International Journal of Hydrogen Energy 59 (2024) 1023–1041

Contents lists available at ScienceDirect

International Journal of Hydrogen Energy

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

Parametric analysis of CO2 hydrogenation via Fischer-Tropsch synthesis: A

review based on machine learning for quantitative assessment
Jing Hu a, 1, Yixao Wang b, 1, Xiyue Zhang c, Yunshan Wang d, Gang Yang d, **, Lufang Shi e,
Yong Sun f, g, *
a
Stanford School of Engineering & Doerr School of Sustainability, Stanford University, California, 94305, USA
b
UCL Institute for Materials Discovery, University College London (UCL), London, WC1H 0AJ, UK
c
Department of Civil and Environmental Engineering, Imperial College London, London, SW7 3LE, UK
d
National Engineering Laboratory of Cleaner Hydrometallurgical Production Technology, Institute of Process Engineering, Chinese Academy of Sciences, Beijing, 100190,
China
e
Each Energy Australia, James Ruse Drive, NSW, 2116, Australia
f
School of Engineering, Edith Cowan University, 270 Joondalup Drive, Joondalup, WA, 6027, Australia
g
Key Laboratory of Carbonaceous Wastes Processing and Process Intensification of Zhejiang Province, University of Nottingham Ningbo, 315100, China

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

Handling Editor: Ramazan Solmaz This review focuses on the parametric impacts upon conversion and selectivity during CO2 hydrogenation via
Fischer-Tropsch (FT) synthesis using iron-based catalyst to provide quantitative evaluation. Using all collected
Keywords: data from reported literatures as training dataset via artificial neural networks (ANNs) in TensorFlow, three
Artificial neural networks categorized parameters (namely: operational, catalyst informatic and mass transfer) were deployed to assess
CO2 hydrogenation
their impacts upon conversions (CO2) and selectivity. The lump kinetic power expressions among literature
Fischer-Tropsch synthesis
reports were compared, and the best fit model is the one that was proposed by this work without arbitrarily
Parametric analysis
Review assuming power values of individual partial pressure (CO and H2). More than five sets of binary parameters were
systematically investigated to find out corresponding evolving patterns in conversion and selectivity. Aided by
machine learning, tailoring product distributions based on specific selectivity or conversion for optimization
purpose is practically achievable by deploying the predictions generated from ANNs in this work.

1. Introduction green hydrogen [9–11]. As a mature technology, the fuels could be

practically and economically transformed to e-gas (methanation) or
Although the idea for direct electrification has been widely accepted e-liquids (Fischer-Tropsch synthesis) for long-range and air transport
by the general-public, going electric in some sectors might not be the [12–15]. Additionally, both bulk and fine chemical manufacturing could
best option based on current conditions [1–3]. The necessity of using be realized by viable Fischer-Tropsch (FT) synthesis technology to
fossil fuels is highly likely to remain during this transitional period, generate various chemical building blocks [16–18]. Thus, the research
approximately 5–10 years ahead [4,5]. Under such circumstances, no and development (R&D) in this direction will surely become meaningful
wonder many conventional oil companies (i.e., Chevron, Occidental during this special energy transitional period. While many technical
Petroleum, Woodside etc.) could still confidently propose the advantages show the rosy prospects, please note particularly that the
middle-term plan of integrating their oil & gas extraction business with catalytic CO2 hydrogenation is still quite challenging due to its intrinsic
carbon capture and storage (CCS) technology to further operate their complexity during the reaction, and it still requires arduous R&D efforts
fossil fuel business arms [6–8]. This surely opens the door for CO2 uti­ [19]. Several routes to convert CO2 into valuable chemicals have been
lizations such as through the route of chemical valorization of CO2 with studied in recent years, ranging from traditional heterogeneous catalysis

* Corresponding author. School of Engineering, Edith Cowan University, University of Nottingham Ningbo China, 270 Joondalup Drive, Joondalup, WA, 6027,
Australia.
** Corresponding author. Institute of Process Engineering, Chinese Academy of Sciences, China.
E-mail addresses: [email protected] (G. Yang), [email protected] (Y. Sun).
1
These two authors contribute equally to this work.

https://doi.org/10.1016/j.ijhydene.2024.02.055
Received 1 December 2023; Received in revised form 14 January 2024; Accepted 5 February 2024
Available online 14 February 2024
0360-3199/© 2024 The Authors. Published by Elsevier Ltd on behalf of Hydrogen Energy Publications LLC. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

to electrocatalysis [20–22], homogeneous catalysis [23,24], and pho­ correlate the objective outputs by the considering the catalytic features
tocatalysis [25–27]. The thermal catalytic CO2 hydrogenation via FT (cheminformatic, material informatic), micro-kinetic and macro-kinetic,
synthesis has attracted many attentions due to its well-studied topic and operational conditions has seldomly been conducted. Thus, the
[28–30], relative higher maturity (solid successful commercialization relatively simple system with less complex intrinsic mechanisms tends to
examples of FT synthesis by Shell and Sasol in last several decades), and be rigorously investigated instead. While for the complex system such as
higher carbon efficiency [31–34]. FT synthesis and CO2 hydrogenation via FT synthesis, big gaps persis­
To better understand a system, the top-down rules, models and tently exist; 2) CO2 hydrogenation via FT synthesis is essentially
simulations are often ideally chosen. Yet, the complexity of a system different to the FT synthesis, because the reactions of CO2 hydrogenation
often makes this venerable attempt as a formidable challenge. Alterna­ (CO formation) and FT synthesis (hydrocarbons formation-olefin, par­
tively, with availability of data and advancement of artificial intelli­ affins, and oxygenates etc.) have been found to occur on different cat­
gence (AI), it offers another route to approach the problems from alytic sites. This complexity of CO2 hydrogenation using iron-based
bottom-up instead, which is to measure or collect lots of data first, catalyst adds more uncertainties. However, please note particularly
and use algorithms to come up with rules, patterns, equations, and sci­ that the purpose of this work is not to challenge this situation, instead,
entific understanding later. To better utilize this powerful tool, some our goal is to sort out the potential data framework with minimum
pioneering works have been done by the scholars. Andrew et al. pro­ datum dimension based on complicated system that are investigated, so
posed a data-information-knowledge hierarchy system using previously that more meaningful predictions and efficient data handling method­
published works, which contains a myriad kind of catalysts for different ologies together with minimum computational power could be identi­
heterogeneous reactions [35], Takahashi et al., reviewed the catalyst fied. Therefore, in this work, we attempted to use the selected three
informatics in the aspects of catalyst data, catalyst design via data sci­ categorized parameters, namely operational parameters, catalyst
ence, and catalyst platform, thus the establishment of a new paradigm of intrinsic parameters and mass transfer features of the system, to inves­
correlating this catalytic informatics for better catalyst design is dis­ tigate the responding evolution patterns of conversion (CO2) and
cussed [36]. In recent works, Liu et al. pinpointed the in-situ carbon selectivity (CO, CH4, olefin SOi and paraffin SPi, where i changes from 2
vacant mediated CO activation mechanism assisted by ML techniques to 10). To further reduce the dimensional complexity, we followed a
during FT synthesis [37]. Motaev et al. attempted to use ML to generate two-step analytical approach. First, the widely used statistical analysis,
a model in hope to link 16 property parameters with 6 objective pa­ namely the analysis of variables (ANOVA), was deployed to prelimi­
rameters during FT synthesis using cobalt catalyst [38]. Because of its narily find out some statistically significant factors (inputs) that could
intrinsic complexity of FT synthesis, more efforts need to be made in potentially influence the conversion and selectivity (outputs). Then, the
correlating features from catalyst (combining the cheminformatic and supervised ML algorithm in TensorFlow was used to find out predictions
material informatic) with other useful operational parameters such as with acceptable leeway so that the responding patterns of conversion
conventional generic description of the process (temperature, pressure and selectivity towards the parametric variations could be captured and
etc.) and mass transfer parameters (i.e., gas hour space velocity-GHSV, analysed. The investigation and assessment of the statistical significance
unitless numbers etc.). Thus, more meaningful, and useful predictions of different conditions upon the catalytic CO2 hydrogenation via FT
could be made with the advancement of ML technology. In this work, we synthesis using iron-based catalyst, to the best of the authors’ knowl­
attempted to explore from bottom-up route, hoping to find out some edge, has not been reported before.
patterns that were produced by the variations of selected inputs during
CO2 hydrogenation via FT synthesis. The choice of how many di­ 2. Reported data collection and data training
mensions that needed to be considered is hard to take when it comes to
assess catalytic CO2 hydrogenation via FT synthesis [15,39]. The factors 2.1. Overview of data
including potential influential impacts, and availability of references
etc., all possibly determine the merits of this trial. Regarding CO2 hy­ All collected data was normalized for the convenience of ANNs
drogenation via FT synthesis using iron-based catalyst, we carefully handling. The data matrix includes 11 inputs (namely stoichiometric
choose those sets of paramount factors as inputs, which are often ratio of Fe, binding energy, loading of Fe weight percentage, particle size
considered as indispensable descriptions for CO2 hydrogenation via FT of catalyst, temperature, GHSV, pressure, TOS, catalytic metal size,
synthesis studies. Broadly speaking, three main categorized inputs were Mears’ criteria, and Thiele modulus) and corresponding 21 outputs (CO2
chosen in this work, they are operational parameters (i.e., pressure, conversion, CO selectivity, CH4 selectivity, olefin selectivity SOi, and
temperature, gas hour space velocity-GHSV, time on stream-TOS), paraffin selectivity SPi, where i changes from 2 to 10). The visualization
catalyst intrinsic feature (i.e., bulk particle size, active metal nano­ of the collected reference data is shown in Fig. 1 as a heatmap, and the
particle size, loading ratio, binding energies estimated from density summary of some selected inputs and outputs are shown in Table 1,
functional theory-DFT approaches etc.), and mass transfer features (i.e., respectively. These data will be used for data training.
Mears’ criteria and Thiele modulus). For outputs, apart from the con­ It is not surprising to find out that the data distribution (both for
version (CO2), the CO selectivity and syncrude composition (olefin and inputs and outputs) among all collected references are not ideally
paraffins with carbon number from C1 to C10) were also considered, distributed, the great challenges i.e., inconsistent standards, experi­
while oxygenates, branched olefins and paraffins were ignored due to mental data collections uncertainties, non-universal referenced baseline
their relatively smaller quantities and scarcity of reported data from conditions, different configured reaction system, and different analytical
literature reports. methods including the calibration methods of using reference gases,
In this work, unlike conventional review papers, which normally might all contribute to this complexity. Yet, the purpose of this work is
summarize the achievement of other scholars in the relevant field in a not to challenge the status quo of the reported data. Instead, our goal is
narrative description, we adopted aforementioned bottom-up data to use bottom-up approach by feeding training data into ANNs to find
driven approach by feeding all collected relevant data into the ANNs to some meaningful patterns for CO2 hydrogenation via FT synthesis using
find out some patterns, rules etc., that have not been reported from iron-based catalyst. The detailed summary of all collected references
previously reviewed perspectives. A couple of reasons might attribute to used for data training are summarized in Supplementary Table S1 (all
the status quo of scarcity of this type of work: 1) to perform ML, one of collected variables of inputs) and Table S2 (all corresponding responses
the foremost tasks is to tailor and identify the feasible and meaningful for outputs), respectively.
data dimensions that could potentially generate useful predictions.
Although the field of catalytic reactions is a conventionally well-
explored domain, how to work out the useful and efficient inputs to

1024
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Fig. 1. Heatmap of dataset from all collected references for data training (both inputs and their corresponding outputs). Prior to be fed into machine learning al­
gorithm, all data need to be normalized and the dimension of minimum and maximum is determined by reported values from all collected references, where XCO2
represent conversion of CO2, SCO refers to CO selectivity, SCH4 refer to methane selectivity, SOi (i refers to carbon varying from 2 to 10) refer to olefin with different
carbon numbers, SPi (i varies from 2 to 10) refer to paraffin with different carbon numbers. Please note that values between 0 and 1 in scale bar refer to the
normalization scale, of which 0 refers to the minimum reported value, and 1 refers to the maximum reported value.

2.2. Data training analysis reported and predicted values were also calculated. The MARR results
suggest that the majority of the reported data point falls into the
During data training, the cross-validation approach was deployed to acceptable leeway of ±25%, with some exceptional higher MARR un­
maximize the dimension of the collected reported dataset. The corre­ certainties appearing at either low or high dimension. Overall, the
sponding mean square error (MSE) and mean average relative residual MARR result indicates a reliability of using this established ANNs in
(MARR) were deployed to analyse the quality between the training data TensorFlow for modeling.
and the model predictions. The expressions are the following:
2
3. Theorical background and methodology
Nexp
1 ∑ )
MSE% = (dexp − dical × 100% (1)
Nexp j=1 i 3.1. Statistically significant analysis

1 ∑
⃒
Nexp (⃒ exp
⃒)
di − dical ⃒ The graphical column and corresponding ANOVA with p-value are
MARR% = × 100% (2) summarized in Fig. 3. The graphical column shows 11 inputs and 21
Nexp j=1 diexp
outputs in data matrix. The inputs are: stoichiometric ratio, binding
energies (estimated using DFT approach), loading percentage, bulk
Where Nexp is the number of reported experimental data (− ), dexp
i and dcal
i particle size, temperature, gas-hour space velocity (GHSV), pressure,
are data reported from references and predictions calculated from ANNs time-on-stream (TOS), active catalytic metal size, Mears’ criteria, and
in TensorFlow, respectively. In ANNs, MSE might potentially refer to the Thiele modulus, respectively. The analysis of variance (ANOVA) was
empirical risk (the average loss on an observed dataset), as an estimate used to preliminarily assess the impacts on the outputs caused by the
of the true MSE. The MSE is a measure of the quality of an estimator. As it variation of all inputs. The p-value of all 21 outputs (which is the ratio of
is derived from the square of Euclidean distance, it is always a positive the mean squares treatment to the mean squares error) were calculated
value that decreases as the error approaches zero. The MSE and MARR to determine if it is statistically significant. Among all investigated 21
analysis of collected dataset are summarized in Fig. 2. The MSE from the outputs, there are 5 outputs (namely XCO2 , SCO , SCH4 , SO2 and SP2 ) being
majority of collected dataset presents less than 25%, indicating a rela­ found to be statistically significant to the variations of inputs (with p-
tively acceptable training outcome. value<0.01). Please be kindly advised that the outputs with p-value
Because MARR possesses the feature of capturing larger uncertainties larger than 0.01 does not necessarily mean that their variations are not
at a relative lower actual value, the corresponding MARR values of all

1025
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Table 1
Summary of collected references of some selected inputs (Temp, Pressure, Mears, Thiele) and outputs (Xco2, SCO, SCH4, SO2, SP2) for data training. Please note
particularly that some parameters being not directly reported from literatures need to be estimated based on their corresponding reported experimental conditions.
Please allow roughly ±25% uncertainties for extracting data sets from literature reports.
Temp/◦ C Pressure/bar Mears/- Thiele/- XCO2/% SCO/% SCH4/% SO2/% SP2/% Ref

320 10 0.34 0.18 29.20 20.20 21.00 10.00 5.00 [40]

320 10 0.42 0.19 32.40 21.40 13.00 8.00 3.00 [40]
330 15 3.58 0.22 40.10 9.50 14.50 6.00 3.00 [41]
330 15 4.22 0.23 43.50 9.20 15.00 5.00 3.00 [41]
340 20 2.00 0.19 35.00 18.00 10.70 6.70 1.30 [42]
340 20 0.15 0.14 42.00 15.00 20.00 15.33 2.73 [43]
320 5 0.68 0.29 54.00 6.29 31.13 16.75 2.15 [44]
320 5 0.50 0.26 46.14 9.57 32.57 15.14 2.35 [44]
300 11 0.70 0.14 31.50 27.00 29.00 0.05 0.07 [45]
300 30 0.14 0.14 40.00 14.20 17.20 15.57 2.83 [46]
350 20 1.96 0.27 44.70 9.40 22.00 15.40 2.37 [47]
320 30 0.16 0.14 40.00 10.00 7.00 6.67 1.11 [48]
300 8 0.64 0.29 20.00 9.00 28.50 6.30 1.26 [49]
340 20 3.57 0.25 45.00 15.00 21.80 10.90 1.90 [50]
340 20 4.08 0.15 30.00 7.70 27.10 5.76 9.81 [51]
360 25 1.24 0.14 25.20 35.60 39.90 1.40 11.30 [52]
360 25 0.44 0.11 15.00 56.60 32.70 0.15 4.85 [52]
340 20 1.25 0.28 23.00 29.00 19.00 17.33 6.67 [53]
340 20 0.13 0.15 38.00 7.30 21.00 8.00 4.67 [53]
300 15 1.04 0.18 35.00 20.00 10.00 20.00 5.18 [54]
300 15 0.53 0.14 25.00 25.00 8.00 15.89 9.29 [54]
320 30 0.23 0.10 35.00 10.00 10.00 11.67 16.67 [55]
320 30 0.23 0.09 35.00 8.00 38.00 3.33 11.00 [55]
320 30 1.09 0.15 20.00 38.00 15.00 13.33 3.33 [56]
330 15 0.37 0.13 41.00 27.50 19.40 14.06 2.06 [57]
300 10 0.26 0.12 36.80 9.40 9.50 4.00 3.00 [58]
340 30 0.37 0.13 41.20 16.30 14.00 11.57 1.50 [59]
250 20 0.76 0.22 33.70 10.60 7.40 5.40 1.08 [60]
320 30 0.44 0.15 41.80 9.70 15.00 13.33 1.67 [61]
320 30 0.43 0.21 34.00 14.30 9.60 12.00 1.50 [62]
300 25 0.17 0.09 41.70 6.00 10.30 7.20 2.07 [63]
300 25 0.14 0.09 37.80 7.30 8.30 5.40 1.73 [63]
320 30 0.12 0.15 35.70 15.10 20.80 7.57 9.83 [46]
330 15 1.60 0.29 47.20 20.00 25.00 4.17 5.50 [64]
320 15 1.28 0.24 52.00 8.50 18.00 17.33 3.33 [65]
400 20 0.39 0.14 36.80 35.00 40.00 3.67 0.61 [66]
340 30 0.11 0.15 45.00 15.00 23.87 11.33 1.89 [67]
320 30 0.11 0.14 27.00 13.00 5.00 3.33 2.67 [68]
320 30 2.45 0.32 40.50 8.30 9.70 9.00 1.80 [69]
310 30 0.52 0.19 32.30 45.00 14.50 12.50 2.50 [70]
300 25 3.82 0.17 30.00 20.00 8.50 18.78 5.22 [71]
300 15 0.27 0.13 13.00 50.00 25.00 5.33 0.67 [72]
320 30 1.12 0.25 47.20 10.20 35.50 4.13 2.11 [73]
330 15 0.26 0.17 35.50 23.50 21.40 13.60 1.30 [74]
300 10 4.44 0.32 31.80 15.10 8.10 5.50 1.00 [75]
300 25 0.11 0.01 32.00 12.00 36.00 10.33 1.67 [76]
320 5 2.0E-3 0.05 27.30 31.90 36.40 2.26 13.33 [77]
320 30 0.04 0.02 38.60 11.70 11.80 10.07 1.33 [78]
290 5 2.57 1.14 30.14 24.28 40.80 9.16 15.07 [79]
350 20 1.46 0.94 45.00 57.00 10.00 15.00 6.66 [80]

important to the outputs. adsorption energies (CO2 and H2) using DFT approach, even though very
Though other outputs i.e., SOi (i > 2) and SPi (i > 2) were found to be large discrepancy might incur. The total energy is expressed by
statistically insignificant to the variations of inputs, the possible reasons assuming all ground-state properties being functionals of the charge
might lie in the complexity of the system (non-linearity) and this density [81]:
ANOVA methodology fails to capture the statistical significance caused
Et [ρ] = T[ρ] + U[ρ] + EXC [ρ] (3)
by the variation of inputs. In this work, we will not only investigate the
impacts of those statistically significant responses, but also some other where Et [ρ] refers to total energy (kJ.mol− 1), T[ρ] refers to the kinetic
selectivity will be explored (CO, CH4, SOi and SPi)-where i varies from 2 energy (kJ.mol− 1) of a system, ρ refers to the density of non-interacting
to 10, in hope to provide a more broader and holistic view of this particles (kg.m− 3), U[ρ] is defined as the classical electrostatic energy
complex system. (kJ.mol− 1) due to Coulombic interactions, EXC [ρ] includes all many-body
contributions to the total energy (kJ.mol− 1), in particular, refers to the
3.2. Density functional theory (DFT) analysis for binding energies exchange and correlation energies. The semi-empirical dispersion-
correction schemes were deployed during the DFT calculation.
In this work, the intrinsic binding energies were estimated based on
the provided compositions from literatures expecting to shed some lights
of its impact on conversion and product distribution. From molecular 3.3. Mass transfer analysis
binding energy perspective, the metallic iron (111) plane (for purpose of
cleaving surfaces of crystallite) was arbitrarily chosen to estimate the In this work, the Mears’ criteria was employed [16,82,83]:

1026
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Fig. 2. MSE and MARR analysis of experimental data and model predictions.

ri ρdp n √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Mears = (4) ΦB MB T
km ci DAB = 7.4 × 10− 8
(9)
ηB vA
where ri refers to the CO2 hydrogenation rate (mol.m− 3.kg−cat1.s− 1), ci
refers to concentration of reagents (mol.m− 3), n is defined as the reac­ where ΦB is the association factor of H2 (− ), vA is the molar volume of
tion order (n = 1, we assume the first order reaction for the simplicity of CO2 in gaseous phase (cm3.mol− 1), ηB is viscosity of H2 (m.s− 1), T refers
calculation), dp is defined as average particle diameter (m). For corre­ temperature (K).
lations of mass transfer coefficient km (m.s− 1) from reactor, the cylin­ Because of the varieties of reactors that existed in different refer­
dered shapes of catalyst were arbitrarily assumed and the Frössling ences, it is very challenging to evaluate the mass transport both radially
correlation was employed as the following [84,85]: and axially within the reactor. One convenient approach to assess the
intra-particle diffusional limitation is to use the dimensionless Thiele
Sh = 2 + 0.6Re1/2 Sc1/3 =
km dp
(5) modulus (Φ):
DAB √̅̅̅̅̅̅̅̅̅̅̅̅̅
ρ(− ri )
ρND2 Φ = Lg (10)
Re = (6) De ci
μ
0.357 Dm 0.641 1/3
DAB,eff B= Re Sc (11)
Sc = (7) εp De
vB
[ ]1/2
where Sh, Re and Sc are Sherwood, Reynolds and Schmidt dimensionless 10− 3 T 1.75 M1H + MCO1

number (− ), respectively, DAB,eff is the effective binary H2 in CO2 Dm = [

2
]2
2
(12)
diffusivity (m2.s− 1), N refers to rotation rate (min− 1), the MA refers to P vH2 + v1/3
1/3
CO2
molecular weight of CO2 (g.mol− 1), vB is kinematic viscosity of hydrogen
(m2.s− 1). To calculate DAB,eff , the correlation between effective binary Where Lg is geometry characteristic length (m), De represents effective
diffusivity and binary diffusivity is depicted as the following: diffusivity (m2.s− 1), Dm represents binary gas diffusivity (m2.s− 1), B is
DAB εp the Biot number (− ), εp (unless specify, we arbitrarily chose 0.6 in all
DAB,eff = (8) calculations if reference did not report this value) is porosity of catalyst
τp
particle prior to reaction (− ), ρ is density of catalyst (kg.m− 3), ci refers to
where εp is porosity (unless specify, we arbitrarily chose 0.6 in all cal­ the concentration of reagents (mol.m3), MH2 and MCO2 refer molecular
culations if reference did not report this value) and τp refers to tortuosity weight of H2 and CO2 (kg.mol− 1), vH2 and vCO2 refer molecular volume of
of catalyst (τp = 2.0, this is an empirical value that are unbiasedly used H2 and CO2 (m3.g− 1.mol− 1), respectively.
for all calculation with consideration for consistency). As H2 is used in
the excess, the DAB (cm2.s− 1) diffusivity of H2 in CO2 was calculated as 3.4. Artificial Neural Networks (ANNs)
the following [86]:
With the advancement of logic chip, especially in the aspects of
availability and cost-effectiveness, the soft computing techniques using

1027
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Fig. 3. Graphical column plot of 11 inputs with 21 outputs, and the preliminary statistical ANOVA of inputs versus outputs. The graphical column data is constructed
according to the Box-Behnken design data matrix.

ANNs have recently experienced a persistent tailwind. The configuration Where x is the input to a neuron. This is also known as a ramp function.
of ANNs originally derives from how neurological system handles the To avoid drying problem, the learning rate is arbitrarily set at a rela­
information. The feature of constructed ANNs relies on neural topologies tively lower rate (0.0001). To consistently compare with the results from
(hidden layers) and learning procedure (weight, activation functions our previously published works using deep learning [88–90], the leaky
etc.). The multilayer perceptron (MLP) is one of the well-formulated and ReLU is not used in this work, let alone the deployment of leaky ReLU
widely studied ANNs in both academia and industry [87]. For MLP would appreciably reduce the overall performance of ML. Ever since its
activation function, both linear and non-linear activation function exist. first release in 2017, TensorFlow (Google Brain’s second-generation
If a MLP has a linear activation function in all neurons, no matters how system), has evolved as an important role player in ML, due to its
complicated the topological architecture is, any number of layers can be extraordinary flexible architecture allowing for the easy deployment of
numerically downsized to a fundamental two-layer input-output model. computation across a variety of platforms (CPUs, GPUs, TPUs), and from
For modeling the complicated system, the nonlinear activation functions desktops to clusters of servers to mobile and edge devices. Therefore, in
are often employed in some layers. The common sigmoid shaped acti­ this work, all ML process was conducted through TensorFlow using
vation function is calculated as the following: Python language. For instance, the input i.e., CO2 conversion (XCO2) was
y(wi ) = 1 + (1 + e− wi − 1
) (13) initially normalized and then fed into algorithm in form of txt via numpy
command. After preprocessing in the data matrix, the normalized data
Where yi is the output of ith node (neuron) and wi is the weighted sum of will be saved as training dataset prior to the prediction. For the estab­
input connection. In recent developments of deep learning, the rectified lishment of hidden layers, five hidden layers were used, the neuron
linear unit (ReLU) is more frequently used as one of the possible ways to nodes at different hidden layers from the first to the fifth are: 64, 256,
overcome the numerical problems related to the sigmoid. In addition, 128, 64 and 32, respectively. Each hidden layer deploys Sigmoid func­
the intrinsic advantages of sparse activation, better gradient propaga­ tion as activation function. The LazyAdam function was used for opti­
tion, faster and effective training of deep neural architectures on large mizer, and both accuracy and mean-square error were deployed as
and complex datasets etc., allow ReLU deep learning achieving quite monitoring indicator in loss function. For model prediction and vali­
many successful applications. As defined as positive part of its argument, dation, the split ratio was arbitrarily set at 0.2 (as our previously
the activation function is calculated as the following: cross-out validation ratio was also set at in this ratio and results were
{ reasonably acceptable) and epoch was set at 10000. Finally, the binary
f(x) = x+ = max(0, x) =
x + |x|
=
x
(14) variations of all inputs (11 parameters) towards the changing patterns of
2 0 XCO2 were automatically generated and plotted with approximately

1028
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Table 2
±25% uncertainties. Summary of mechanisms for FT synthesis, where S represents the active catalytic
site, - refers to not available, AHM represents Alkylidene–hydride–methylidyne,
PD refers to product distribution, D refers to Deuterium, DFT refers to density
4. Results and discussion
functional theory, ASF refers to Anderson–Schulz–Flory, CLD refers to chain
length dependent, O/P represents olefin to paraffin ratio, and * refers to the
4.1. Reaction kinetics and mechanism comprehensive product distribution model.
Mechanism Monomer Initiator Supporting Reference
It is commonly believed that the CO2 hydrogenation reaction evidence
occurred through two major steps: the reverse water-gas shift (RWGS)
Carbide S–CH2 Carbide formation [92]
reaction and subsequent FT synthesis, while the direct reaction is
–
on several transient
generally found to be marginal [91]. For the RWGS reaction, two main metals
reaction mechanisms dominate. They are the redox mechanism (fit for Enol S–CH–OH CHOH Initiation caused by [93]
the reaction over metal oxides) and associative mechanism (fit for the alcohol and
iron-based catalyst by introducing the formation of active intermediate aldehyde
Formate CO S–OH Co–MgO catalyst [94]
formate). Among RWGS and FT synthesis, the appreciably lower energy and chain
barrier for RWGS is tandem with fast kinetics of CO2 conversion during lengthening
RWGS, especially on iron-based catalyst, making the reactions of FT CO insertion S–CO S–CH3 Transients/back- [95]
synthesis as the main rate determining step (RDS) of the overall re­ transients of
gaseous products
actions. By considering the main products, the following reactions are
Alkyl S–CH2 S–CH3 C2 are mainly [96]
summarized: ethylene in inert gas
Alkenyl S–CH2 S–CH2– Similar PD with 1- [97]
(15)
–CH2
CO2 + H2 ↔ CO + H2 O alkene predominate
AHM S–CH + S–
–CH–CH2–S CH + H as monomer [98]
1 S–H by D trace analysis
CO + 2H2 → Cn H2n + H2 O (16)
n H-assisted CO S–CO + – DFT and isotopic [99]
dissociation S–H studies
2n + 1 1 CO Insertion- S–CH2+ S–CH3 Alcohol probes co- [100]
CO + H2 → Cn H2n+2 + H2 O (17) carbide S–CO feeding and dual
n n
ASF superposition
1 n− 1 CO Insertion- S–CO S–CH3 Exponential [101]
CO + 2H2 → Cn H2n+1 OH + H2 O (18) CLD model* descending O/P
n n ratio
Carbide-non- S–CH2 – O/P independent to [12]
Where eq.(15)–(18) refer to the CO formation via RWGS reaction, and CLD model* carbon numbers on
formation of 1-alkenes, n-alkanes and alcohols via FT synthesis. Please alloy catalyst
note particularly that most kinetic studies tend to ignore the modeling of
oxygenates (alcohol), branched olefin/paraffins due to their negligible
concentrations comparing with other linear hydrocarbons (olefin and [102,103]. For Fe-based catalyst, three different active sites mechanisms
paraffins). In this work, considering the availability from literature re­ have been proposed, of which the Fe3O4 is responsible for RWGS, the
ports, the product distribution is mainly focused on linear paraffin and carbide/Fe attributes for the formation of main hydrocarbons via chain
olefin with the carbon number varying from 1 to 10. propagation, and the Fe metal is responsible for olefin re-adsorption and
Unlike the conventional polymerization reaction, during the stage of secondary hydrogenation [104,105]. Both carbide/Fe and metal Fe are
FT synthesis, the feed (CO and H2) needs to be firstly converted to a critical for the formation of the main products, which is essential to the
monomer and initiator on the active catalytic site, and then it proceeds FT synthesis pathway. When a constant chain growth probability α for
to the polymerization for FT synthesis. Broadly speaking, two types of all growing species is assumed (− ), the famous Anderson-Schulz-Flory
mechanisms are widely accepted, namely one-path monomer/initiator (ASF) distribution is calculated as the following:
(i.e., carbide, enol, alkenyl, CO insertion etc.) and two-paths for initia­ xn = (1 + α)αn− 1
(19)
tion/propagation (i.e., H-assisted CO dissociation, CO insertion-
carbide). They are briefly summarized in Table 2. The intrinsic bene­ Where n refers to carbon number (− ), xn is molar fraction of the products
fits of using the top-down rules, models and simulations are obvious. If at nth carbon number (− ). There are two popular top-down models
top-down model yields the reasonable regression fit against experi­ based on mechanistic hypothesis, which are power law models (lumped
mental datasets, the underlining mechanism could be potentially and type) and Langmuir-Hinshelwood-Hougen-Watson (LHHW) models,
logically applied in explaining the experimental observations. Most of respectively. Though LHHW models provide more information on in­
these mechanisms from Table 2 to the largest extent, obtained some dividual product formation rate, its intrinsic complexities such as
supporting evidences from reported works. However, all these listed Byzantine theories for active intermediates assumptions, and very
mechanisms did not consider the impact of catalyst support on the limited reported works etc., all limits its direct applications in straight-
initiation and chain propagation, which leaves more space for the future forward sizing the reactor [106,107]. The power law models, on the
works to further narrow the gap between experimental results and the other hand, still shows its merit due to its intrinsic simplicity for the
proposed top-down models. It is also confident that more accurate and reactor design [84,108]. In this work, a few reported power law kinetic
complex mechanisms such as those two-pathway mechanisms (com­ models together with generic rate expression for both CO and CO2
bines some simultaneously independent pathways-initiation, chain (without prior power number assumption) were used against reported
termination and propagation) will be more likely to be discovered by dataset for data regression. The selected models from data regression
considering the intrinsic complexity of CO2 hydrogenation via FT syn­ and their corresponding regressed parameters are shown in Fig. 4 and
thesis using iron-based catalyst as the advancement of theorical tools Table 3, respectively. Among these collected power law models, their R2
and experimental analytical instruments. behaves differently. It is obvious to find out that the generic power law
Regarding the mechanisms of the CO2 hydrogenation reactions, the expression proposed in this work (without prior assumed power
works of Schulz and Visconti and their collaborators have made sub­ numbers for both CO and H2 partial pressure), seems to yield best model
stantial contributions to the understanding of reaction mechanisms fit (R2 = 0.8181), while other models with pre-assumed power number

1029
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Fig. 4. Data regressions from reported lump power kinetics expressions, where all data regressions were performed through global optimization via Levenberg-
Marquardt (LM) algorithms. The objective functions were set at less than 1E-10 as the converge condition. The maximum itineration of 1000 was set for all data
regression. The rco was estimated at the reported TOS period from literatures.

1030
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Table 3
The reported kinetic rate expressions during the FT synthesis, where R2 refers to the correlation coefficient and α, β, γ δ are regression constants of power law models.
Please note particularly that some rate values were not directly reported from literature collected, thus, the estimation based upon their corresponding experimental
conditions are needed. For rate estimation, their corresponding period for kinetics were arbitrarily set at their reported time on stream-TOS.
Equation Catalyst α β γ δ Reference R [2] Mechanism

(− PCO PH2 ) Fe 180.01 19.96 – – [110] 0.5827 Enol

r =α
(1 + βPCO )2
(− PCO PH2 0.74 ) Fe 60.11 7.73 – – [111] 0.3117 Empirical
r =α
(1 + βPCO )2
(− PCO PH2 0.5 ) Fe 0.05 − 0.96 – – [112] 0.6518 Carbide
r =α
(1 + βPCO )2
(− PCO PH2 ) (− PCO ) Fe 2.37E4 − 1.32E-5 1.44 – [113] 0.5399 Empirical
r =α +β
(1 + γPCO )2 (1 + γPCO )2
(αPH2 + β)PCO Fe/Co 2.23E9 − 1.25E8 1.77E2 – [114] 0.5295 Carbide
r = −
(1 + γPCO )2
(− PCO PH2 1.5 ) Fe/Co 3.78E4 5.77E2 – – [107] 0.3564 Empirical
r =α
(1 + βPCO )2
(− PCO 0.5 PH2 0.75 ) Co 7.56 2.01 – – [115] 0.6836 CO-dissociation
r =α
(1 + βPCO 0.5 )2
PH2 β PCO γ Fe 2.5E-4 4.36 1.62 1.03 This work 0.8181 Empirical
rCO = α
(1 + δPCO )2
PH2 β PCO2 γ Fe 0.45 0.74 0.38 0.06 This work 0.2015 Empirical
rCO2 = α
(1 + δPCO2 )2

Fig. 5. Predictions using binding energy (eV) and active metal size (nm) for the conversion (CO2) and selectivity (CO, CH4). The binding energies between the
reported catalyst and reagents (CO2 and H2) were calculated based on the provided information from each literature. The crystallite lattice was established according
to the provided information from each literature (stoichiometric ratios, the amounts of promotors etc). Prior to DFT calculation, all constructed structures were
optimized by the DMoI3 algorithm. The active metal size in each literature was estimated from direct TEM measurement or XRD approximations if this value was not
directly reported from literature.

1031
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

for either H2 or CO partial pressure seem to yield a relatively poorer fit. dissociates into HCO* and O*, then the newly produced HCO* needs to
From mathematical perspective, empirical power number assumptions migrate from Fe3O4 site to the adjacent carbide/Fe site to continue the
for the partial pressure (CO and H2) are not favorable in improving the initiation and chain growth reactions. A further dissociation to CH*
prediction accuracy of kinetic models. In addition, the partial pressure produces the monomer necessary for chain growth or CH4 formation.
ratio between H2 and CO of this work is around 2.7 (4.36/1.62), which Indeed, the CH4 selectivity depends on the competition between the
indicates that H2 needs to be stoichiometrically surplus so that the re­ CH*+H* (termination) and CH*+CH* (chain propagation). Conditions
action (eq. (16) and eq. (17)) could be skewed towards the formation of that favor skewing from CH4 termination to chain propagation will
hydrocarbons during FT synthesis, which also agrees with experimental suppress the methane selectivity. Reducing active metal size, increasing
results. This is also the reason that the ratio of H2/CO2 is often kept at 3 operational pressure, and lowering H2/COx ratio, are found to help
in most of the reported works. In addition, the target vs calculation reducing methane selectivity practically. After the chain initiation, more
(Fig. 4) suggests that the proposed models from previous scholars tend to complicated chain reactions (two mechanisms, namely, the linear chain
overshoot the calculated value, especially at the lower range of dataset. prolongation and chain branching) will occur, of which are summarized
From mechanism perspective, CO dissociation mechanism tends to yield by many very good reviews on FT synthesis [120,121]. For example, the
a relatively higher R2 = 0.6836 followed by carbide mechanism with R2 linear chain prolongation is the reaction between an alkylene and a
= 0.6518. It is crucial to address that most of previously reported power methyl species to form an alkyl intermediate that is attached to the
law rate models was derived from FT synthesis condition (CO + H2), not active site with its α-C atom. Branching reactions follow the same
from CO2 hydrogenation condition (CO2+H2), and this might possibly mechanism, but the alkyl intermediate is attached with its β-C atom, and
explain why the correlating CO2 rate with partial pressure of CO2 and H2 further growth of these intermediates leads to the formation of the
in a power law expression in this work tends to yield a poorer R2 = branched products. The branching probability has been observed being
0.2015. Indeed, the challenge of direct mirroring the similar mathe­ decreased as an increase of carbon numbers. This is due to spatial con­
matical kinetic expressions from FT synthesis to the CO2 hydrogenation straints, which are stronger for branching reactions and bigger size of
surely exists. Yet, it agrees with the fact that the RWGS and FT synthesis the intermediate species.
are different reactions and occurred on different catalytic sits, which
have been reported by many scholars [109]. Catalyst wise, the selected 4.2. Operational parameters
models are independent to the type of catalyst (either Co or Fe), sug­
gesting its potentially universal applications in kinetic modeling works. The FT synthesis is sensitive to the operating conditions. Two most
Besides kinetics of power law expression, the monomer formation frequently mentioned measurable operational parameters are tempera­
(either CO* or CHx*) during CO2 hydrogenation is a critical step, which ture and pressure. For example, temperature affects desorption and
is also likely and closely linked with RWGS. Therefore, investigating the hydrogenation rates, which inevitably influences the subsequent chain
conversion of (CO2) and selectivity (CO and CH4), is meaningful in un­ propagation and the degree of hydrogenation. It also affects reaction
derstanding the chain initiation mechanism. During the initial reactions, phase and steam quality. During the CO2 hydrogenation using Fe cata­
it is believed that both the binding energy and the size of the active lyst, the range of temperature is often kept at 300–360 ◦ C (some re­
metal play the important roles [116–118]. Therefore, we attempted to searchers reported the exceptional high temperature reaching over
use these two parameters, which represents the intrinsic features of 400 ◦ C, but it is not common practice among most scholars), which is
catalyst, to investigate their influences upon conversion and selectivity often classified as high temperature FT (HTFT). Pressure influences the
(CO, CH4). According to Wang et al. the Fe (111) facet was suggested to productivity as well. The ambient pressure process was developed by
be the most active site for CO2 hydrogenation [119]. Therefore, the facet Germany in the 1930s, which was called German normal-pressure pro­
(111) in all DFT calculations was arbitrarily chosen among all references cess, but it achieved low CO selectivity and poor productivity (the
to make a consistent comparison. The impacts of these two parameters product yield was only 0.1 kg h− 1 kg− 1 catalyst). By increasing pressure,
are summarized in Fig. 5. For CO2 conversion, two higher conversion the volumetric productivity and liquid product are enhanced, which in
regions, namely higher binding energy-smaller metal size and larger turn favours the removal of the products synthesized on the surface of
metal size-higher binding energy, are identified as being ideal for larger catalyst. Moreover, the higher H2 partial pressure enhances the lifetime
CO2 conversion. It is more reasonable to accept the region with higher of catalyst (two to three times longer than that of conditions that were
binding energy-smaller metal size and reject the region with larger metal operated at lower pressure). However, the increased productivity leads
size-higher binding energy. The binding energy (>-10 eV) and metal size to the increased heat release per unit volume, which needs the
(<10 nm) are favorable for the adsorption of CO2 to become active demanding heat exchange surface of the reactor during CO2 hydroge­
COO* on the surface of the catalyst. The hydrogenation leads to the nation. The ML predictions of the influences of pressure and temperature
formation of the intermediate HCOO*, with activation energy varies in upon conversion and some selectivity are depicted in Fig. 6.
the range of 45–80 kJ mol− 1, which generally agrees with the reported When optimize a complicated reaction system, both thermodynamic
range from other literatures. Then the direct dissociation leads to the and kinetic often need balancing adroitly. The kinetic is enhanced by the
formation of CH2* and other by-products (CO, formate, methanol, and increased temperature, while this increased temperature results in
formaldehyde) via side reactions. This also orchestrates with kinetic adverse effect on adsorption equilibrium constant, which is critical for
mechanism studies in Table 2 (CO-dissociation mechanism). The CO the availabilities of reagents to the surface of catalyst. Therefore, to
selectivity behaves quite differently from that of CO2 conversion. The better take advantage of the kinetics and thermodynamics simulta­
higher SCO appears at the lower binding energy-relatively larger metal neously, a general industrial practise is to setup a two-stage process-a
size region. Many factors might contribute to this baffling result. For high-temperature reaction zone (HTRZ) coupled with a low-temperature
instance, the surface reactions among RWGS and subsequent FT syn­ reaction zone (LTRZ). For CO2 conversions, the higher conversion falls
thesis on different active catalytic sizes are complicated. The higher CO2 into higher temperature range together with medium pressure region.
binding energy seems to marginally contribute to CO selectivity during This means that the boost of kinetics by increasing temperature during
FT synthesis. In another word, the increased CO2 adsorption and RWGS is crucial. In addition, a favorable CO2 conversion is observed
dissociation might hinder the adsorption of CO and H2. Additionally, the with an increase of pressure (H2 partial pressure) during the CO2 hy­
nano size impact seems to be more pronounced during RWGS than that drogenation, which agrees with eq. (15). For CO selectivity, the higher
of subsequent FT synthesis. This seems to indirectly align with the SCO appears at relatively lower temperature together with medium
assumption that RWGS and FT synthesis occurred on different catalytic pressure region. This indicates that adsorption of CO on the surface of Fe
sites (RWGS occurs on Fe3O4, while subsequent FT synthesis occurs on catalyst is paramount to the subsequent FT synthesis. For example, an
carbide/Fe and Fe metal accordingly). In this vein, the HCOO* increased temperature from 200 to 600 ◦ C could lead to an adverse effect

1032
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

Fig. 6. Predictions based on P (pressure) and T (temperature) made by ML using literature reported data, where Xco2 refers to CO2 conversion, Sco and SCH4 refers to
CO and CH4 selectivity, Si refers to selectivity for paraffin or olefin with specific carbon number, and i refers to the carbon number. Due to the limited space, the olefin
(C2–C5, C9–C10) and paraffin (C1–C5, C9–C10) were selected for purpose of indicating the trend.

on equilibrium constant by a factor of about 80 [122]. Thus, the relative products. For paraffin selectivity with different carbon numbers, the
lower temperature is preferred particularly for improving the CO relative higher selectivity region seems to appear at the high
selectivity. The spectrum of product obtained from FT synthesis is broad pressure-low temperature region, though SP5 appears to be a little bit off
and varies from methane (CH4) to heavier hydrocarbons. It is often this trend. This might be due to uncertainties caused by reference reports
found that the HTFT process tends to produce higher CH4 selectivity and overall global optimization during ML. Nevertheless, the trend is
(sometimes reaches >10%). It has been shown that expensive tail gas obvious. Regarding olefin selectivity, no appreciable trend based on
recovery procedures can be replaced by simple recovery of H2 from tail searching for relative larger olefin selectivity is captured. It is commonly
gas if the CH4 selectivity from Fe-based catalyst can be suppressed to believed that the chain growth probability for paraffin and olefin are
1.0–1.5 wt%. However, achieving such low CH4 selectivity for industrial different. The sum of reaction is often defined as the following if other
FT synthesis process has yet been reported among all commercial pro­ products such as oxygenate are ignored:
cesses [123]. Comparing to other paraffins with different carbon
ri = ri,P + ri,O (20)
numbers, the manipulation of CH4 selectivity requires a different
approach, this is due to different formation mechanism on the surface of
Where ri is sum of reaction for an ith given number of carbons (kg.s− 1),
catalyst and its different thermodynamic favorable formation conditions
ri,P and ri,O are paraffin and olefin formation rate at ith carbon number
comparing with other paraffins and olefins. Albeit two low CH4 selec­
(kg.s− 1). One of the most suitable empirical correlations to describe
tivity regions are found on the simulation result, namely low
olefin probabilities with both temperature and pressure is depicted as
temperature-high pressure region and high temperature-medium pres­
the following:
sure region, it is more reasonable to accept low temperature-high
pressure region while reject high temperature-medium pressure re­ 1
αi = ( ) (21)
gion. This is due to the facts a) lower temperature suppresses the CH4
a
PHi
1 + k′ exp − EAi 2

selectivity, b) higher pressure facilitates the formation of liquid

0i RT bi
PCO
2

1033
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

emphasized that, almost all literatures (we collected so far) did not
Where αi is growth probability (− ), k′0i refers to adsorption constant (kJ.
provide this Mears estimation. Thus, the uncertainties, or discrepancy by
mol− 1), EAi is activation energy for an ith given number of carbons (kJ.
calculating these unitless numbers for each reference potentially remain
mol− 1), PH2 and PCO2 are partial pressures (atm), ai and bi are constants
high. It is quite interesting to find out some calculated Mears values are
(− ). Obviously, this inexplicit non-linear expression suggests complex
bigger than 0.15 in a few reported literatures. Indeed, this lies in the fact
relationship between the temperature and pressure. In addition, the
that larger numbers of literatures did not aim to investigate the kinetic
active reactivity of olefin also adds its complexity to the product dis­
studies. For internal diffusion, one can use Thiele modulus to estimate
tribution with different carbon numbers.
the impact of internal diffusion upon the overall rate. If Thiele modulus
is small, it indicates that the surface reaction is rate-limiting. For CO2
4.3. Mass transfer impact conversion, the higher conversion appears at two regions: medium
Mears together with medium Thiele modulus, and higher Mears together
To look at the mass transfer impact upon reaction system among all with low Thiele modulus (this is not reasonable and can be rejected)
available references that were collected, this review tries to use both region. This indicates that the proper control for both external and in­
Mears’ criteria and Thiele modulus to find out any potential impacts ternal diffusion will enhance the availability of CO2 on the surface
upon the conversion and product distribution. The ANNs prediction of catalyst.
the influences of Mears’ criteria and Thiele modulus upon conversion For CO selectivity, the preferable region falls into both lower Mears
and some selectivity are depicted in Fig. 7. For external diffusion (intra- together with lower Thiele modulus region. This indicates that both
particle mass transport), the Mears’ criteria is a useful tool to assess external and internal diffusion are equally important for CO selectivity.
whether the mass transfer from bulk gas phase to catalyst surface can be This seems to be in line with the assumption that RWGS and FT synthesis
ignored via referencing on the measured rate of reaction. If the quantity occur at different catalytic sites. The molecule CO, once is reduced from
of Mears is less than 0.15, then the external diffusion could be ignored, RWGS reaction, needs to diffuse to other catalytic site to proceed the
or the rate of reaction is governed by the reaction. It should be

Fig. 7. Predictions (Mears and Thiele modulus) made by ML using literature reported data, where Xco2 refers to the CO2 conversion, Sco and SCH4 refers to CO and
CH4 selectivity, Si refers to selectivity for paraffin or olefin with specific carbon number, and i refers to the carbon number. Due to the limited space, the olefin
(C2–C5, C9–C10) and paraffin (C1–C5, C9–C10) with selected carbon numbers are listed for purpose of indicating the trend.

1034
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

subsequent FT reaction. For CH4 selectivity, the corner with relative olefin and paraffin). Experiments that aim to collect data especially for
lower Mears and Thiele modulus value is found to associate with rela­ kinetic data collection need cautious efforts to maintain the reaction rate
tively lower methane rate, thus the enhancement of external and in­ to be dominated by intrinsic kinetic not by the mass transfer. Otherwise,
ternal mass transfer will be favorable to reduce methane selectivity. For regardless of top-down models and bottom-up data driven correlation,
other paraffin with different carbon numbers, it shows a recognizable the parameters regressed from either proposed or correlated models
pattern for favorable region (Mears<0.15, Thiele modulus<0.2). Obvi­ could not essentially reflect the intrinsic kinetic features of the reactions.
ously, the overall rate is governed by kinetics if reaction conditions were Taking olefin to paraffin ratio (OPR) during the stage of FT synthesis for
kept this region. For olefin with different carbon numbers, no appre­ example, with an increase of carbon numbers (i > 5), it is widely
ciable patterns were identified. Possible reason might be because the accepted that the Van der Waals force gradually becomes phenomenal
olefin formations are also associated with active intermediate re- for the longer chain 1-olefin re-adsorption and secondary reactions
adsorption and chain termination. And these reactions might occur at [124]. If comprehensive kinetic models were fed against experimental
different catalytic active sites. In addition, it also needs to be addressed data, whether the kinetic constants generated from data regression
that not all literature works were done for the purpose of conducting could reflect the intrinsic kinetic or not, depends on how well the mass
kinetic modelling, thus it results in many experimental conditions being transfer might be manipulated. If the objective is to tailor the product
held at mass transfer-controlled region instead. In addition, for kinetic distribution towards waxy hydrocarbons by CO2 hydrogenation via FT
studies especially for relatively slow kinetics or low reaction rate, the synthesis using iron-based catalyst, the mass transfer (both external and
mass transfer resistance surely plays a critical role in determining the internal diffusion) management become crucial. Thus, in these sce­
quality of data. Comparing among the products, as the chain length narios, the mass transfer resistance associated parameters need to be
increases, the mass transfer resistance will become more phenomenal implemented and monitored.
largely being caused by the space constrain. In this vein, mass transfer
resistance is more critical to these longer chain hydrocarbons (both

Fig. 8. Predictions (TOS and GHSV) made by ML by using literature reported data, where Xco2 refers to CO2 conversion, Sco and SCH4 refers to CO and CH4
selectivity, Si refers to selectivity for paraffin or olefin with specific carbon number, and i refers to the carbon number. Due to the limited space, the olefin (C2–C5,
C9–C10) and paraffin (C1–C5, C9–C10) with selected carbon numbers are listed for purpose of indicating the trend.

1035
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

4.4. Other parameters a pattern, evolving from left-hand corner (higher TOS and lower GHSV)
via middle range to the edging region with medium TOS and larger
Regarding to operational conditions, apart from pressure and tem­ GHSV, except O5 results. The possible reason for this discrepancy of SO5
perature, the impacts of other parameters on conversion and selectivity (pentene) might lie in its physical feature, of which its flashing point is
were also explored. As suggested from results of ANOVA, it is quite close to the ambient temperature. This inevitably increases the loss of
challenging to find out obvious significant statistics upon these impacts. sample collection during trap swap and trap dismantling. In fact, this
For example, the gas hour space velocity (GHSV) and time on stream often contributes to big uncertainties during the liquid sample collection
(TOS) are also quite important descriptive parameters for FT synthesis, in practice. Maintaining medium TOS (150 h) and larger GHSV
which is frequently mentioned in FT synthesis studies. The impacts of (>30000 L kg− 1 h− 1) tend to enhance the olefin selectivity, especially
TOS with GHSV upon conversion and selectivity with different carbon with longer chain. The trend of paraffin selectivity responding to the
numbers are summarized in Fig. 8. The TOS is a useful tool to gauge variations of TOS and GHSV as an increase of carbon numbers from C2 to
stability of catalyst. The GHSV inversely impacts the conversion, of C10 is also appreciable. Starting from edge area (medium TOS and larger
which the lower GHSV leads to a higher conversion, and the vice versa. GHSV), the paraffin with longer chain tends to be formed at the corner of
The simulation results are shown in Fig. 9. For CO2 conversion, the lower TOS (<50 h) and larger GHSV (>30000 L kg− 1 h− 1), both of which
optimal area exists in the middle range of TOS and GHSV. The higher CO contributes to the relative lower deactivation results during FT
selectivity seems to appear at larger GHSV and medium TOS. As the synthesis.
RWGS and FT synthesis occur at different catalytic sites, the responses of Apart from those above operational parameters, we also attempted
CO2 and CO towards the variations of TOS and GHSV behave differently. to investigate the responses from the variations of particle size of cata­
For CH4 selectivity, the edging area in the middle range of TOS and large lyst (macroscopic level in micrometre scale) and active metal size
range of GHSV is found to associate with relatively lower CH4 selec­ (microscopic level in nanometre scale). For CO2 conversion, the larger
tivity. For olefin (from C2 to C10), the higher selectivity seems to follow conversion exists in the middle range of particle size and larger metal

Fig. 9. Predictions (Metal size/nm and Particle size/μm) made by ML using literature reported data, where Xco2 refers to CO2 conversion, Sco and SCH4 refers to CO
and CH4 selectivity, Si refers to selectivity for paraffin or olefin with specific carbon number, and i refers to the carbon number. Due to the limited space, the olefin
(C2–C5, C9–C10) and paraffin (C1–C5, C9–C10) with selected carbon numbers are listed for purpose of indicating the trend.

1036
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

nanoparticle size. A common belief is that the smaller particle size (in and more supportive data. For CH4 selectivity, according to prior
micrometric scale) tends to remove the impact of mass transfer resis­ research on FT synthesis [125], the smaller cobalt nanoparticle size (less
tance. This generally aligns with the previous results of Mears’ criteria than 10 nm) supported on mesoporous silica tends to increase CH4
and Thiele modulus analysis. Regarding the nanoparticle size, the selectivity, especially at higher temperature. This CH4 selectivity trend
pattern becomes baffling. While cobalt and ruthenium nanoparticles was also found in metal alloys catalyst such as Co–Ni during FT reaction
exhibit a clear particle size effect, there have been some contradictories [17]. Thus, reducing active metal size (<10 nm) will adversely lead to an
in the aspects of activities and selectivity trends for Fe-based FT catalysts increase of CH4 selectivity. The changing patterns of olefin with the
[116]. Scholars commonly agreed that these discrepancies are most optimal selectivity appear to evolve from the central area for carbon
likely to be attributed to different operational conditions possessing with number from 2 to 4, which corresponds to medium particle size and
different experimental uncertainties [117]. Another possible reason nanoparticle size, to the edge area locating at medium particle size and
might also lie in the fact that RWGS and FT synthesis occur on different small nanoparticle size for carbon number of 5–10. This indicates that
catalytic sites and some intermediates from reactions of RWGS and FT the medium particle size (300 μm) and smaller metal nanoparticle size
synthesis might not be critically sensitive to the nano-sized active metal. (<10 nm) favor the relative larger selectivity for longer chain olefin (C5
What works for CO2 conversion does not necessarily suit for CO selec­ to C10). The changing patterns for paraffin with optimal selectivity
tivity. Agreeing with our prior simulation results (Fig. 5), the ideal CO behave differently. Starting from edge region (medium particle size and
selectivity region appear at the right corner of the region corresponding small nanoparticle size) for C2 paraffin, the larger paraffin selectivity
to the larger particle size and larger nanoparticle size. The underlying with an increase of carbon numbers (C3–C10) seem to evolve into the
information is quite controversial to the current knowledge in enhancing region with larger particle size and medium nanoparticle size. There­
catalytic selectivity by downsizing the catalytic metal particles. As fore, to boost longer chain paraffin selectivity, the condition with the
mentioned from above, CO acts as an intermediate in this complex re­ particle size (>400 μm) and nanoparticle size (>50 nm) is preferred.
action during CO2 hydrogenation and size of active metal particles
might not be critical to those intermediates generated from secondary 4.5. Evaluation and optimization
reactions. However, it is quite interesting that the ML results also present
another corner (small particle size and small nanoparticle size) yielding Optimization of complex system is challenging. Unfolding those
larger CO selectivity. How to choose the correct arrangement of size statistically significant outputs in responding to the variations of inputs
parameters (particle size and active metal size) and reject the unrea­ clarify how significantly the system responds to the variation of the
sonable size parameters, definitely requires further experimental works selected variables. From ANOVA results, the key significant responses,

Fig. 10. Statistical significance binary impacts distribution for conversion (CO2) and selectivity (CO–SCO, methane-SCH4, ethylene-SO2 and ethane-SP2), A-L repre­
sents: A is stoichiometric ratio of Fe (− ), B is calculated binding energy from DFT (eV), C is loading percentage of active metal (%), D is particle size (μm), E is
temperature (◦ C), F is GHSV (L.h− 1.kg− 1), G is pressure (bar), H is TOS (h), J is the averaged metal size (nm), K is Mears number (− ), L is Thiele modulus (− ). Note:
the binary impacts with red transparent circle are those p-value less than 0.1, which is assumed to be statistically significant to the outputs. (For interpretation of the
references to colour in this figure legend, the reader is referred to the Web version of this article.)

1037
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

namely the conversion (CO2) and selectivity (CO, methane, ethylene and that have significant potentials for further development. First, the cur­
ethane), were chosen to investigate their responses towards the varia­ rent experimental works are mostly designed for reporting the break­
tions of all binary parameters (binary combination of all inputs). Their throughs from catalyst cheminformatics, very few considerations were
results are shown in Fig. 10. taken for establishing databank for the collective purposes. Inspired
In this work, the p-value of binary parameters less than 0.1 was from many commercial labs that millions of tests were conducted and
assumed to be significant to the response [87]. Regarding CO2 conver­ shared daily throughout the world, the improved data infrastructure
sion, there are 6 binary inputs showing p-value being less than 0.1, that allows uploading data to cloud-based infrastructure is highly rec­
which indicates significant response introduced by the variations of ommended for the future research and development. Secondly, it often
binarily combined operational parameters. The statistically significance takes arduous efforts for data scientists or theoreticians to determine and
of these investigated combined parameters towards the CO2 conversion process the dimension of datum matrix. With the advancement of ML
was sorted and results are shown in Fig. S1. Among them, the most two technology, more and more reliable and easy-to-use ML algorithms have
significant binary inputs are the Fe stoichiometric ratio (− )/averaged become available to the researchers both in academia and industry. This
metal size (nm) and particle size (μm)/Thiele modulus (− ), respectively. indicates that downstream of AI technology starts to outweigh its up­
To improve the CO2 conversion, the adroitly manipulating the param­ stream technology (i.e., data matrix dimensional identification and
eters of Fe stoichiometric ratio (10–20), average metal size (8–10 nm), optimization, data handling etc.). The ripples of AI applications are
particle size (400–600 μm), and Thiele Modulus (0.5–0.7) are preferable gradually dissipating from optimizing AI algorithm itself (downstream)
for maximizing CO2 conversion. For CO selectivity, it is found that the to size and process the dimensional data matrix (upstream). For
combination of particle size (<100 μm) and Thiele Modulus (<0.1) are instance, more flocks of researchers with particularly specific field
the most significant influences, thus keeping particle size being less than expertise (not just researchers with only computer science back­
100 μm and Thiele Modulus being less than 0.1 are favorable for grounds), who possess rich experiences in their specific fields, will
maximizing CO selectivity. For methane selectivity, minimizing it to participate in the realm of upstream dimensional data matrix identifi­
10% or less is preferable [126]. The pressure (bar)/TOS (h) and TOS cation and optimization. In particularly for CO2 hydrogenation via FT
(h)/average metal size (nm) are one of the most significant parameters synthesis using iron catalyst, the bottleneck might fall into the effec­
that influence the CH4 selectivity. Keeping the combined parameters tively and efficiently determining and cleaning the dimensions of data
(pressure (20 bar)/TOS (200 h) and TOS (200 h)/average metal size matrix. Indeed, this upstream of pre-processing and cleaning the data
(<10 nm)) in the suggested range are preferable for a relative lower will significantly determine the quality of learning results. Thirdly, the
methane selectivity. For ethylene selectivity, the temperature/TOS is integration of information between experimentalists and theoreticians
found to be most significant inputs. Manipulating the range of temper­ could be greatly encouraged by publishers, who could provide some
ature (300 ◦ C)/TOS (50 h) tends to lead to larger ethylene selectivity types of impetus such as waiving open access fees for those researchers
(+20%). For ethane selectivity, promotor loading percentage with their willingness to share or publish their raw data in the databank
(>5%)/Mears (<0.1) tends to lead to a larger selectivity (10%). Overall, with standardized data framework. Lastly, perhaps the most pivotal part
manipulating complex system with large dimensional variables is chal­ is the development of open-source databases and software tools. The
lenging. Indeed, the responding patterns of the conversion (CO2) and DECADE, is one of the most well-known prototype knowledge-based
selectivity (CO, methane, Oi and Pi, where i varies from 2 to 10) in system for catalyst selection and it was firstly reported in 1987 [127,
responding to the parametric variations are, in the many cases, intrin­ 128]. However, because of lack of financial incentives and no accom­
sically counter-active to each other. This makes the tailoring of product panying open-source ecosystem, the development of this type of soft­
distributions of CO2 hydrogenation via FT synthesis using iron-based ware tools for catalyst selection and optimization is still at the early
catalyst extremely challenging. There are a few reasons that are highly phase. Yet, the situations have improved during the last decade. The
likely to attribute to this intrinsic complexity of system (CO2 hydroge­ significant advances in infrastructure for open-source software devel­
nation via FT synthesis using iron-based catalyst): 1) the feature of re­ opment such as Github and open-source algorithms such as TensorFlow,
action is complicated, reflecting on its occurrence of catalytic reactions facilitate the rapid construction of ecosystem with open-source software.
(the reactions of CO2 hydrogenation and subsequent FT synthesis) on Thus, the underlining latencies i.e., higher chances for citation, data
different catalytic active sites. Thus, most patterns responding to the reproducibility and reliability, long-tail effect (some less investigated
variations of parameters are associated to the so-call secondary re­ but still important catalytic research area), could be unleashed to the
actions, and this type of high-order nonlinearity might yield some largest extent [35,129–131]. It is also worth noting that Python lan­
counter-active results; 2) feature of data collected is complicated, guage has become a widely adopted programming language among ML
reflecting on inconsistent standards for reporting datum for this process. community. Therefore, learning this powerful language will provide
For instance, many critical and necessary information i.e., average profound benefits in both academic and industrial community.
particle size of catalyst, GHSV, TOS, mass transfer coefficients (because
of different configurations of the system) etc., could not be obtained 6. Conclusions
from many reported works. This creates great challenge for establishing
the efficient data frameworks for ML; 3) the feature of experimental In summary, the collected reported data of CO2 hydrogenation via FT
operation is complicated, reflecting on great uncertainties in experi­ synthesis were fed into ANNs to investigate the parametric impacts upon
mental data collections. For example, there are no recommended stan­ the conversion (CO2) and selectivity (CO, methane, SOi and SPi, where i
dardized methods for reporting the baseline conditions and guidance for varies from 2 to 10). The dimensions of investigated data include 11
methodology of conducting gas analysis including the reference gas, inputs and corresponding 21 outputs. The lump power kinetic expres­
calibration methods and setting for refinery gas analyzer (RGA) etc; 4) sions from prior reports together with the general one (without pre-
The amount of data is still very limited in this specific topic. To fully assumed power values for individual partial pressure) proposed in this
unleash the latency of ML, the current amount of datum is not sufficient, work were fed with collected data and rigorously compared. A better fit
especially those datum from industrial scale or even polit scale are very kinetic model was found to be the generic kinetic expression proposed in
scarce. To validate or solve those baffling discrepancies need more data this work (with R2>80%). Several statistically significant binary pa­
to work on it. rameters (i.e., binding energy (eV)/average metal size (nm), P (bar)/T
(◦ C), Mears (− )/Thiele Modulus (− ), TOS (h)/GHSV (SL.kg− 1.h− 1),
5. Future opportunities particle size (μm)/average metal size (nm)) were explored to investigate
their impacts upon CO2 conversion and selectivity (CO, CH4, olefin and
Reflecting the results achieved from this work, there are a few areas paraffin, with i carbon numbers varies from 2 to 10). Albeit the

1038
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

responding patterns of the conversion and selectivity in responding to [16] Sun Y, Jia Z, Yang G, Zhang L, Sun Z. Fischer-Tropsch synthesis using iron based
catalyst in a microchannel reactor: performance evaluation and kinetic modeling.
the parametric variations mostly intrinsically counter-interact with each
Int J Hydrogen Energy 2017;42(49):29222–35.
other, even with some baffling results, the predictions obtained from [17] Sun Y, Yang G, Zhang L, Sun Z. Fischer-Tropsch synthesis in a microchannel
ANNs offers a practical guidance in adroitly manipulating the specific reactor using mesoporous silica supported bimetallic Co-Ni catalyst: process
selectivity or conversion. optimization and kinetic modeling. Chem Eng Process 2017;119:44–61.
[18] Sage V, Sun Y, Hazewinkel P, Bhatelia T, Braconnier L, Tang L, Chiang K,
Batten M, Burke N. Modified product selectivity in Fischer-Tropsch synthesis by
catalyst pre-treatment. Fuel Process Technol 2017;167:183–92.
Declaration of competing interest [19] Jiang YJ, Wang KZ, Wang Y, Liu ZH, Gao XH, Zhang JL, Ma QX, Fan SB, Zhao TS,
Yao M. Recent advances in thermocatalytic hydrogenation of carbon dioxide to
The authors declare that they have no known competing financial light olefins and liquid fuels via modified Fischer-Tropsch pathway. J CO2 Util
2023:67.
interests or personal relationships that could have appeared to influence [20] Chung W, Jeong W, Lee J, Kim J, Roh K, Lee JH. Electrification of CO2 conversion
the work reported in this paper. into chemicals and fuels: gaps and opportunities in process systems engineering.
Comput Chem Eng 2023:170.
[21] Gao JJ, Shiong SCS, Liu Y. Reduction of CO2 to chemicals and fuels:
Acknowledgements thermocatalysis versus electrocatalysis. Chem Eng J 2023:472.
[22] Over H. Surface chemistry of ruthenium dioxide in heterogeneous catalysis and
This work was funded by commercial project by Beijing Zhong Dian electrocatalysis: from fundamental to applied research. Chem Rev 2012;112(6):
3356–426.
Hua Yuan Environment Protection Technology Co., Ltd. [23] Sen R, Goeppert A, Prakash GKS. Homogeneous hydrogenation of CO2 and CO to
(E01211200005), National Key R&D Program of China methanol: the renaissance of low-temperature catalysis in the context of the
(2018YFC1903500), Ningbo Science and Technology Innovation 2025 methanol economy. Angew Chem, Int Ed 2022;61(42).
[24] Fors SA, Malapit CA. Homogeneous catalysis for the conversion of CO2, CO,
Key Project (2020Z100) and Ningbo NSF key project 2023J024. The CH3OH, and CH4 to C2+Chemicals via C-C bond formation. ACS Catal 2023;13
kind help for further language polishing from Adrian Styles and Slavko (7):4231–49.
Nikolic are appreciated. All the critical comments raised from anony­ [25] Xue RT, Ge P, Xie J, Hu ZY, Wang XK, Li PQ. Controllable CO2 reduction or
hydrocarbon oxidation driven by entire solar via silver quantum dots direct
mous reviewers in significant improving the quality of manuscript are
photocatalysis. Small 2023.
highly appreciated. [26] Gao YZ, Liu SQ, Su WP. CO2-facilitated radical sequential (3+2) annulation of
1,6-enynes via cooperation of sulfinate catalysis and photocatalysis. Green Chem
2023.
Appendix A. Supplementary data [27] Li Z, Mao CL, Pei QJ, Duchesne PN, He T, Xia MK, Wang JT, Wang L, Song R,
Jelle AA, Meira DM, Ge QJ, Ghuman KK, He L, Zhang XH, Ozin GA. Engineered
Supplementary data to this article can be found online at https://doi. disorder in CO2 photocatalysis. Nat Commun 2022;13(1).
[28] Xu SS, Chansai S, Xu SJ, Stere CE, Jiao YL, Yang SH, Hardacre C, Fan XL. CO
org/10.1016/j.ijhydene.2024.02.055.
poisoning of Ru catalysts in CO2 hydrogenation under thermal and plasma
conditions: a combined kinetic and diffuse reflectance infrared fourier transform
References spectroscopy-mass spectrometry study. ACS Catal 2020;10(21):12828–40.
[29] Wang JJ, Wang XX, AlQahtani MS, Knecht SD, Bilen SG, Chu W, Song CS.
Synergetic effect of non-thermal plasma and supported cobalt catalyst in plasma-
[1] Borlaug B, Muratori M, Gilleran M, Woody D, Muston W, Canada T, Ingram A,
enhanced CO2 hydrogenation. Chem Eng J 2023:451.
Gresham H, McQueen C. Heavy-duty truck electrification and the impacts of
[30] Xu WQ, Ramirez PJ, Stacchiola D, Rodriguez JA. Synthesis of alpha-MoC1-x and
depot charging on electricity distribution systems. Nat Energy 2021;6(6):673–82.
beta-MoCy catalysts for CO2 hydrogenation by thermal carburization of Mo-oxide
[2] Wei W, Ramakrishnan S, Needell ZA, Trancik JE. Personal vehicle electrification
in hydrocarbon and hydrogen mixtures. Catal Lett 2014;144(8):1418–24.
and charging solutions for high-energy days. Nat Energy 2021;6(1):105–+.
[31] Dry ME. Recent developments in Sasol fischer-tropsch technology. Abstr Pap Am
[3] Gourich W, Chan ES, Ng WZ, Obon AA, Maran K, Ong YH, Lee CL, Tan JLY,
Chem Soc 1986;191:64. Fuel.
Song CP. Life cycle benefits of enzymatic biodiesel co-produced in palm oil mills
[32] van Helden P, Prinsloo F, Van den Berg JA, Xaba B, Erasmus W, Claeys M, van de
from sludge palm oil as renewable fuel for rural electrification. Appl Energy
Loosdrecht J. Cobalt-nickel bimetallic Fischer-Tropsch catalysts: a combined
2022:325.
theoretical and experimental approach. Catal Today 2020;342:88–98.
[4] Liu ML, Wu JT, Lang ZK, Meng XY. Long-term energy-environment-economic
[33] van de Loosdrecht JV, Ciobica IM, Gibson P, Govender NS, Moodley DJ, Saib AM,
programming under carbon neutrality target: a study on China’s regional energy
Weststrate CJ, Niemantsverdriet JW. Providing fundamental and applied insights
transition pathways and CO2 mitigation strategies. Energy Sources Part B 2023;
into fischer-tropsch catalysis: sasol-eindhoven university of technology
18(1).
collaboration. ACS Catal 2016;6(6):3840–55.
[5] Nascimento VT, Gimenes PA, Udaeta MEM, Gimenes ALV, Riboldi VB, Ji T.
[34] Eisberg N. Business Shell pulls out of US GTL project. Chem Ind-London 2014;78
Transition mapping for modern energy service provision under uncertainty: a
(1). 12-12.
case study from Brazil. Util Pol 2023:84.
[35] Medford AJ, Kunz MR, Ewing SM, Borders T, Fushimi R. Extracting knowledge
[6] Adams M. Occidental ramps up effort to acquire Anadarko Petroleum. Oil Gas J
from data through catalysis informatics. ACS Catal 2018;8(8):7403–29.
2019;117(5):25–6.
[36] Takahashi K, Takahashi L, Miyazato I, Fujima J, Tanaka Y, Uno T, Satoh H,
[7] Hough J. Top CEOS of 2023:interview with Vicki Hollub. Barron Magazine 2023:
Ohno K, Nishida M, Hirai K, Ohyama J, Nguyen TN, Nishimura S, Taniike T. The
18–20.
rise of catalyst informatics: towards catalyst genomics. ChemCatChem 2019;11
[8] Fuentes X. Proportionality analysis and disproportionate damages: Occidental
(4):1146–52.
petroleum corporation and Occidental exploration and production company v the
[37] Liu QY, Shang C, Liu ZP. Active site for CO activation in Fe-catalyzed fischer-
republic of Ecuador. J World Invest Trade 2015;16(2):315–24.
tropsch synthesis from machine learning. J Am Chem Soc 2021;143(29):
[9] Estevez R, Aguado-Deblas L, Bautista FM, Lopez-Tenllado FJ, Romero AA,
11109–20.
Luna D. A review on green hydrogen valorization by heterogeneous catalytic
[38] Motaev Kirill, Molokeev Maxim, Sultanov Bulat, Kharitontsev Vladimir,
hydrogenation of captured CO2 into value-added products. Catalysts 2022;12
Matigorov Alexey, Palianov M. Application of machine learning to
(12).
fischer–tropsch synthesis for cobalt catalysts. Ind Eng Chem Res 2023;62(48):
[10] Cordero-Lanzac T, Ramirez A, Cruz-Fernandez M, Zander HJ, Joensen F,
20658–66.
Woolass S, Meiswinkel A, Styring P, Gascon J, Olsbye U. A CO2 valorization plant
[39] Sun Y, Yang G, Sun G, Sun Z, Zhang L. Performance Study of stirred tank slurry
to produce light hydrocarbons: kinetic model, process design and life cycle
reactor and fixed-bed reactor using bimetallic Co–Ni mesoporous silica catalyst
assessment. J CO2 Util 2023:67.
for fischer–tropsch synthesis. Environ Prog Sustain Energy 2018;37(1):553–61.
[11] Iaquaniello G, Setini S, Salladini A, De Falco M. CO2 valorization through direct
[40] Han Y, Fang CY, Ji XW, Wei J, Ge QJ, Sun J. Interfacing with carbonaceous
methanation of flue gas and renewable hydrogen: a technical and economic
potassium promoters boosts catalytic CO2 hydrogenation of iron. ACS Catal 2020;
assessment. Int J Hydrogen Energy 2018;43(36):17069–81.
10(20):12098–108.
[12] Sun Y, Wang YX, He J, Yusuf A, Wang YS, Yang G, X X. Comprehensive kinetic
[41] Zhang C, Xu MJ, Yang ZX, Zhu MH, Gao J, Han YF. Uncovering the electronic
model for acetylene pretreated mesoporous silica supported bimetallic Co-Ni
effects of zinc on the structure of Fe5C2-ZnO catalysts for CO2 hydrogenation to
catalyst during Fischer-Trospch synthesis. Chem Eng Sci 2021;246:116828–44.
linear alpha-olefins. Appl Catal B Environ 2021:295.
[13] Sun Y, Lin Z, Peng SH, Sage V, Sun Z. A critical perspective on CO2 conversions
[42] Xu Y, Zhai P, Deng YC, Xie JL, Liu X, Wang S, Ma D. Highly selective olefin
into chemicals and fuels. J Nanosci Nanotechnol 2019;19:1–13.
production from CO(2)Hydrogenation on iron catalysts: a subtle synergy between
[14] Sun Y, Wei J, Zhang JP, Yang G. Optimization using response surface
manganese and sodium additives. Angew Chem, Int Ed 2020;59(48):21736–44.
methodology and kinetic study of Fischer-Tropsch synthesis using SiO2 supported
[43] Wang JJ, You ZY, Zhang QH, Deng WP, Wang Y. Synthesis of lower olefins by
bimetallic Co-Ni catalyst. J Nat Gas Sci Eng 2016;28:173–83.
hydrogenation of carbon dioxide over supported iron catalysts. Catal Today 2013;
[15] Sun Y, Yang G, Zhang L, Sun Z. Fischer-Trospch synthesis using iron-based
215:186–93.
catalyst in a microchannel reactor: hybrid lump kinetic with ANNs/RSM. Chem
Eng Process 2017;122:181–9.

1039
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

[44] Zhang JL, Lu SP, Su XJ, Fan SB, Ma QX, Zhao TS. Selective formation of light [71] Schmidt C, Kureti S. CO2 conversion by fischer-tropsch synthesis using Na-
olefins from CO2 hydrogenation over Fe-Zn-K catalysts. J CO2 Util 2015;12: modified Fe catalysts. Chem-Ing-Tech 2022;94(11):1747–55.
95–100. [72] Skrypnik AS, Petrov SA, Kondratenko VA, Yang QX, Lund H, Matvienko AA,
[45] Satthawong R, Koizumi N, Song CS, Prasassarakich P. Comparative study on CO2 Kondratenko EV. Descriptors affecting methane selectivity in CO2 hydrogenation
hydrogenation to higher hydrocarbons over Fe-based bimetallic catalysts. Top over unpromoted bulk iron(III)-Based catalysts. ACS Catal 2022;12(18):
Catal 2014;57(6–9):588–94. 11355–68.
[46] Wei J, Sun J, Wen ZY, Fang CY, Ge QJ, Xu HY. New insights into the effect of [73] Liu JH, Li B, Cao JY, Song CS, Guo XW. Effects of indium promoter on iron-based
sodium on Fe3O4-based nanocatalysts for CO2 hydrogenation to light olefins. catalysts for CO2 hydrogenation to hydrocarbons. J CO2 Util 2022:65.
Catal Sci Technol 2016;6(13):4786–93. [74] Xu MJ, Cao CX, Xu J. Understanding kinetically interplaying reverse water-gas
[47] Jiang JD, Wen CY, Tian ZP, Wang YC, Zhai YP, Chen LG, Li YP, Liu QY, Wang CG, shift and Fischer-Tropsch synthesis during CO2 hydrogenation over Fe-based
Ma LL. Manganese-Promoted Fe3O4 microsphere for efficient conversion of CO2 catalysts. Appl Catal a-Gen 2022:641.
to light olefins. Ind Eng Chem Res 2020;59(5):2155–62. [75] Brubach L, Hodonj D, Biffar L, Pfeifer P. Detailed kinetic modeling of CO2-based
[48] Liang BL, Duan HM, Sun T, Ma JG, Liu X, Xu JH, Su X, Huang YQ, Zhang T. Effect fischer-tropsch synthesis. Catalysts 2022;12(6).
of Na promoter on Fe-based catalyst for CO2 hydrogenation to alkenes. Acs [76] Panzone C, Philippe R, Nikitine C, Vanoye L, Bengaouer A, Chappaz A,
Sustain Chem Eng 2019;7(1):925–32. Fongarland P. Catalytic and kinetic study of the CO2 hydrogenation reaction over
[49] Sun Y, Yang G, Wen C, Zhang L, Sun Z. Artificial neural networks with response a Fe-K/Al2O3 catalyst toward liquid and gaseous hydrocarbon production. Ind
surface methodology for optimization of selective CO2 hydrogenation using K- Eng Chem Res 2021;60(46):16635–52.
promoted iron catalyst in a microchannel reactor. J CO2 Util 2018;23:10–21. [77] Liu B, Geng SS, Zheng J, Jia XL, Jiang F, Liu XH. Unravelling the new roles of Na
[50] Numpilai T, Witoon T, Chanlek N, Limphirat W, Bonura G, Chareonpanich M, and Mn promoter in CO2 hydrogenation over Fe3O4-based catalysts for enhanced
Limtrakul J. Structure activity relationships of Fe-Co/K-Al2O3 catalysts calcined selectivity to light -olefins. ChemCatChem 2018;10(20):4718–32.
at different temperatures for CO2 hydrogenation to light olefins. Appl Catal a-Gen [78] Liang BL, Sun T, Ma JG, Duan HM, Li L, Yang XL, Zhang YR, Su X, Huang YQ,
2017;547:219–29. Zhang T. Mn decorated Na/Fe catalysts for CO2 hydrogenation to light olefins.
[51] Al-Dossary M, Ismail AA, Fierro JLG, Bouzid H, Al-Sayari SA. Effect of Mn loading Catal Sci Technol 2019;9(2):456–64.
onto MnFeO nanocomposites for the CO2 hydrogenation reaction. Appl Catal B [79] Zhang ZZ, Wei CY, Jia LY, Liu YY, Sun C, Wang P, Tu WF. Insights into the
Environ 2015;165:651–60. regulation of FeNa catalysts modified by Mn promoter and their tuning effect on
[52] Chew LM, Kangvansura P, Ruland H, Schulte HJ, Somsen C, Xia W, Eggeler G, the hydrogenation of CO2 to light olefins. J Catal 2020;390:12–22.
Worayingyong A, Muhler M. Effect of nitrogen doping on the reducibility, activity [80] Ding J, Huang L, Gong WB, Fan MH, Zhong Q, Russell AG, Gu H, Zhang HJ,
and selectivity of carbon nanotube-supported iron catalysts applied in CO2 Zhang YL, Ye RP. CO2 hydrogenation to light olefins with high-performance
hydrogenation. Appl Catal a-Gen 2014;482:163–70. Fe0.30Co0.15Zr0.45K0.10O1.63. J Catal 2019;377:224–32.
[53] You ZY, Deng WP, Zhang QH, Wang Y. Hydrogenation of carbon dioxide to light [81] Frenkel DSB. Understanding molecular simulation: from algorithms to
olefins over non-supported iron catalyst. Chin J Catal 2013;34(5):956–63. applications. San Diego: Academic Press; 2002.
[54] Fedorov A, Lund H, Kondratenko VA, Kondratenko EV, Linke D. Elucidating [82] Mears DE. Tests for transport limitations in experimental catalytic reactors. Ind
reaction pathways occurring in CO2 hydrogenation over Fe-based catalysts. Appl Eng Chem Process Des Dev 1971;10(4):541–&.
Catal B Environ 2023:328. [83] Sun Y, He J, Wang YS, Yang G, Sun GZ, Sage V. Experimental and CFD study of
[55] Li YB, Zeng L, Pang G, Wei XE, Wang MH, Cheng K, Kang JC, Serra JM, H2S oxidation by activated carbon prepared from cotton pulp black liquor.
Zhang QH, Wang Y. Direct conversion of carbon dioxide into liquid fuels and Process Saf Environ 2020;134:131–9.
chemicals by coupling green hydrogen at high temperature. Appl Catal B Environ [84] Fogler HS. Elements of chemical reaction engineering. New Jersey: Prentice-Hall;
2023:324. 2004.
[56] Singh G, Khurana D, Khan TS, Ghosh IK, Chowdhury B, Khodakov AY, Bordoloi A. [85] Sun Y, Yang G, Sun GZ, Sun Z, Zhang L. Performance study of stirred tank slurry
Insight into Mn enhanced short-chain olefins selectivity in CO2 hydrogenation reactor and fixed-bed reactor using bimetallic Co-Ni mesoporous silica catalyst
over Na-CuFeO2 catalyst. Appl Surf Sci 2023:616. for Fischer-Tropsch synthesis. Environ Prog Sustain Energy 2018;37(1):553–61.
[57] Chen BY, Dobele G, Plavniece A, Volperts A, Tamasauskaite-Tamasiunaite L, [86] Portha JF, Allain F, Coupard V, Dandeu A, Girot E, Schaer E, Falk L. Simulation
Norkus E, Chen CL, Lin YC. Catalytic hydrogenation of CO2 to light olefins by and kinetic study of transesterification of triolein to biodiesel using modular
using K-doped FeCx catalysts derived from the Fe-chitosan complex. Int J reactors. Chem Eng J 2012;207:285–98.
Hydrogen Energy 2023;48(11):4276–86. [87] Wang YX, Tang M, Ling J, Wang Y, Liu Y, Jin H, He J, Sun Y. Modeling
[58] Brubach L, Trutzler D, Hodonj D, Pfeifer P. Influence of recycle operation on the biohydrogen production using different data driven approaches. Int J Hydrogen
catalytic hydrogenation of CO2 to long-chain hydrocarbons. Ind Eng Chem Res Energy 2021;46(58):29822–33.
2022. [88] Sun Y, Hu J, Yusuf A, Wang YX, Jin H, Zhang XY, Liu YY, Wang YS, Yang G, He J.
[59] Wen CY, Xu XL, Song XB, Lu LY, Zhuang XZ, Jin K, Jiang Q, Zhang XH, Chen LG, A critical review on microbial degradation of petroleum-based plastics:
Wang CG, Ma LL. Selective CO2 hydrogenation to light aromatics over the Cu- quantitatively effects of chemical addition in cultivation media on biodegradation
modified Fe-Based/ZSM-5 catalyst system. Energy Fuel 2023;37(1):518–28. efficiency. Biodegradation 2022;33(1):1–16.
[60] Wang W, Toshcheva E, Ramirez A, Shterk G, Ahmad R, Caglayan M, Cerrillo JL, [89] Wang YX, Hu J, Zhang XY, Yusuf A, Qi BB, Jin H, et al. Kinetic study of product
Dokania A, Clancy G, Shoinkhorova TB, Hijazi N, Cavallo L, Gascon J. Bimetallic distribution using various data-driven and statistical models for fischer–tropsch
Fe-Co catalysts for the one step selective hydrogenation of CO2 to liquid synthesis. ACS Omega 2021;6(41):27183–99.
hydrocarbons. Catal Sci Technol 2023;13(5):1527–40. [90] Liu YY, Liu JZ, He HZ, Yang SR, Wang YX, Hu J, Jin H, Cui TX, Yang G, Sun Y.
[61] Yuan F, Zhang GH, Zhu J, Ding FS, Zhang AF, Song CS, Guo XW. Boosting light A review of enhancement of biohydrogen productions by chemical addition using
olefin selectivity in CO2 hydrogenation by adding Co to Fe catalysts within close a supervised machine learning method. Energies 2021;14(18):5916–32.
proximity. Catal Today 2021;371:142–9. [91] Riedel T, Schaub G, Jun KW, Lee KW. Kinetics of CO2 hydrogenation on a K-
[62] Wei J, Ge QJ, Yao RW, Wen ZY, Fang CY, Guo LS, Xu HY, Sun J. Directly promoted Fe catalyst. Ind Eng Chem Res 2001;40(5):1355–63.
converting CO2 into a gasoline fuel. Nat Commun 2017;8. [92] Tavasoli A, Abbaslou RMM, Dalai AK. Deactivation behavior of ruthenium
[63] Jiang F, Liu B, Geng SS, Xu YB, Liu XH. Hydrogenation of CO2 into hydrocarbons: promoted Co/-γ-AlO catalysts in Fischer-Tropsch synthesis. Appl Catal a-Gen
enhanced catalytic activity over Fe-based Fischer-Tropsch catalysts. Catal Sci 2008;346(1–2):58–64.
Technol 2018;8(16):4097–107. [93] Davis BH. Fischer-Tropsch Synthesis: reaction mechanisms for iron catalysts.
[64] To DT, Juan JC, Tsai MH, Wang CH, Pao CW, Chen CL, Lin YC. Conversion of CO2 Catal Today 2009;141(1–2):25–33.
to light hydrocarbons by using FeCx catalysts derived from iron nitrate Co- [94] Schulz H. Principles of Fischer-Tropsch synthesis-Constraints on essential
pyrolyzing with melamine, bulk g-C3N4, or defective g-C3N4. Catal Surv Asia reactions ruling FT-selectivity. Catal Today 2013;214:140–51.
2023;27(3):260–9. [95] Schweicher J, Bundhoo A, Kruse N. Hydrocarbon chain lengthening in catalytic
[65] Liu Q, Ding J, Wang RN, Zhong Q. FeZnK/SAPO-34 catalyst for efficient CO hydrogenation: evidence for a CO-insertion mechanism. J Am Chem Soc 2012;
conversion of CO2 to light olefins. Catal Lett 2023;153(1):54–61. 134(39):16135–8.
[66] Witoon T, Numpilai T, Nueangnoraj K, Cheng CK, Chareonpanich M, Limtrakul J. [96] Maitlis PM. Fischer-Tropsch, organometallics, and other friends. J Organomet
Light olefins synthesis from CO2 hydrogenation over mixed Fe-Co-K supported on Chem 2004;689(24):4366–74.
micro-mesoporous carbon catalysts. Int J Hydrogen Energy 2022;47(100): [97] Overett MJ, Hill RO, Moss JR. Organometallic chemistry and surface science:
42185–99. mechanistic models for the Fischer-Tropsch synthesis. Coord Chem Rev 2000;
[67] Ye DP, Tang WX, Zhang T, Lv L, Zou ZP, Gupta RK, Tang SW. Enhancing the 206:581–605.
synergism of Fe(3)O(4)and Fe5C2 to improve the process of CO2 hydrogenation [98] Ciobîca IM, Kramer GJ, Ge Q, Neurock M, van Santen RA. Mechanisms for chain
to olefines. Colloid Surface 2022:654. growth in Fischer-Tropsch synthesis over Ru(0001). J Catal 2002;212(2):136–44.
[68] Liang Jie, Wang Xin-yu, Xin-hua GAO, Ju-mei TIAN, Bin Duan WZ, Jiang Yong- [99] Loveless BT, Buda C, Neurock M, Iglesia E. CO chemisorption and dissociation at
jun, Reubroycharoen Prasert, Zhang Jian-li, Zhao T-s. Effect of Na promoter and high coverages during CO hydrogenation on Ru catalysts. J Am Chem Soc 2013;
reducing atmosphere on phase evolution of Fe-based catalyst and its CO2 135(16):6107–21.
hydrogenation performance. J Fuel Chem Technol 2022;50(12):1573–80. [100] Chai JC, Jiang JD, Gong Y, Wu P, Wang AN, Zhang XB, Wang T, Meng XK, Lin Q,
[69] Orege JI, Wei J, Han Y, Yang M, Sun XT, Zhang JX, Amoo CC, Ge QJ, Sun J. Lv YJ, Men Z, Wang P. Recent mechanistic understanding of fischer-tropsch
Highly stable Sr and Na co-decorated Fe catalyst for high-valued olefin synthesis synthesis on Fe-carbide. Catalysts 2023;13(7).
from CO2 hydrogenation. Appl Catal B Environ 2022:316. [101] Kuipers EW, Vinkenburg IH, Oosterbeek H. Chain-length dependence of alpha-
[70] Si ZY, Wang LK, Han Y, Yu JF, Ge QJ, Zeng CY, Sun J. Synthesis of alkene and olefin readsorption in fischer-tropsch synthesis. J Catal 1995;152(1):137–46.
ethanol in CO2 hydrogenation on a highly active sputtering CuNaFe catalyst. Acs [102] Schulz H. Confinements on growth sites of Fischer-Tropsch synthesis: manifesting
Sustain Chem Eng 2022. in hydrocarbon chain branching, vol. 258. Abstr Pap Am Chem S; 2019.

1040
J. Hu et al. International Journal of Hydrogen Energy 59 (2024) 1023–1041

[103] Schulz H. Confinements on growth sites of Fischer-Tropsch synthesis, manifesting [117] Jones VK, Neubauer LR, Bartholomew CH. Effects of crystallite size and support
in hydrocarbon-chain branching - nature of growth site and growth-reaction. Appl on the Co hydrogenation activity selectivity properties of Fe carbon. J Phys Chem-
Catal a-Gen 2020:602. Us 1986;90(20):4832–9.
[104] Han SJ, Hwang SM, Park HG, Zhang CD, Jun KW, Kim SK. Identification of active [118] Watanabe M, Sei H, Stonehart P. A breakthrough to the crystallite size effect of
sites for CO2 hydrogenation in Fe catalysts by first-principles microkinetic platinum catalysts supported on carbon. J Electrochem Soc 1988;135(3). C157-
modelling (vol 8, pg 13014, 2020). J Mater Chem A 2020;8(35). 18385-18385. C157.
[105] Nie XW, Wang HZ, Janik MJ, Guo XW, Song CS. Computational investigation of [119] Wang H, Nie X, Chen Y, Guo X, Song C. Facet effect on CO2 adsorption,
Fe-Cu bimetallic catalysts for CO2 hydrogenation. J Phys Chem C 2016;120(17): dissociation and hydrogenation over Fe catalysts: insight from DFT. J CO2 Util
9364–73. 2018;26:160–70.
[106] Teng BT, Chang J, Zhang CH, Cao DB, Yang J, Liu Y, Guo XH, Xiang HW, Li YW. [120] Rommens KT, Saeys M. Molecular views on fischer-tropsch synthesis. Chem Rev
A comprehensive kinetics model of Fischer-Tropsch synthesis over an industrial 2023;123(9):5798–858.
Fe-Mn catalyst. Appl Catal a-Gen 2006;301(1):39–50. [121] Bukur DB, Todic B, Elbashir N. Role of water-gas-shift reaction in Fischer-Tropsch
[107] Mousavi S, Zamaniyan A, Irani M, Rashidzadeh M. Generalized kinetic model for synthesis on iron catalysts: a review. Catal Today 2016;275:66–75.
iron and cobalt based Fischer-Tropsch synthesis catalysts: review and model [122] Peter M, Maitlis AdK. Greener fischer-tropsch processes for fuels and feedstocks.
evaluation. Appl Catal a-Gen 2015;506:57–66. London: Wiley-VCH; 2013. p. 390.
[108] Elbashir NO, Chatla A, Lemonidou A, Spivey JJ. Reaction engineering & catalysis [123] Yang J, Ma WP, Chen D, Holmen A, Davis BH. Fischer-Tropsch synthesis: a review
issue in honor of professor dragomir bukur: introduction and review. Catal Today of the effect of CO conversion on methane selectivity. Appl Catal a-Gen 2014;470:
2020;343:1–7. 250–60.
[109] Panzone C, Philippe R, Chappaz A, Fongarland P, Bengaouer A. Power-to-Liquid [124] Yan F, Qian WX, Sun QW, Zhang HT, Ying WY, Fang DY. Product distributions
catalytic CO2 valorization into fuels and chemicals: focus on the Fischer-Tropsch and olefin-to-paraffin ratio over an iron-based catalyst for Fischer-Tropsch
route. J CO2 Util 2020;38:314–47. synthesis. React Kinet Mech Catal 2014;113(2):471–85.
[110] Sarup B, Wojciechowski BW. Studies of the Fischer− Tropsch synthesis on a cobalt [125] Melaet G, Lindeman AE, Somorjai GA. Cobalt particle size effects in the fischer-
catalyst II. Kinetics of carbon monoxide conversion to methane and to higher tropsch synthesis and in the hydrogenation of CO2 studied with nanoparticle
hydrocarbons. Can J Chem Eng 1989;67:62–74. model catalysts on silica. Top Catal 2014;57(6–9):500–7.
[111] Yates IC, Satterfield CN. Intrinsic kinetics of the Fischer− Tropsch synthesis on a [126] Patanou E, Tsakoumis NE, Myrstad R, Blekkan EA. The impact of sequential H-2-
cobalt catalyst. Energy Fuel 1991;5:168–73. CO-H-2 activation treatment on the structure and performance of cobalt based
[112] Botes FG, Breman BB. Development and testing of a new macro kinetic expression catalysts for the Fischer-Tropsch synthesis. Appl Catal a-Gen 2018;549:280–8.
for the iron-based low-temperature fischer-tropsch reaction. Ind Eng Chem Res [127] Banaresalcantara R, Ko EI, Westerberg AW, Rychener MD. Decade - a hybrid
2006;45(22):7415–26. expert system for catalyst selection .2. Final architecture and results. Comput
[113] Ojeda M, Nabar R, Nilekar AU, Ishikawa A, Mavrikakis M, Iglesia E. CO activation Chem Eng 1988;12(9–10):923–38.
pathways and the mechanism of fischer-tropsch synthesis. J Catal 2010;272: [128] Banaresalcantara R, Westerberg AW, Ko EI, Rychener MD. Decade - a hybrid
287–97. expert system for catalyst selection .1. Expert system consideration. Comput
[114] Ojeda M, Nabar R, Nilekar AU, Ishikawa A, Mavrikakis M, Iglesia E, Catal J. CO Chem Eng 1987;11(3):265–77.
activation pathways and the mechanism of Fischer-Tropsch synthesis. J Catal [129] Mai HX, Le TC, Chen DH, Winkler DA, Caruso RA. Machine learning for
2010;272(2):287–97. electrocatalyst and photocatalyst design and discovery. Chem Rev 2022;122(16):
[115] Botes FG, Dyk Bv, McGregor C. The development of a macro kinetic model for a 13478–515.
commercial Co/Pt/Al2O3 fischer-tropsch catalyst. Ind Eng Chem Res 2009;48 [130] Motagamwala AH, Dumesic JA. Microkinetic modeling: a tool for rational catalyst
(23):10439–47. design. Chem Rev 2021;121(2):1049–76.
[116] Jung HJ, Walker PL, Vannice MA. Co hydrogenation over well-dispersed carbon- [131] Yan TX, Chen XY, Kumari L, Lin JL, Li ML, Fan Q, Chi HY, Meyer TJ, Zhang S,
supported iron catalysts. J Catal 1982;75(2):416–22. Ma XB. Multiscale CO electrocatalysis to C products: reaction mechanisms,
catalyst design, and device fabrication. Chem Rev 2023;123(17):10530–83.

1041