other products, % αa
( )
C2-C4 28.1 56.7 15.2 0.47
(1 − γ)Wn (α)
Sn = (1 − SCO ) , at n = 3, 4 … ∞,
C5-C15 4.0 59.7 36.3 0.80 1 − (1 − δ)W2 (α)
C8-C16 2.6 40.7 56.7 0.84
a where Si – Ci selectivity, SCO – CO selectivity, n – carbon number, α –
chain growth probability
chain growth probability. The δ parameter represents a fraction of C2
responding for a decrease of the ethylene selectivity compared with the
ratio, catalyst composition etc.). Concerning catalytic data there are
value calculated by the ASF distribution. The empirical parameter γ
different ways presenting this information in the articles. In general,
responds an excess of CH4 selectivity.
researchers report only selected data e.g., only C1, C2–4 and C5+ selec­
One can see that only 9% of the experiments are reported with the
tivity and CO2 conversion, CO selectivity, and olefins/paraffins ratio.
full information about hydrocarbon distribution (up to C12). In this part
Some works (around 23%) have the information about the chain growth
of article, only this data subset was discussed and analyzed.
probability. What is reported varies between papers, which creates some
According to equations 4–6 hydrocarbon selectivity depends on only
difficulties in directly comparing the catalysts. But as will be seen below,
4 parameters: SCO , α, γ, δ wherein CO selectivity is usually known. It is
even this information is enough to extract some interesting catalytic
worth noting that all the parameters are in a range from 0 to 1. The other
features of CO2-FT catalysts. Fig. 2 shows the circle diagram of the in­
parameters of the equations can be calculated by minimization of the
formation that was reported to describe the hydrocarbon distribution in
following cost function:
the analyzed papers. It is also worth mentioning alcohol formation for
∑ ( exp )2
CO2-FTS which were reported in ~ 22% articles of the database (for Si − Spred
i
example in [7]). However, most articles do not contain any information
about alcohols or other oxygenates. The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm [58] was
For comparing the hydrocarbon distributions from different experi­ used for minimization of the cost function. The chain growth probability
ments, the data was converted to one line with fixed parameters using can be also calculated from the slope of a curve describing hydrocarbon
linear transformation. Detailed description of this approach can be distribution for C3+ in corresponding ASF coordinates if the information
found in Supporting information. Fig. 3 shows mean hydrocarbon dis­ about selectivity distribution exists. The parameter α calculated in this
tribution of the CO2-FT catalysts plotted in ASF coordinates. The 99.7% way allows one to calculate the other parameters γ and δ. The advantage
confidence interval also is presented on Fig. 3. One can clearly see that of the approach via optimization is that it opens the opportunity to
C3+ hydrocarbons correspond to the ASF distribution but CH4 and C2 do calculate the parameters even when data are given in various ways (S1,
not. Thus, using only the ASF distribution does not allow one to describe S2 …, or ranges S2–4, S2–6 …). Both approaches give similar results (see
the experimentally observed hydrocarbon distribution of CO2-FT cata­ Fig. S4). It is worth noting that the majority information about hydro­
lysts accurately. It is necessary to consider deviations at least of C1 and carbons is given in the form C1, C2–4 and C5+ selectivity (see Fig. 2). It is
C2. Empirically, the hydrocarbon selectivity can be described by the impossible to calculate the three parameters of equations 4–6 by using
following equations: these data with only two degrees of freedom. However, it was found that
( ) a decrease of C2 selectivity is around 10% (the δ parameter) for most
(1 − γ)W1 (α) catalysts. By setting δ to 0.1 for those data, α and γ parameters can be
S1 = (1 − SCO ) γ +
1 − (1 − δ)W2(α) calculated from the data having only a two degree of freedom. This
( ) approach was checked by the data with the information about hydro­
(1 − γ)(1 − δ)W2 (α)
S2 = (1 − SCO ) carbons distribution for which these parameters can be calculated
1 − (1 − δ)W2 (α) directly (see the Supporting information). Thus, using this approxima­
tion allowed us to estimate the parameters α and γ even for the data

3
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

Fig. 3. Hydrocarbon distributions of the CO2-FT catalysts plotted in ASF coordinates for all hydrocarbons (left), olefins (middle), and paraffins (right).

presented in the form of C1, C2–4 and C5+ selectivity or similar

representation.

3.3. Descriptor selection analysis

The descriptor selection (feature selection) is the most important

stage for building reliable machine learning models. There are different
approaches for finding and determining important descriptors, for
example Pearson’s R correlation coefficient, lasso regression, random
forest, gradient boosting etc. [59,60]. For understanding the influence of
reaction and catalyst parameters on performance and catalytic activity
(CO2 conversion; CO selectivity, the chain growth probability etc.) the
random forest regression was used in this work. The first stage is to
present the catalytic data as vectors using the set of descriptors. The
following descriptors were considered:

• Temperature reaction, pressure reaction, H2:CO2 ratio, GHSV,

reduction temperature
• Bulk or supported catalyst, presence of alkali or transition metals Fig. 4. Histogram of the distribution of alkaline containing/free catalysts and
(has K or has Me) mean values of CO selectivity and CO2 conversion as a function of the ther­
modynamic ratio η.
• Type of support (Al2O3, SiO2 etc.)
• Cu content, Mn content etc.
• Alkali metals content worth noting that the portion of the alkaline-doped catalysts in all cat­
alysts is around 63% whereas it is around 88% in the case of η > 1. The
Bulk or supported catalyst, presence of alkali or transitional metals, Fisher’s exact test [61] was used for validating the influence catalyst
and similar categorial descriptors were presented as 0 and 1. Type of doping by alkali metal on thermodynamic ratio. A significance level was
supported was encoded by one-hot (one-of-K) encoding [53]. Other selected equal to 0.01. The two samples of experimental data with η ≤ 1
descriptors were presented as real number (temperature, Mn content and η > 1 were created. Elements of each one were equal to 1 or 0 (1 if a
etc.). The performance of models obtained by 10-fold cross validation is catalyst consists of alkali metals). The calculated p-value was 2.4⋅10− 5
shown in the Supplementary information (see Fig. S20-S21). (< 0.01) that allows one to reject a null hypothesis that both distribu­
For analyzing the deviation of the descriptor contributions two ap­ tions are equal. It can be concluded that an increased value of thermo­
proaches were used. The values of targets parameters were changed by dynamic ratio is characteristic of the catalysts doped by alkali metals
adding random noise in a range ± 5% with subsequent using 10-fold that was proven by statistics. Possible reasons for η > 1 are discussed
cross validation. The obtained arrays of descriptor contributions were below.
used for calculation of the mean values and the deviations. 5% and 95% The presented data allow to conclude that the equilibrium of the
quantities were used for determination of a low and high values of de­ RWGS reaction for CO2-FTS on Fe-based catalysts is not established in
viations. For calculating the correlation of descriptors with target data general. Fig. 4 also shows the mean value CO selectivity SCO and CO2
(positive or negative) a linear regression with L2-regularization conversion XCO2 for each group (bars of the histogram). One can see that
(parameter of regularization was set 10− 3) was chosen as the first there are correlations between the thermodynamic ratio η and SCO , XCO2
approximation. in the case of η ≤ 1. A decrease of CO selectivity accompanied by an
increase CO2 conversion is known to be a consequence of two stages
4. Analysis and discussion mechanism of CO2-FTS. It was found and proven by the analysis of the
dependence of products selectivity on CO2 conversion [10, 34, 62–64].
4.1. The approach to equilibrium of RWGS reaction An increase of thermodynamic ratio is accompanied by an increase of
XCO2 and a decrease of SCO . A high CO2 conversion corresponds to a high
Fig. 4 shows the distribution histogram of the thermodynamic ratio η RWGS reaction rate and, thus, the thermodynamic equilibrium of the
for all experimental data. One can see that there is data (its portion is RWGS reaction may be achieved.
around 18%) with a thermodynamic ratio η value of more than 1 and The cases of η > 1.0 may results from other reactions leading to
that this is more common for catalysts doped by alkaline metals. It is

4
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

either reduced reagents pressure (CO2, H2) or to increased production influence of the methanation reaction is weak. The Reaction 3 (see
pressure (CO, H2O). Such a reaction could be coke formation and sub­ Fig. 5) is a direct hydrogenation of CO2 to hydrocarbons that can lead to
sequent reaction carbon dioxide: η > 1. Thus, the analysis of the literature data of CO2-FT catalysts in­
dicates the presence of a direct path of CO2 hydrogenation to
C + CO2 ⇄2CO,
hydrocarbons.
that has a thermodynamic limitation. The equilibrium constant is
around 10− 8 at 250 ºC [65]. The calculated pressure of CO at this tem­ 4.2. Analysis of CO-FT reaction
perature and a pressure of CO2 of 15 bar is only 3.98⋅10− 4 bar which is
orders of magnitude below the experimental observation. It is The hydrocarbons selectivity of Fe-based catalysts for CO2 hydro­
well-known that the Boudouard Reaction usually plays a significant role genation was shown to correspond to the ASF distribution except for a
only at higher temperature (T > 900 ºC) [66]. The calculated values for decrease of ethylene selectivity and an increase of CH4 one (see Fig. 3).
the thermodynamic ratio would be lower if alcohols are formed. How­ This result suggests that for the calculation chain growth probability
ever, at typical alcohol selectivity (5% MeOH) the thermodynamic ratio only C3+ hydrocarbons should be used. In practice, the various hydro­
would hardly change. For example, in the case of using ZnFeOx-5.76Na carbons ranges were used for calculating chain growth probability. So,
catalyst for CO2-FTS [67] oxygenates selectivity was 4.9%. Assuming for examples, in work [73] C1-C5 were used, C1-C7 [74], C2-C7 [75],
that the main product of oxygenates are alcohols, thermodynamic ratio C3-C10 [76]. There are also some works without description of the
was calculated and was 0.48. If alcohol selectivity was ignored the calculation methods like here [44,77].
estimated value would be 0.52. In the case K-Fe/ZrO2 catalyst [7], the It is worth mentioning that some papers report deviations from ASF
values of ratio are 0.99 and 1.33. Methanol selectivity was 20% in last distribution, like for Fe-Mn-K catalyst in this work [9] or for Co/MnOx in
case. this one [72]. Thus, one should carefully check the applicability of the
The reason of an increase of η may be methanation reaction (direct ASF distribution for describing hydrocarbon selectivity. In the case of
CO2 hydrogenation to CH4). Only the catalysts with η > 1.5 were our data, this ASF approximation is reasonable for most CO2-FT catalysts
considered for further analysis to avoid an influence of systematic errors as proven by the analysis shown in Fig. 3 and Fig. S3.
in the experimental data (measurement the reaction temperature, There are some reasons for the increased methane selectivity in
determination of product concentrations etc.) and in the estimation of comparison to the theoretical ASF distribution (see Fig. 3). The first one
thermodynamic ratio (see Table S1). Firstly, it is worth pointing out the is hydrogenolysis of olefins [31,34]. The other reason may be direct
Co-containing catalysts investigated in works [68,69]. In this work [69], hydrogenation of CO2 (or CO) to CH4 not via FT that were discussed in
the calculated values of η are in a range from 1.53 to 5.83 wherein excess several works [14,34,62,64]. Fig. 3 also presents olefin and paraffin
of CH4 selectivity γ are in (0.06, 0.90). In the work [68], the thermo­ distribution in corresponding coordinates. It shows a decrease in
dynamic ratio is around 1.85 with γ value is 0.95 for the Co-based ethylene selectivity compared with the value calculated by the ASF
catalyst. The Co-based catalysts are known to be hardly active in distribution that does exist in the case of paraffin distribution. It is worth
RWGS reaction and but to have high hydrogenation ability. This may noting the larger confidence intervals for the paraffins compared to
explain the high values of η an increase of which may be also relate to olefins. It may relate to a higher error of quantifying long paraffins in
methanation reaction. The high hydrogenation ability is confirmed by presence of the corresponding olefins which when mainly olefins are
the data about the fraction of olefins in C2-C4 hydrocarbons that are formed (an excess of olefins is common for the data published with the
practically absent for these Co-containing catalysts. Secondly, a rise of full hydrocarbon distribution information). Also, for C12+ paraffins it
the thermodynamic ratio (in a range of 1.72–2.98) is also found for becomes experimentally challenging to prevent partial condensation
Fe-based catalyst supported on carbon nanotubes [70,71] doped with or which could affect the accuracy of online analysis data. Concerning
without alkaline metals. Wherein the values of γ are low (not more than ethylene selectivity, a decrease may be explained by the high reactivity
0.17) and even equal to zero in some cases. A similar behavior is also of C2H4 compared to higher α-olefins [55,78].
found for several bulk and supported Fe-based catalysts doped by The hydrocarbon selectivity of CO2-FT catalysts can be described by
alkaline metals [11,17,31,55]. The data with high values of η and low the suggested empirical equations 4–6 that depend on 4 parameters SCO,
ones of γ cannot be explained by the methanation reaction. However, a α, γ, and δ. CO selectivity is usually known. The α parameter is the chain
direct hydrogenation of CO2 to hydrocarbons may explain this phe­ growth probability that defines the selectivity of CO2-FT catalysts. The γ
nomenon as was discussed in [14]. So, Zhenhong He et al. [72] parameter describes the excess of CH4 selectivity. The parameter δ de­
demonstrated for Co/MnOx catalysts this direct CO2 hydrogenation to fines a decrease of the ethylene selectivity. It was set to 0.1 to estimate α
hydrocarbons without CO production. They suggested that the reduction and γ values for cases where data were presented with only two degrees
of adsorbed CO2 by hydrogen to CH2 and CH3 monomers proceeds via of freedom. It is worth mentioning, another important characteristic of
formation of COδ− 2 , HCOO , − CH2OH, and/or CH3O intermediates.
− −
the CO2-FT catalysts that is olefin-paraffin ratio (or the portion of olefins
Fig. 5 presents different pathways for CO2 hydrogenation. The in hydrocarbons) which was also included in the analysis. The rate of
pathways 1–2 are the two-stage CO2-FT process via RWGS reaction and CO2 conversion was used as a measure of catalyst activity. It was
subsequent hydrogenation of CO to hydrocarbons. They cannot describe calculated by the following Equation:
the increase of the thermodynamic ratio for those cases where the

Fig. 5. Different pathways for CO2/CO hydrogenation. * - surface intermediates.

5
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

f • XCO2 olefins content, GHSV and temperature slightly influence it. One can see
r= ,
mcat a decrease of the olefin content with rising pressure and H2:CO2 ratio. It
relates to an increase hydrogen pressure resulting to acceleration of
where mcat – mass of catalyst; f – molar flow rate of CO2 at reactor inlet; hydrogenation reactions. The excess of CH4 selectivity γ also weakly
XCO2 – CO2 conversion. depends on reaction condition. One may point at certain tendency that γ
For analyzing the alcohol production, alcohol content walc. was rises at high temperature which is observed for most catalysts.
calculated by the following equation: The influence of reaction conditions on other catalyst parameters
ṅalc. such as CO2 conversion XCO2 , CO selectivity, CH4 selectivity, C2+
walc. =
ṅalc. + ṅHC
• 100%, selectivity, and thermodynamic ratio is shown in the Supporting infor­
mation (Fig. S6-S9). In the case of XCO2 , there is a significant increase
where ṅalc. – molar flow of alcohol in the flow from reactor; ṅHC – molar with rising temperature, pressure and H2:CO2 ratio and a decrease with
flow of hydrocarbons in the flow from reactor. We assumed that the space velocity that relates to a change of reaction rate. CO selectivity
main product of alcohols and oxygenates is methanol. strongly decreases with rising pressure and H2:CO2 ratio resulting in an
In the next sections of this work, the influence of reaction condition, increase of C2+ selectivity. In the case of temperature influence on CO
catalyst composition, catalyst preparation and activation on the sug­ and C2+ selectivity, non-linear dependencies with an extremum are
gested parameters of CO2-FT catalysts is analyzed. found for most catalysts.
In summary, the differences in terms of hydrocarbon distribution
4.2.1. Influence of reaction condition between catalysts are larger than the changes induced by varying the
The influence of reaction condition such as temperature, pressure, reaction conditions such as temperature, pressure, H2:CO2 ratio, and
H2:CO2 ratio, and GHSV on catalyst performance has frequently been GHSV. Therefore, one can conclude that parameters such as catalyst
investigated in the literature [27,28,45,55,68,71,79–81]. Using this composition, method of catalyst preparation, nature of active sites, and
literature data, the suggested parameters for CO2-FT catalysts such as α, pretreatment conditions define α mainly.
γ, olefin content of CO2-FT catalysts were calculated. The values of these The presented data also demonstrates an advantage of using the
parameters are shown in Fig. 6 as a function of the reaction conditions. suggested parameters for describing CO2-FT catalysts. For example,
From Fig. 6 one can see that reaction parameters have a weak in­ Fig. S6 shows the dependencies of the catalysts main parameters at
fluence on the chain growth probability. It practically does not depend different temperature for K-Fe catalysts from article [80] (green dots)
on GHSV and H2:CO2 ratio. There is a little decrease of α at pressures and for Mn-Fe catalyst from [81] (purple dots). One can see that CH4
below 10 bar. It is worth noting various dependences between α and selectivity is higher for the Mn-containing catalyst. For the latter the
reaction temperature for different catalysts. Despite this, the values of α high CH4 selectivity is due to the lower chain growth probability.
at different temperature are in a narrow range for each catalyst. For However, the higher value of excess of methane selectivity γ for
example, values of α are in a range 0.75–0.85 at temperatures between K-containing catalyst clearly shows that an undesired (non-ASF) route
260 and 340 ºC for bulk K-Fe catalyst [80]. In the work of Shunwu Wang for methane formation is a bigger problem for this catalyst than for the
et al. [71], α are in a range 0.66–0.73 at 270–380 ºC. In the case of other (Mn-containing) one.

Fig. 6. The dependencies between the parameters of different CO2-FT catalysts and reaction temperature, pressure, and H2:CO2 ratio [27,45,55,68,71,79–81]. The
same color and shape mean one specific catalysts from one article.

6
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

Concerning the catalyst deactivation under CO2-FT condition, there of doping by Li is weaker than for other alkali metals.
are only few works where it was investigated. Due to lack of data The influence of K or Na content on the characteristic CO2-FT pa­
deactivation could not be analyzed in the same way as activity and rameters is presented in Fig. 9. One can see that adding alkali metals
selectivity. The reason of deactivation may be carbon deposit on catalyst greatly increases the catalysts´selectivity to olefins and hydrocarbons
surface which was reported for iron-based catalysts doped by alkali with high carbon number. Simultaneously they reduce the loss of CO2
promoter (K, Rb, Cs) [82,83]. An oxidation-reduction treatment can via additional methane formation (see the decrease of excess of CH4
regenerate the deactivated Fe-K/alumina catalyst [82]. Another reason selectivity γ in Figs. 8 and 9). Moreover, doping Fe-based catalysts by
of deactivation is phase transformation of FeCx to FeOx by oxidation alkali metals increases CO2 conversion (see Fig. S10–13), thus, improve
with CO2 and H2O [32,84]. The dopants like Na and Zn were shown to catalysts performance. It is interesting that doping by K can increase
be able to suppress the oxidation process [32]. It is also worth catalyst performance even for CO2-FT catalysts based on Fe-containing
mentioning the growth of iron carbide crystallinity resulting in industrial waste that was demonstrated by Artem Russkikh et al. [25].
component separation as one of reason of catalyst deactivation that was It emphasizes the general effectivity of using alkali dopants. It was also
found for Fe-Cu-K-Al catalyst in the work [85]. found that doping by alkaline metals increases the alcohol production of
CO2-FT catalysts. So, the mean value of alcohol content walc. for alkaline
4.2.2. Influence of catalyst composition doped catalysts calculated from the obtained database is around 18.5%
This section addresses the influence of catalyst composition on the and higher than the value for alkaline-free catalysts (4.3%). It can be
key parameters of CO2-FT catalysts. Many different compositions of clearly seen in the dependency between alcohol content and CO2 con­
catalysts are reported in the literature. The dominating main component version (see Fig. S20).
is Fe. In some works (for example this [69]) Co-based catalysts were There are series of work where the effect of alkali dopants was
investigated. The most popular dopants are alkali metals like K and Na. investigated and analyzed. Numpilai Thanapha et al. [8] explained the
Fig. 7 shows distribution of elements in CO2-FT catalysts from our effect of K by the enhanced bonding strength of adsorbed CO2 and H2
database plotted on the periodic table. Mn, Co, and Cu are found to be leading to a decrease of the hydrogenation activity of K-containing
the most employed dopants except for alkali metals. Wherein 70% of all catalyst. This concept was proven by in situ XPS and DFT calculation in
the doped catalysts consist of these ones. Less popular are Zn, B, Mo, Zr, [87] and DFT calculation in [89]. Besides affecting adsorption, alkali
and La. metals promote formation and stability of Fe5C2. This was reported for
Fig. 8 shows values of the chain growth probability, olefins content, alkali-modified catalysts where activity directly correlated with alkali
and an excess of CH4 selectivity for catalysts with different alkali dop­ content and amount of iron carbide [26,90]. It is noteworthy that Fe5C2
ants and tested under the same reaction condition. One can see that K- can be transformed to FeOx with a loss of CO2-FT activity under reaction
doped catalysts are optimal due to providing highest values of α and ω, condition [27] which further confirms the key role of alkali metals. But
and a low value of γ. This is in line with prior work which reported on the other hand, there are other works about active CO2-FT catalysts
potassium as the preferable alkali dopant [24]. Doping by alkali metals that do not require doping by alkali metals [34,62]. Andrey Skrypnik
results in a rise of CO2 conversion and a decrease of CO selectivity, and, et al. reported around 60 wt% of Fe2C5 for a catalyst obtained from iron
therefore, in an increase of hydrocarbons production. More detailed (II) oxalate [62]. This indicates on an important role of preparation
results on the influence of K can be found in the Supporting information methods and catalyst pretreatment in the synthesis of highly-active and
(Fig. S10). It is worth mentioning that doping by Li increased the chain selective CO2-FT catalysts.
growth probability and olefins content for some authors [24,86] Concerning other dopants, the most popular transitional metals like
although the opposite was reported in [7]. However, the positive effect Mn, Co, and Cu were analyzed in more detail (see Fig. S14-S19). It was

Fig. 7. Distribution of elements used for preparation of CO2-FT catalysts plotted in the periodical table. % means a portion of catalysts with this element in the
obtained database. The catalysts doped by alkali metals present separately and are labeled by blue color.

7
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

Fig. 8. Values of the chain growth probability, olefins content, and an excess of CH4 selectivity for the catalysts with different alkali dopants [7,24,45,86]. Fe means
a catalyst without alkali dopants. The same color means one specific catalysts from one article.

Fig. 9. Values of the chain growth probability, olefins content, and an “extra” increase of CH4 selectivity for the catalysts with different molar ratio Me/Fe (where Me
is K (circle) and Na (square)) [29,45,87,88]. The same color means one specific catalysts from one article.

shown that Mn can increase olefin content even without alkali dopants. data analysis is potassium. Due to its beneficial effects, it is in frequent
In the cases of Cu and Co, a reduced olefins production was observed use for CO2-FT catalysts.
with and without alkali dopants. For Co-containing catalysts, the high
value of the excess of methane parameter relates to their high hydro­
genation ability. In conclusion, the most promising dopant for 4.3. Descriptor importance analysis
improving activity and selectivity CO2-FT catalysts is K based on our
For analyzing the influence of reaction condition, composition, and

Fig. 10. The histogram of descriptors importance for chain growth probability calculated using random forest regression model. Blue bars are positive correlations,
red bars – negative correlations. * except for Li, s type of supports.

8
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

other parameters of catalysts (type of support, reduction temperature mediocre. Several reasons may be responsible for this. One of them is the
etc.) in common, descriptor-importance (or feather-importance) anal­ experimental errors in the literature data because the catalytic data were
ysis based on random forest regression model was used in this work. obtained in various scientific groups using different laboratory equip­
Fig. 10 presents the descriptor contributions of a random forest regres­ ment. Another reason is the likely existence of hidden parameters
sion model for the prediction of the chain growth probability. The influencing the catalytic performance. Catalysis is defined by many
analogue analyses for the other catalytic parameters can be found in the factors including structure of active site, their number, the presence of
Supporting information (Fig. S23-S25). The alkali content shows a sig­ additional sites for CO2/H2 adsorption, catalyst stability etc. These
nificant positive correlation with the chain growth probability (Fig. 10). factors can be determined not only by parameters used here as de­
This means that an increase of potassium amount in the catalyst is scriptors (catalyst composition, reaction condition, support type etc.)
accompanied by an increase of the α value. For reaction temperature, but also by other catalyst preparation parameters, nature of supports,
there is a negative correlation. purity of reagents etc. Because the information about catalyst in litera­
The descriptor-importance analysis indicates that the most impor­ ture is often incomplete, such parameters were not included into the
tant descriptor defining the chain growth probability is the alkaline models.
metal content. The next-important descriptors are the reaction temper­
ature and Fe content. The slight influence of the reaction temperature on 5. Conclusions
α was discussed before. Alkali dopants also are the most dominant
descriptor for olefin production (see Fig. S24, bottom). This key The analysis of hydrocarbons distribution showed that their selec­
importance of presence and content of alkali metal for the selectivity of tivity corresponds to the ASF distribution except for CH4 and C2H4. Our
CO2-FT catalyst proves the conclusions made before. In, the descriptor- proposed model based on ASF distribution modified by two additional
importance analysis confirms that alkaline metal content is more parameters was demonstrated to describe hydrocarbons of CO2-FT cat­
important than other descriptors such as a type of support, doping by alysts. The increase of methane selectivity reported for some catalysts
transitional metals, reduction temperature, supported or bulk catalysts may relate to direct hydrogenation CO2 to CH4 or splitting α-olefins. A
etc. decrease of ethylene selectivity may be connected with its high reac­
In the case of reaction rate, supports based on carbon (active carbon, tivity compared to other α-olefins. A lesson learned from this literature
carbon nanotubes etc.) were found to be the most promising one for data analysis, is that the chain growth probability should be calculated
achieving high rates of CO2 conversion (see Fig. S23, top). The next- from the hydrocarbon distribution C3+; C1 and C2 should not be included
important parameters are iron content, reaction temperature. Impor­ but reported separately.
tance of carbon supports can be proved by calculating mean values of The analysis of RWGS reaction carrying out in the CO2-FT process
CO2 conversion rate for different type of catalysts. So, the mean CO2 with different catalysts points to the existence of different pathways of
conversion rate for the catalysts supported on carbon is around CO2 hydrogenation. The direct CO2 converting to hydrocarbons via FT
1070 ml⋅h− 1⋅g− 1. In the case of bulk catalysts, the mean value is mechanism seems to play a role at least for some catalyst systems.
500 ml⋅h− 1⋅g− 1. For catalysts supported on Al2O3 the one is – Validation of this pathway by detailed kinetic modeling and further
370 ml⋅h− 1⋅g− 1. It allows one to conclude that carbon supports are experiments would be of great interest. It was also found that the
promising ones for CO2-FT catalysts. equilibrium of RWGS reaction is not established for most CO2-FT cata­
It is worth mentioning that transition metals slightly influence on lysts. Therefore, the rates of RWGS and FTS are similar in magnitude.
parameters of CO2-FT catalysts except for Co metal (see Fig. S24 and The data analysis allowed us to conclude that reaction parameters
S25). The last one relates to high hydrogenation activity and methane such as temperature, pressure, H2:CO2 ratio slightly influence the chain
formation that try to be avoided and reduced in the CO2-FT process. It growth probability, olefine content, and excess of CH4 selectivity. These
explains a large contribution of Co content for prediction of CH4 selec­ key performance parameters of CO2-FT catalysts are predominantly
tivity and excess CH4 selectivity. defined by catalyst composition, method of catalyst preparation, nature
The model´s performance for predicting the chain growth probability of active sites and pretreatment conditions.
is presented in Fig. 11. The performance of other models can see in the With respect to the catalyst composition, it was shown that doping
Supplementary information (see Fig. S24 and S25). Fig. 11 shows that catalysts by alkali metals like K or Na is the most effective measure to
the generalization of the obtained random forest regression model is improve catalyst performance. It increases the chain growth probability
as well as the olefins content and reduces the excess of CH4 selectivity.
Other factors that influence catalyst selectivity but to a smaller extend
are type of support, doping by transitional metals, reduction tempera­
ture, and reaction conditions.

Funding

Financial support from Federal Ministry of Education and Research

(BMBF) in project InnoSyn (FKZ: 03SF0616B) and Deutsche For­
schungsgemeinschaft (DFG, German Research Foundation) in project
NFDI4Cat (number 670389-NFDI 2/1) is gratefully acknowledged.

CRediT authorship contribution statement

Aleksandr Fedorov: Conceptualization, Methodology, Formal

analysis, Investigation, Software, Data curation, Writing – original draft,
Visualization. David Linke: Conceptualization, Supervision, Project
administration, Writing – review & editing.

Fig. 11. Performance of random forest regression model for prediction of the Declaration of Competing Interest
chain growth probability. The dependence between experimental values of
alpha and predicted values for test set on 10-fold cross validation. The authors declare that they have no known competing financial

9
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

interests or personal relationships that could have appeared to influence [25] A. Russkikh, G. Shterk, B.H. Al-Solami, B.A. Fadhel, A. Ramirez, J. Gascon, Turning
waste into value: potassium-promoted red mud as an effective catalyst for the
the work reported in this paper.
hydrogenation of CO2, ChemSusChem 13 (11) (2020) 2981–2987.
[26] B. Liang, H. Duan, T. Sun, J. Ma, X. Liu, J. Xu, X. Su, Y. Huang, T. Zhang, Effect of
Appendix A. Supporting information Na promoter on Fe-based catalyst for CO2 hydrogenation to alkenes, ACS Sustain.
Chem. Eng. 7 (1) (2019) 925–932.
[27] J. Liu, A. Zhang, X. Jiang, M. Liu, J. Zhu, C. Song, X. Guo, Direct transformation of
Supplementary data associated with this article can be found in the carbon dioxide to value-added hydrocarbons by physical mixtures of Fe5C2 and K-
online version at doi:10.1016/j.jcou.2022.102034. Modified Al2O3, Ind. Eng. Chem. Res. 57 (28) (2018) 9120–9126.
[28] B. Liang, T. Sun, J. Ma, H. Duan, L. Li, X. Yang, Y. Zhang, X. Su, Y. Huang,
T. Zhang, Mn decorated Na/Fe catalysts for CO2 hydrogenation to light olefins,
References Catal. Sci. Technol. 9 (2) (2019) 456–464.
[29] S.-R. Yan, K.-W. Jun, J.-S. Hong, M.-J. Choi, K.-W. Lee, Promotion effect of Fe–Cu
catalyst for the hydrogenation of CO2 and application to slurry reactor, Appl. Catal.
[1] M. Hermesmann, K. Grübel, L. Scherotzki, T.E. Müller, Promising pathways: The
A: Gen. 194–195 (2000) 63–70.
geographic and energetic potential of power-to-x technologies based on
[30] A. Aitbekova, E.D. Goodman, L. Wu, A. Boubnov, A.S. Hoffman, A. Genc, H. Cheng,
regeneratively obtained hydrogen, Renew. Sustain. Energy Rev. 138 (2021),
L. Casalena, S.R. Bare, M. Cargnello, Engineering of ruthenium–iron oxide colloidal
110644.
heterostructures: improved yields in CO2 hydrogenation to hydrocarbons, Angew.
[2] D.Y.C. Leung, G. Caramanna, M.M. Maroto-Valer, An overview of current status of
Chem. Int. Ed. 58 (48) (2019) 17451–17457.
carbon dioxide capture and storage technologies, Renew. Sustain. Energy Rev. 39
[31] S.-C. Lee, J.-H. Jang, B.-Y. Lee, J.-S. Kim, M. Kang, S.-B. Lee, M.-J. Choi, S.-
(2014) 426–443.
J. Choung, Promotion of hydrocarbon selectivity in CO2 hydrogenation by Ru
[3] M.D. Garba, M. Usman, S. Khan, F. Shehzad, A. Galadima, M.F. Ehsan, A.
component, J. Mol. Catal. A Chem. 210 (1) (2004) 131–141.
S. Ghanem, M. Humayun, CO2 towards fuels: a review of catalytic conversion of
[32] C. Zhang, C. Cao, Y. Zhang, X. Liu, J. Xu, M. Zhu, W. Tu, Y.-F. Han, Unraveling the
carbon dioxide to hydrocarbons, J. Environ. Chem. Eng. 9 (2) (2021), 104756.
role of zinc on bimetallic Fe5C2–ZnO catalysts for highly selective carbon dioxide
[4] R. Guil-López, N. Mota, J. Llorente, E. Millán, B. Pawelec, J.L.G. Fierro, R.
hydrogenation to high carbon α-Olefins, ACS Catal. 11 (4) (2021) 2121–2133.
M. Navarro, Methanol synthesis from CO2: a review of the latest developments in
[33] Y. Xu, P. Zhai, Y. Deng, J. Xie, X. Liu, S. Wang, D. Ma, Highly selective olefin
heterogeneous catalysis, Materials 12 (23) (2019) 3902.
production from CO2 hydrogenation on iron catalysts: a subtle synergy between
[5] C. Panzone, R. Philippe, A. Chappaz, P. Fongarland, A. Bengaouer, Power-to-Liquid
manganese and sodium additives, Angew. Chem. Int. Ed. 59 (48) (2020)
catalytic CO2 valorization into fuels and chemicals: focus on the Fischer-Tropsch
21736–21744.
route, J. CO2 Util. 38 (2020) 314–347.
[34] M. Albrecht, U. Rodemerck, M. Schneider, M. Bröring, D. Baabe, E.V. Kondratenko,
[6] M.V. Landau, N. Meiri, N. Utsis, R. Vidruk Nehemya, M. Herskowitz, Conversion of
Unexpectedly efficient CO2 hydrogenation to higher hydrocarbons over non-doped
CO2, CO, and H2 in CO2 hydrogenation to fungible liquid fuels on Fe-based
Fe2O3, Appl. Catal. B Environ. 204 (2017) 119–126.
catalysts, Ind. Eng. Chem. Res. 56 (45) (2017) 13334–13355.
[35] R.W. Dorner, D.R. Hardy, F.W. Williams, H.D. Willauer, K and Mn doped iron-
[7] J. Wang, Z. You, Q. Zhang, W. Deng, Y. Wang, Synthesis of lower olefins by
based CO2 hydrogenation catalysts: detection of KAlH4 as part of the catalyst’s
hydrogenation of carbon dioxide over supported iron catalysts, Catal. Today 215
active phase, Appl. Catal. A: Gen. 373 (1) (2010) 112–121.
(2013) 186–193.
[36] R.W. Dorner, D.R. Hardy, F.W. Williams, H.D. Willauer, C2-C5+ olefin production
[8] T. Numpilai, N. Chanlek, Y. Poo-Arporn, C.K. Cheng, N. Siri-Nguan, T. Sornchamni,
from CO2 hydrogenation using ceria modified Fe/Mn/K catalysts, Catal. Commun.
M. Chareonpanich, P. Kongkachuichay, N. Yigit, G. Rupprechter, J. Limtrakul,
15 (1) (2011) 88–92.
T. Witoon, Tuning interactions of surface-adsorbed species over Fe− Co/K− Al2O3
[37] T. Toyao, Z. Maeno, S. Takakusagi, T. Kamachi, I. Takigawa, K.-i Shimizu, Machine
catalyst by different K contents: selective CO2 hydrogenation to light olefins,
learning for catalysis informatics: recent applications and prospects, ACS Catal. 10
ChemCatChem 12 (12) (2020) 3306–3320.
(3) (2020) 2260–2297.
[9] B. Yao, T. Xiao, O.A. Makgae, X. Jie, S. Gonzalez-Cortes, S. Guan, A.I. Kirkland, J.
[38] V. Tshitoyan, J. Dagdelen, L. Weston, A. Dunn, Z. Rong, O. Kononova, K.A. Persson,
R. Dilworth, H.A. Al-Megren, S.M. Alshihri, P.J. Dobson, G.P. Owen, J.M. Thomas,
G. Ceder, A. Jain, Unsupervised word embeddings capture latent knowledge from
P.P. Edwards, Transforming carbon dioxide into jet fuel using an organic
materials science literature, Nature 571 (7763) (2019) 95–98.
combustion-synthesized Fe-Mn-K catalyst, Nat. Commun. 11 (1) (2020) 6395.
[39] A. Smith, A. Keane, J.A. Dumesic, G.W. Huber, V.M. Zavala, A machine learning
[10] U. Rodemerck, M. Holeňa, E. Wagner, Q. Smejkal, A. Barkschat, M. Baerns,
framework for the analysis and prediction of catalytic activity from experimental
Catalyst development for CO2 hydrogenation to fuels, ChemCatChem 5 (7) (2013)
data, Appl. Catal. B Environ. 263 (2020), 118257.
1948–1955.
[40] K. Suzuki, T. Toyao, Z. Maeno, S. Takakusagi, K.-i Shimizu, I. Takigawa, Statistical
[11] A. Ramirez, A. Dutta Chowdhury, A. Dokania, P. Cnudde, M. Caglayan, I. Yarulina,
analysis and discovery of heterogeneous catalysts based on machine learning from
E. Abou-Hamad, L. Gevers, S. Ould-Chikh, K. De Wispelaere, V. van Speybroeck,
diverse published data, ChemCatChem 11 (18) (2019) 4537–4547.
J. Gascon, Effect of zeolite topology and reactor configuration on the direct
[41] U. Zavyalova, M. Holena, R. Schlögl, M. Baerns, Statistical analysis of past catalytic
conversion of CO2 to light olefins and aromatics, ACS Catal. 9 (7) (2019)
data on oxidative methane coupling for new insights into the composition of high-
6320–6334.
performance catalysts, ChemCatChem 3 (12) (2011) 1935–1947.
[12] J. Wei, R. Yao, Q. Ge, Z. Wen, X. Ji, C. Fang, J. Zhang, H. Xu, J. Sun, Catalytic
[42] E.V. Kondratenko, M. Schlüter, M. Baerns, D. Linke, M. Holena, Developing
hydrogenation of CO2 to isoparaffins over Fe-Based multifunctional catalysts, ACS
catalytic materials for the oxidative coupling of methane through statistical
Catal. 8 (11) (2018) 9958–9967.
analysis of literature data, Catal. Sci. Technol. 5 (3) (2015) 1668–1677.
[13] P. Gao, S. Li, X. Bu, S. Dang, Z. Liu, H. Wang, L. Zhong, M. Qiu, C. Yang, J. Cai,
[43] Q. Yang, A. Skrypnik, A. Matvienko, H. Lund, M. Holena, E.V. Kondratenko,
W. Wei, Y. Sun, Direct conversion of CO2 into liquid fuels with high selectivity over
Revealing property-performance relationships for efficient CO2 hydrogenation to
a bifunctional catalyst, Nat. Chem. 9 (10) (2017) 1019–1024.
higher hydrocarbons over Fe-based catalysts: Statistical analysis of literature data
[14] T. Riedel, G. Schaub, K.-W. Jun, K.-W. Lee, Kinetics of CO2 hydrogenation on a K-
and its experimental validation, Appl. Catal. B Environ. 282 (2021), 119554.
promoted Fe catalyst, Ind. Eng. Chem. Res. 40 (5) (2001) 1355–1363.
[44] T. Herranz, S. Rojas, F.J. Pérez-Alonso, M. Ojeda, P. Terreros, J.L.G. Fierro,
[15] R.E. Owen, D. Mattia, P. Plucinski, M.D. Jones, Kinetics of CO2 hydrogenation to
Hydrogenation of carbon oxides over promoted Fe-Mn catalysts prepared by the
hydrocarbons over iron–silica catalysts, ChemPhysChem 18 (22) (2017)
microemulsion methodology, Appl. Catal. A Gen. 311 (2006) 66–75.
3211–3218.
[45] A. Ramirez, L. Gevers, A. Bavykina, S. Ould-Chikh, J. Gascon, Metal organic
[16] R. Satthawong, N. Koizumi, C. Song, P. Prasassarakich, Light olefin synthesis from
framework-derived iron catalysts for the direct hydrogenation of CO2 to short
CO2 hydrogenation over K-promoted Fe–Co bimetallic catalysts, Catal. Today 251
chain Olefins, ACS Catal. 8 (10) (2018) 9174–9182.
(2015) 34–40.
[46] M. Owens, The definitive guide to SQLite, Apress.
[17] J. Zhang, S. Lu, X. Su, S. Fan, Q. Ma, T. Zhao, Selective formation of light olefins
[47] R.D. Hipp, SQLite.
from CO2 hydrogenation over Fe–Zn–K catalysts, J. CO2 Util. 12 (2015) 95–100.
[48] G. Van Rossum, F.L. Drake, Python 3 Reference Manual, CreateSpace.
[18] G.V. Schulz, Über die Beziehung zwischen Reaktionsgeschwindigkeit und
[49] M. Bayer, SQLAlchemy, in: A. Brown, G. Wilson (Eds.), The Architecture of Open
Zusammensetzung des Reaktionsproduktes bei Makropolymerisationsvorgängen,
Source Applications Volume II: Structure, Sca le, and a Few More Fearless Hacks,
Z. für Phys. Chem. 30B (1) (1935) 379–398.
aosabook.org, 2012.
[19] P.J. Flory, Molecular size distribution in linear condensation polymers1, J. Am.
[50] P. Virtanen, R. Gommers, T.E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau,
Chem. Soc. 58 (10) (1936) 1877–1885.
E. Burovski, P. Peterson, W. Weckesser, J. Bright, S.J. van der Walt, M. Brett,
[20] R.B. Anderson, R.A. Friedel, H.H. Storch, Fischer-Tropsch reaction mechanism
J. Wilson, K.J. Millman, N. Mayorov, A.R.J. Nelson, E. Jones, R. Kern, E. Larson, C.
involving stepwise growth of carbon Chain, J. Chem. Phys. 19 (3) (1951) 313–319.
J. Carey, İ. Polat, Y. Feng, E.W. Moore, J. VanderPlas, D. Laxalde, J. Perktold,
[21] G. Henrici-Olivé, S. Olivé, The Fischer-Tropsch synthesis: molecular weight
R. Cimrman, I. Henriksen, E.A. Quintero, C.R. Harris, A.M. Archibald, A.H. Ribeiro,
distribution of primary products and reaction mechanism, Angew. Chem. Int. Ed.
F. Pedregosa, P. van Mulbregt, A. Vijaykumar, A.P. Bardelli, A. Rothberg,
Engl. 15 (3) (1976) 136–141.
A. Hilboll, A. Kloeckner, A. Scopatz, A. Lee, A. Rokem, C.N. Woods, C. Fulton,
[22] E.S. Lox, G.F. Froment, Kinetics of the Fischer-Tropsch reaction on a precipitated
C. Masson, C. Häggström, C. Fitzgerald, D.A. Nicholson, D.R. Hagen, D.
promoted iron catalyst. 2. Kinetic modeling, Ind. Eng. Chem. Res. 32 (1) (1993)
V. Pasechnik, E. Olivetti, E. Martin, E. Wieser, F. Silva, F. Lenders, F. Wilhelm,
71–82.
G. Young, G.A. Price, G.-L. Ingold, G.E. Allen, G.R. Lee, H. Audren, I. Probst, J.
[23] F.A.N. Fernandes, Polymerization kinetics of Fischer-Tropsch reaction on iron
P. Dietrich, J. Silterra, J.T. Webber, J. Slavič, J. Nothman, J. Buchner, J. Kulick, J.
based catalysts and product grade optimization, Chem. Eng. Technol. 28 (8) (2005)
L. Schönberger, J.V. de Miranda Cardoso, J. Reimer, J. Harrington, J.L.
930–938.
C. Rodríguez, J. Nunez-Iglesias, J. Kuczynski, K. Tritz, M. Thoma, M. Newville,
[24] Z. You, W. Deng, Q. Zhang, Y. Wang, Hydrogenation of carbon dioxide to light
M. Kümmerer, M. Bolingbroke, M. Tartre, M. Pak, N.J. Smith, N. Nowaczyk,
olefins over non-supported iron catalyst, Chin. J. Catal. 34 (5) (2013) 956–963.

10
A. Fedorov and D. Linke Journal of CO2 Utilization 61 (2022) 102034

N. Shebanov, O. Pavlyk, P.A. Brodtkorb, P. Lee, R.T. McGibbon, R. Feldbauer, [71] S. Wang, T. Wu, J. Lin, Y. Ji, S. Yan, Y. Pei, S. Xie, B. Zong, M. Qiao, Iron–potassium
S. Lewis, S. Tygier, S. Sievert, S. Vigna, S. Peterson, S. More, T. Pudlik, T. Oshima, on single-walled carbon nanotubes as efficient catalyst for CO2 hydrogenation to
T.J. Pingel, T.P. Robitaille, T. Spura, T.R. Jones, T. Cera, T. Leslie, T. Zito, heavy olefins, ACS Catal. 10 (11) (2020) 6389–6401.
T. Krauss, U. Upadhyay, Y.O. Halchenko, Y. Vázquez-Baeza, C. SciPy, SciPy 1.0: [72] Z. He, M. Cui, Q. Qian, J. Zhang, H. Liu, B. Han, Synthesis of liquid fuel via direct
fundamental algorithms for scientific computing in Python, Nat. Methods 17 (3) hydrogenation of CO2, Proc. Natl. Acad. Sci. 116 (26) (2019) 12654–12659.
(2020) 261–272. [73] X. Su, J. Zhang, S. Fan, Q. Ma, T.-S. Zhao, Effect of preparation of Fe–Zr–K catalyst
[51] C.R. Harris, K.J. Millman, S.J. van der Walt, R. Gommers, P. Virtanen, on the product distribution of CO2 hydrogenation, RSC Adv. 5 (98) (2015)
D. Cournapeau, E. Wieser, J. Taylor, S. Berg, N.J. Smith, R. Kern, M. Picus, 80196–80202.
S. Hoyer, M.H. van Kerkwijk, M. Brett, A. Haldane, J.F. del Río, M. Wiebe, [74] R. Satthawong, N. Koizumi, C. Song, P. Prasassarakich, Bimetallic Fe–Co catalysts
P. Peterson, P. Gérard-Marchant, K. Sheppard, T. Reddy, W. Weckesser, H. Abbasi, for CO2 hydrogenation to higher hydrocarbons, J. CO2 Util. 3–4 (2013) 102–106.
C. Gohlke, T.E. Oliphant, Array programming with NumPy, Nature 585 (7825) [75] N. Boreriboon, X. Jiang, C. Song, P. Prasassarakich, Fe-based bimetallic catalysts
(2020) 357–362. supported on TiO2 for selective CO2 hydrogenation to hydrocarbons, J. CO2 Util.
[52] W. McKinney, others, Data structures for statistical computing in python, Austin, 25 (2018) 330–337.
TX, pp. 51–56. [76] L. Guo, J. Sun, X. Ji, J. Wei, Z. Wen, R. Yao, H. Xu, Q. Ge, Directly converting
[53] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. carbon dioxide to linear α-olefins on bio-promoted catalysts, Commun. Chem. 1 (1)
Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. (2018) 11.
Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: Machine Learning [77] M.J. Bradley, R. Ananth, H.D. Willauer, J.W. Baldwin, D.R. Hardy, F.W. Williams,
in Python, Journal of Machine Learning Research 12 2825–2830. The effect of copper addition on the activity and stability of iron-based CO₂
[54] P.T. Inc, Collaborative data science, 〈https://plot.ly〉. hydrogenation catalysts, Molecules 22 (9) (2017) 1579.
[55] C.G. Visconti, M. Martinelli, L. Falbo, A. Infantes-Molina, L. Lietti, P. Forzatti, [78] T. Riedel, H. Schulz, G. Schaub, K.-W. Jun, J.-S. Hwang, K.-W. Lee,
G. Iaquaniello, E. Palo, B. Picutti, F. Brignoli, CO2 hydrogenation to lower olefins Fischer–Tropsch on Iron with H2/CO and H2/CO2 as synthesis gases: the episodes
on a high surface area K-promoted bulk Fe-catalyst, Appl. Catal. B Environ. 200 of formation of the Fischer–Tropsch regime and construction of the catalyst, Top.
(2017) 530–542. Catal. 26 (1) (2003) 41–54.
[56] S. Krishnamoorthy, A. Li, E. Iglesia, Pathways for CO2 formation and conversion [79] S. Hu, M. Liu, F. Ding, C. Song, G. Zhang, X. Guo, Hydrothermally stable MOFs for
during fischer–tropsch synthesis on iron-based catalysts, Catal. Lett. 80 (1) (2002) CO2 hydrogenation over iron-based catalyst to light olefins, J. CO2 Util. 15 (2016)
77–86. 89–95.
[57] J.M. Moe, Design of water-gas shift reactors, Chem. Eng. Prog. 58 (1962) 3. [80] F. Jiang, B. Liu, S. Geng, Y. Xu, X. Liu, Hydrogenation of CO2 into hydrocarbons:
[58] J. Nocedal, S. Wright, Numerical Optimization, Springer, New York, 2006. enhanced catalytic activity over Fe-based Fischer–Tropsch catalysts, Catal. Sci.
[59] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning, Springer Technol. 8 (16) (2018) 4097–4107.
New York. [81] J. Jiang, C. Wen, Z. Tian, Y. Wang, Y. Zhai, L. Chen, Y. Li, Q. Liu, C. Wang, L. Ma,
[60] V. Svetnik, A. Liaw, C. Tong, J.C. Culberson, R.P. Sheridan, B.P. Feuston, Random Manganese-promoted Fe3O4 microsphere for efficient conversion of CO2 to light
forest: a classification and regression tool for compound classification and QSAR olefins, Ind. Eng. Chem. Res. 59 (5) (2020) 2155–2162.
modeling, J. Chem. Inf. Comput. Sci. 43 (6) (2003) 1947–1958. [82] J.S. Hwang, K.-W. Jun, K.-W. Lee, Deactivation and regeneration of Fe-K/alumina
[61] R.A. Fisher, On the interpretation of χ2 from contingency tables, and the catalyst in CO2 hydrogenation, Appl. Catal. A Gen. 208 (1) (2001) 217–222.
calculation of P, J. R. Stat. Soc. 85 (1) (1922) 87–94. [83] W.D. Shafer, G. Jacobs, U.M. Graham, H.H. Hamdeh, B.H. Davis, Increased CO2
[62] A.S. Skrypnik, Q. Yang, A.A. Matvienko, V.Y. Bychkov, Y.P. Tulenin, H. Lund, S. hydrogenation to liquid products using promoted iron catalysts, J. Catal. 369
A. Petrov, R. Kraehnert, A. Arinchtein, J. Weiss, A. Brueckner, E.V. Kondratenko, (2019) 239–248.
Understanding reaction-induced restructuring of well-defined FexOyCz [84] Y. Zhang, C. Cao, C. Zhang, Z. Zhang, X. Liu, Z. Yang, M. Zhu, B. Meng, J. Xu, Y.-
compositions and its effect on CO2 hydrogenation, Appl. Catal. B Environ. 291 F. Han, The study of structure-performance relationship of iron catalyst during a
(2021), 120121. full life cycle for CO2 hydrogenation, J. Catal. 378 (2019) 51–62.
[63] J. Zhu, G. Zhang, W. Li, X. Zhang, F. Ding, C. Song, X. Guo, Deconvolution of the [85] J.-S. Hong, J.S. Hwang, K.-W. Jun, J.C. Sur, K.-W. Lee, Deactivation study on a
particle size effect on CO2 hydrogenation over iron-based catalysts, ACS Catal. 10 coprecipitated Fe-Cu-K-Al catalyst in CO2 hydrogenation, Appl. Catal. A: Gen. 218
(13) (2020) 7424–7433. (1) (2001) 53–59.
[64] J. Liu, G. Zhang, X. Jiang, J. Wang, C. Song, X. Guo, Insight into the role of Fe5C2 in [86] T. Riedel, M. Claeys, H. Schulz, G. Schaub, S.-S. Nam, K.-W. Jun, M.-J. Choi,
CO2 catalytic hydrogenation to hydrocarbons, Catal. Today 371 (2021) 162–170. G. Kishan, K.-W. Lee, Comparative study of Fischer–Tropsch synthesis with H2/CO
[65] S. Daniel, NIST Standard Reference Simulation Website - SRD 173, National and H2/CO2 syngas using Fe- and Co-based catalysts, Appl. Catal. A Gen. 186 (1)
Institute of Standards. (1999) 201–213.
[66] J. Hunt, A. Ferrari, A. Lita, M. Crosswhite, B. Ashley, A.E. Stiegman, Microwave- [87] B. Liu, S. Geng, J. Zheng, X. Jia, F. Jiang, X. Liu, Unravelling the new roles of Na
specific enhancement of the carbon–carbon dioxide (Boudouard) reaction, J. Phys. and Mn promoter in CO2 hydrogenation over Fe3O4-based catalysts for enhanced
Chem. C 117 (51) (2013) 26871–26880. selectivity to light α-Olefins, ChemCatChem 10 (20) (2018) 4718–4732.
[67] X. Cui, P. Gao, S. Li, C. Yang, Z. Liu, H. Wang, L. Zhong, Y. Sun, Selective [88] J. Wei, J. Sun, Z. Wen, C. Fang, Q. Ge, H. Xu, New insights into the effect of sodium
production of aromatics directly from carbon dioxide hydrogenation, ACS Catal. 9 on Fe3O4-based nanocatalysts for CO2 hydrogenation to light olefins, Catal. Sci.
(5) (2019) 3866–3876. Technol. 6 (13) (2016) 4786–4793.
[68] S. Geng, F. Jiang, Y. Xu, X. Liu, Iron-based Fischer–Tropsch synthesis for the [89] X. Liu, C. Zhang, P. Tian, M. Xu, C. Cao, Z. Yang, M. Zhu, J. Xu, Revealing the effect
efficient conversion of carbon dioxide into isoparaffins, ChemCatChem 8 (7) of sodium on iron-based catalysts for CO2 hydrogenation: insights from calculation
(2016) 1303–1307. and experiment, J. Phys. Chem. C 125 (14) (2021) 7637–7646.
[69] R.E. Owen, P. Plucinski, D. Mattia, L. Torrente-Murciano, V.P. Ting, M.D. Jones, [90] N. Fischer, R. Henkel, B. Hettel, M. Iglesias, G. Schaub, M. Claeys, Hydrocarbons
Effect of support of Co-Na-Mo catalysts on the direct conversion of CO2 to via CO2 hydrogenation over iron catalysts: the effect of potassium on structure and
hydrocarbons, J. CO2 Util. 16 (2016) 97–103. performance, Catal. Lett. 146 (2) (2016) 509–517.
[70] D.L. Williamson, C. Herdes, L. Torrente-Murciano, M.D. Jones, D. Mattia, N-Doped
Fe@CNT for combined RWGS/FT CO2 hydrogenation, ACS Sustain. Chem. Eng. 7
(7) (2019) 7395–7402.

11