You are on page 1of 17

Applied Thermal Engineering 229 (2023) 120581

Contents lists available at ScienceDirect

Applied Thermal Engineering


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

Selecting an optimum kinetic model for gas–solid adsorption based on


statistical approaches and a model selection criterion
Md. Matiar Rahman a, c, Abu Zar Shafiullah c, Mahua Jahan Rupa a, e,
Shamal Chandra Karmaker a, b, c, Shahadat Hosan a, b, Animesh Pal d, Bidyut Baran Saha a, b, *
a
International Institute for Carbon-Neutral Energy Research, Kyushu University, 744 Motooka, Nishi-ku, Fukuoka 819-0395, Japan
b
Mechanical Engineering Department, Kyushu University, 744 Motooka, Nishi-ku, Fukuoka 819-0395, Japan
c
Department of Statistics, University of Dhaka, Dhaka-1000, Bangladesh
d
Department of Nuclear Engineering, University of Dhaka, Dhaka-1000, Bangladesh
e
Department of Mathematics, University of Barishal, Barishal-8254, Bangladesh

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

Keywords: Adsorption heat transformation (AHT) techniques can help to safeguard the environment by reducing the use of
Adsorption fossil fuels. When constructing an adsorption-based dynamic heat pump system, analyzing the kinetic charac­
Bootstrap teristics of the adsorbate–adsorbent pair is important. The recently established time-adapted linear driving force
Information criteria
(LDF) and commonly used models such as LDF, modified LDF, and Fickian diffusion (FD) models are available in
Kinetic model
Statistical tool
the literature for analyzing the kinetic characteristics, making it challenging for researchers to select the opti­
Time-adapted LDF mum one for simulating the complete real system. This study applies a non-parametric statistical approach
known as bootstrapping for simulating the replica of experimental kinetics data for all studied pairs. The kinetics
models mentioned above are used to correlate the bootstrap samples. The estimated values of the statistical
parameters are compared, and the optimum kinetics model for gas–solid physical adsorption is proposed. The
parameters of the model are estimated using the generalized reduced gradient (GRG) non-linear optimization
method. For each bootstrap sample, the optimum model is selected based on the minimum values, which means
less information loss, of a set of information-based model selection criteria (ICs). The time-adapted LDF incurs
smaller values of IC compared to other kinetics models. The p-value of the proportion test for water adsorption
onto AQSOA-Z01, RD silica gel, and Al-Fum is very small, which is less than 0.01. So, the test is significant at a
1% level, which implies that time adapted LDF model is significantly optimum. This is the first time that a
rigorous statistical optimization technique has been used to identify the optimum kinetic model for gas–solid
adsorption. The current findings are critical for a comprehensive analysis of an adsorption system and its design.

isotherm and kinetic models can be used to explore the adsorption


properties. The isotherm model identifies the equilibrium absorption
1. Introduction uptake at a particular temperature, while the kinetic model determines
the cycle time of the process. The kinetic model estimates the adsorption
Adsorption heat transformation (AHT) systems, for instance, heating or desorption cycle time required to complete processes by determining
and cooling, that can be powered by waste heat or renewable-resource the rate at which vapor diffuses into the pores or removes from the
and operate at temperatures as low as 60 ◦ C, are attracting a lot of pores. As a result, in order to simulate and develop a realistic adsorption
attention from scientists as a means to reduce world electricity demand cooling system, an appropriate model for calculating adsorption kinetics
[1]. It is also employed in the following applications: gas storage/cap­ is required [8]. Several authors have estimated the kinetic parameters of
ture [2], purification [3], gas separation [4], desalination [5], heavy various working pairs utilized in the sorption process, and several ki­
metal/liquid removal [6], and so on. The characteristics of the netics models were employed to analyze the experimental data [9-19].
adsorbate-adsorbent combinations determine the efficiency of adsorp­ Table 1 shows the studied different kinetics models and their
tion systems. The most significant factors in choosing a pair are their characteristics.
physical, chemical, and thermodynamic characteristics [7]. Adsorption

* Corresponding author at: International Institute for Carbon-Neutral Energy Research, Kyushu University, 744 Motooka, Nishi-ku, Fukuoka 819-0395, Japan.
E-mail address: saha.baran.bidyut.213@m.kyushu-u.ac.jp (B.B. Saha).

https://doi.org/10.1016/j.applthermaleng.2023.120581
Received 14 February 2022; Received in revised form 2 March 2023; Accepted 13 April 2023
Available online 18 April 2023
1359-4311/© 2023 Elsevier Ltd. All rights reserved.
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

transfer. The same author [36] developed mathematical models for a


Nomenclature batch process to forecast concentration distributions for an active in­
gredient’s (vancomycin) adsorption on a typical hydrophobic-molecule
a Fitting parameters of time-adapted LDF model (s− 2) adsorbent. A heuristic algorithm was used to estimate the kinetic pa­
b Fitting parameters of time-adapted LDF model (s− 1) rameters by maximizing the coefficient of determination.
c Fitting parameters of time-adapted LDF model (-) However, the LDF model [37,38] is the most extensively used for
Ds Diffusion coefficient [m2/s]. explaining gas–solid sorption dynamics. At constant pressure, the LDF
Ei Expected frequency model undertakes that the particle temperature of adsorbent is always
F0 Shape factor of the particle similar. Several researchers updated the LDF model based on the char­
H0 Null hypothesis acteristics of the adsorption pair. For example, Sultan et al. [25] intro­
H1 Alternative hypothesis duced a dimensionless power ‘m’ as a variable. The LDF model was used
n1 Number of data points [-] by Liaw et al. [39] to study the parabolic concentration curve of a
Oi Observed frequency spherical particle. Sircar and Hufton [40] investigated a relationship
p Number of parameters in the model [-] between the diffusion time constant and the total mass transfer coeffi­
RP Adsorbent particle radius [m] cient of the LDF model, removing the usual constraint between them.
Wcal Calculated uptake [kg/kg] Sircar and Hufton [28] studied at the interplay of the LDF and FD
Wexp Experimental uptake [kg/kg] models, but due to integrations, local adsorption aspects were over­
W(t) Instantaneous uptake [kg/kg] looked. Lagergreen’s pseudo-first-order model and Glueckauf’s LDF
w(∞) Equilibrium uptake [kg/kg] model were thoroughly explored by Rodrigues and Silva [41]. They
observed that both techniques were similar only in the limiting condi­
Greek Letters tion of infinite bath systems or zero adsorption. The final and initial fluid
α Level of significance phase concentrations, on the other hand, are practically similar. The LDF
χ2 Chi-square value of the overall test has some limitations; it sometimes overestimates experimental kinetics
θ fractional uptake (kg/kg) at a specific time and underestimates at the rest periods [42]. Both the
LDF and FD models are successful in predicting the long-time experi­
mental results, however, they fail to predict the adsorption’s beginning
phase of adsorption. The linearized FD model causes this occurrence; the
Pal et al. [26] examined both FD and LDF kinetic models for ethanol
interception would be (ln 6/π2), implying that there will always be a
vapor adsorption onto activated carbons made from waste palm trunk
constant value. The semi-infinite model, on the other hand, is good at
(WPT) and mangrove (M). Ibrahim et al. [24] studied FD model to
forecasting the beginning phase of adsorption. The semi-infinite kinetic
determine the adsorption kinetics of ethanol adsorption onto phenol
model is clearly imprecise in predicting the experimental data for the
resin-based adsorbents for a cooling system. The author employed two
AQSOA-Z01/water pair beyond 400 s and entirely fails to forecast the
types of KOH-treated adsorbents, KOH4-PR and KOH6-PR. Sultan et al.
behavior for the RD silica gel/water pair [33].
[25] examined FD, LDF, and semi-infinite kinetic model for water
Recently Rupa et al. [33] established a time-adapted LDF model for
adsorption onto polymer-based adsorbents. Vancomycin adsorption on
gas–solid adsorption. The time-adapted LDF model was compared to the
polymeric adsorbent was investigated by Blaz Likozar et al. [34]. Here,
other kinetic models for various types of adsorption pairs. They studied
vancomycin adsorption from broth supernatant, diluted broth, and the
AC/ethanol, AC/CO2, MOF/water, MOF/methanol, zeolite/adsorbate
whole broth was examined in terms of equilibrium and kinetics. The
(water and CO2), silica gel/water, aminated solid/CO2, immobilized
RSFK, the HSDM, and the BPM models were used to study it. The HSDM
polyethyleneimine-mesoporous silica/CO2 adsorption pair. As can be
provided the best response. Vancomycin adsorption was investigated by
seen from the above literature reviews, it is evident that many adsorp­
authors [35] using two different polymeric resins (Amberlite XAD16N
tions kinetic models have been suggested for specific pairs for gas–solid
and 1600 N). They developed a fixed-bed column model and verified it
adsorption, but there is no widely accepted model for the general
in a laboratory experiment. The consistency between the estimated and
gas–solid adsorption process. Researchers face difficulties in choosing
measured data was significantly improved by applying structural
the best kinetic model to describe their experimental findings. In almost
adsorption/desorption kinetics while accounting for the resin particle
all cases, to the best of our knowledge, the authors did not consider the
size distribution, as opposed to changing the impact of external mass
information loss of the fitted model, and they used only root mean

Table 1
Studied adsorption kinetics models and their characteristics [20].
Model name Mathematical expression Advantages Limitations Remarks Ref(s).
√̅̅̅̅̅̅̅
Semi-infinite 2A Ds t Best to calculate the initial phase Not suitable for high-pressure Only acceptable for semi-infinite medium for
[21,22]
model θ =
V π ( of adsorption ratios any particle shape
)
Fickian 6 Ds t Can predict long time region Inadequate for the initial Convenient for intra-particle diffusion
θ = 1 − 2 exp − F0 π2 2 [23-
diffusion π Rp kinetics data phase of adsorption
(FD) model 27]
)
Linear driving (
Ds t Easy to calculate, suitable for Under-predicts at the initial Often used model by the researchers
θ = 1 − exp − F0 2 [28-
force (LDF) Rp breakthrough curves stage and over-predicts at the
model saturated stage of dynamic 32]
adsorption data
Modified LDF ( ( )m ) Dimensionless parameter m has No physical significance of Very few works can be found in the literature
Ds t [25]
θ = 1 − exp − F0 been introduced to correlate the the parameter m has been utilizing this model
R2p
experimental precisely. reported
Time-adapted θ = To describe the transient Time dependency of a and b is Relatively new and gaining popularity as this
[33]
LDF model ( ( )(at2 +bt+c) ) behavior more accurately, two not well defined yet. model can predict the adsorption dynamics
Ds t
1 − exp − F0 time-dependent variables a and accurately both during the initial unsaturated
R2p
b are introduced. condition as well as during the later period
when the adsorbent is near its saturated
condition.

2
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Table 2 square deviation (RMSD) error analysis in their investigations.


Adsorption pair, data sources, and surface properties of adsorbents. The major purpose of this investigation is to use a variety of statis­
SL. Adsorption pair Data sources Surface properties of tical approaches and a model selection criterion to determine the
No. adsorbent optimal kinetic model for gas adsorption on solids. This is the first time
1 AQSOA-Z01/water Teo et al. [48] PV: 0.071 cm3/g; BET-SA: that a rigorous statistical optimization technique has been used to
189.6 m2/g, and average identify the optimum kinetic model for gas–solid adsorption. Parameter
PS: 11.78 Å. effect adjustment and information loss are considered in this investi­
gation to find the optimal kinetics for gas–solid adsorption, which also
2 RD silica gel/water Sun et al. [7], Islam PV: 0.4 cm3/g; BET-SA:
et al. [49], Chua et al. 820 m2/g, and average PS:
makes it significant from the statistical analysis of the best-fitting model.
[50] 2.2 nm. A non-linear optimization method called generalized reduced gradient
(GRG) is employed to optimize the adsorption kinetic data. The root
3 Aluminum Teo et al. [51] PV: 0.926 cm3/g; BET-SA: mean square deviation (RMSD) is employed as an error evaluation
fumarate (Al- 792.26 m2/g, and average
function. As far as we are concerned, the RMSD distribution is unknown.
Fum)/water PS: 10.91 Å.
Therefore, we must use inferential methods using non-parametric ap­
4 AQSOA-Z02/water Teo et al. [48] PV: 0.27 cm3/g; BET-SA: proaches, such as the bootstrap technique, to determine the distribution
717.8 m2/g, and average of RMSD. In the bootstrap method, 10,000 resamples from the original
PS: 11.84 Å. sample are taken, giving us a distribution of RMSD. Then this informa­
5 M− AC/ethanol Pal et al. [26] PV: 2.18 cm3/g; BET-SA:
tion is used in model selection criteria to find an optimum kinetic model.
2924 m2/g; and average Thirteen gas–solid adsorption kinetic experimental data are taken from a
PS: 1.47 nm. published paper with the aim of finding the optimum kinetic model for
6 WPT-AC/ethanol Pal et al. [26] PV: 2.51 cm3/g; BET-SA: this investigation. Among them, nine pairs are chosen from cooling
2927 m2/g; and average
systems with different refrigerants, while the remaining four are from
PS: 1.68 nm.
7 MIL-101-GP/ Solovyeva et al. [52] PV: 1.61 cm3/g; BET-SA: CO2 capture and storage applications. At first, all kinetic models are
methanol 2550 m2/g; average PS: tried to analyze the experimental data, but only the most likely optimal
not available. models, i.e., FD, LDF, and time-adapted LDF models, are chosen for
8 MIL-101-T(5)/ Solovyeva et al. [52] PV: 1.61 cm3/g; BET-SA: further examination. Statistical techniques such as bootstrap sampling,
methanol 2870 m2/g; average PS:
overall, and pairwise proportion tests were employed to select the
not available.
9 MIL-101-T(10)/ Solovyeva et al. [52] PV: 1.43 cm3/g; BET-SA: suitable kinetic model. The selection rates were determined from n =
methanol 2610 m2/g; average PS: 10000 bootstrap samples, and the statistical tests for the significance of
not available. differences in selection rates further confirmed the optimal kinetic
10 Natural zeolite tuff Ammendola et al. PV: not available; BET-SA:
model.
/CO2 [53] 141 m2/g, and average PS
range: 2–7 Å.
11 CS-H2O/CO2 Alvarez et al. [54] PV: 0.53 cm3/g; BET-SA: 2. Kinetic models
998 m2/g; and average PS:
not availabe.
The vapor diffusion rate into the pores is measured by adsorption
12 CS-C2O/CO2 Alvarez et al. [54] PV: 0.48 cm3/g; BET-SA:
1045 m2/g; and average kinetics, which aids in calculating the cycle time required to complete
PS: not availabe. the desorption/adsorption processes. Therefore, when simulating and
13 Calgon BPL/CO2 Alvarez et al. [54] PV: 0.50 cm3/g; BET-SA: designing actual adsorption chillers, an accurate model to predict the
1129 m2/g; and average adsorption kinetics is essential [8]. Numerous authors have experi­
PS: not availabe.
mentally investigated the adsorption kinetics of various working pairs
utilized in the adsorption system, and a variety of kinetics models [9-
Note: PV: Pore volume; BET-SA: BET surface area; PS: Pore size. 11,14,19] have been used to fit those data. To characterize the kinetics
of different gasses adsorbing onto various adsorbents, various models

Table 3
Fitting parameters of a time-adapted LDF model for studied adsorption pairs.
Time-adapted LDF model LDF FD

Adsorption pairs Tads [◦ C] Ds/R2p [s¡1] a [s¡2] b [s¡1] c [-] Ds/R2p [s¡1] Ds/R2p [s¡1]

AQSOA-Z01/water 60 5.42E-05 1.02E-09 − 1.5E-05 1.01 6.0E-05 4.4E-06


RD silica gel/water 30 8.33E-04 7.02E-07 − 8.6E-05 0.6 0.26 6.64E-05
Aluminum fumarate/water 60 3.29E-03 1.10E-05 − 5.90E-04 1.01 3.97E-03 2.92E-04
Aluminum fumarate/water 25 5.12E-04 1.10E-05 − 5.90E-04 1.01 4.90E-04 4.20E-05
AQSOA-Z02/water 60 6.60E-05 5.72E-09 − 6.5E-05 1.15 8.00E-05 –
AQSOA-Z02/water 45 2.75E-05 5.72E-09 − 6.5E-05 1.15 2.84E-05 –
AQSOA-Z02/water 35 1.13E-05 5.72E-09 − 6.5E-05 1.15 1.12E-05 –
AQSOA-Z02/water 25 6.95E-06 5.72E-09 − 6.5E-05 1.15 6.8E-06 –
M− AC/ethanol 50 2.71E-04 1.1E-07 − 3.1E-05 0.75 2.21E-04 1.70E-05
M− AC/ethanol 70 3.33E-04 1.0E-07 − 3.0E-05 0.75 3.23E-04 2.50E-05
WPT-AC/ethanol 30 3.96E-04 2.21E-06 − 1.1E-03 0.82 3.85E-04 3.92E-04
WPT-AC/ethanol 50 6.82E-04 2.21E-06 − 1.0E-03 0.82 6.74E-04 6.71E-04
MIL-101-GP/methanol – 1.05E-03 5.01E-07 − 4.7E-04 0.89 1.07E-03 5.51E-05
MIL-101-T(5)/methanol – 6.51E-04 5.11E-07 − 4.7E-04 0.90 6.49E-04 3.36E-05
MIL-101-T(10)/methanol – 6.58E-04 5.12E-07 − 4.7E-04 0.90 5.63E-04 3.28E-05
Natural zeolite tuff /CO2 25 3.96E-02a 4.23E-03b − 1.21E-01a 1.5 4.55E-02a 1.85E-03a
CS-H2O/CO2 30 5.81E-02a 9.74E-04b − 5.02E-03a 0.9 6.92E-02a 6.01E-03a
CS-C2O/CO2 30 5.44E-02a 9.62E-04b − 5.10E-02a 1.4 7.34E-02a 3.17E-03a
Calgon BPL/CO2 30 5.80E-02a 9.72E-04b − 5.00E-02a 0.8 5.34E-02a 2.92E-03a

Note: a [Unit] = [min− 1


] and b [Unit] = [min− 2
].

3
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 1. Adsorption kinetics of (a) AQSOA-Z01/water, (b) RD silica gel/water, and (c) Al-Fum/water pairs using RMSD errors.

4
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Table 4
Results of all ICs, p-value of proportion tests for AQSOA-Z01/water, RD silica gel/water, Al-Fum/water, and AQSOA-Z02/water pairs of probable kinetic models
employing n = 10000 bootstrap samples.
AQSOA-Z01/water

Model Error mean AIC BIC AICc mAIC AIC3 CAIC ABIC

LDF 0.0745 –223.16 − 214.78 –222.43 − 199.16 − 219.16 − 210.78 − 227.36


FD 0.0838 − 218.08 − 211.79 − 217.65 − 206.08 − 215.08 − 208.79 − 221.23
Time-adapted LDF 0.0638 ¡234.48 ¡228.20 ¡234.06 –222.48 ¡231.48 ¡225.20 ¡237.64
Overall and pairwise tests P-value
Overall 2.0E-15 2.1E-14 2.2E-15 2.0E-13 2.1E-15 2.0E-14 2.1E-15
LDF vs time-adapted LDF 2.0E-04 2.1E-05 2.2E-03 1.4E-04 1.8E-04 1.9E-03 2.1E-04
FD vs time-adapted LDF 1.5E-08 2.1E-09 1.1E-08 2.0E-07 2.0E-08 1.7E-09 2.1E-11
RD silica gel/water
LDF 0.0745 –223.16 − 214.78 –222.43 − 199.16 − 219.16 − 210.78 − 227.36
FD 0.0838 − 218.08 − 211.79 − 217.65 − 206.08 − 215.08 − 208.79 − 221.23
Time-adapted LDF 0.0638 ¡234.48 ¡228.20 ¡234.06 –222.48 ¡231.48 ¡225.20 ¡237.64
Overall and pairwise tests P-value
Overall 2.5E-11 1.9E-12 2.8E-16 2.4E-12 1.8E-13 2.4E-12 2.3E-16
LDF vs time-adapted LDF 2.1E-05 2.5E-07 1.8E-04 1.9E-03 2.8E-05 2.9E-04 2.7E-03
FD vs time-adapted LDF 1.6E-09 1.6E-10 1.7E-07 2.8E-09 2.2E-06 1.4E-12 2.7E-13
Al-Fum/water
LDF 0.0745 –223.16 − 214.78 –222.43 − 199.16 − 219.16 − 210.78 − 227.36
FD 0.0838 − 218.08 − 211.79 − 217.65 − 206.08 − 215.08 − 208.79 − 221.23
Time-adapted LDF 0.0638 ¡234.48 ¡228.20 ¡234.06 –222.48 ¡231.48 ¡225.20 ¡237.64
Overall and pairwise tests P-value
Overall 2.2E-16 2.3E-13 2.7E-14 1.0E-15 1.9E-16 2.7E-12 2.8E-13
LDF vs time-adapted LDF 2.1E-05 2.5E-04 2.9E-04 1.3E-03 2.8E-03 2.9E-04 2.8E-05
FD vs time-adapted LDF 1.6E-07 2.5E-10 2.1E-07 2.1E-08 2.7E-09 1.0E-8 2.8E-12
AQSOA-Z02/water
LDF 3.11 22.35 25.36 23.56 24.65 27.65 28.51 24.61
Time-adapted LDF 0.98 4.23 19.50 6.35 27.31 6.24 18.50 2.81
Overall test P-value
LDF vs time-adapted LDF 2.2E-03 2.3E-04 2.2E-04 1.7E-04 1.9E-05 2.8E-04 2.7E-04

*Bold values represent the smallest ICs.

are available. The pseudo-first-order equation proposed by Lagergreen


[43] is said to be the initial model considering the adsorption rate
2.2. Fickian diffusion (FD) model
dependent on the adsorption capacity. However, the linear driving force
(LDF) and Fickian diffusion (FD) model is the most often used model for
The FD model can be employed to predict the adsorption kinetics of
characterizing the gas–solid adsorption dynamics. The next paragraphs
various adsorption pairs [44]. Assuming purely radial diffusion by a
explain the LDF, FD, and recently developed time-adapted LDF models.
spherical shape adsorbent particle, the diffusion equation can be
expressed as follows [21]-
2.1. Linear driving force (LDF) model
( ) ( )
dw 1 ∂ ∂w
The LDF is a widely used model to illustrate the interaction between = 2 r 2 Ds (5)
dt r ∂r ∂r
adsorbate and adsorbent during the adsorption process. Glueckauf and
Coates were the first to present this model’s approximation [37]. The For constant diffusivity (Ds ), equation (5) simplifies to
following equation can be used to determine the rate of adsorption of an (
dw
) ( 2
∂ w 2 ∂w
)
adsorption pair. = Ds + (6)
dt ∂r2 r ∂r
(dw/dt) = Ks av (w(∞) − w(t)) (1) where, w(r, t) represents the adsorbed phase concentration.
Although the diffusivity is concentration dependent, the assumption of a
where ksav is the overall mass transfer coefficient. w(t) represents instant
constant diffusivity is still an acceptable approximation provided that
uptake [kg/kg], w(∞) represents the saturation uptake [kg/kg]. This
the uptake curve is measured over a small differential change in
equation is originally proposed by Gleuckauf [27], where the shape
adsorbed phase concentration [21]. If the uptake of sorbate by the
factor (F0) is 15 for the spherical shape of an adsorbent particle with a
adsorbent is small relative to the total quantity of sorbate introduced
homogeneous structure.
into the system, the ambient sorbate concentration will remain essen­
F0 Ds tially constant following the initial step, change, and the appropriate
Ks av = (2)
R2p initial and boundary conditions are;
( )
Rp denotes adsorbent particle radius [m], and Ds signifies the diffu­ ∂w
w(r, 0) = w(initial), w(r, t) = w(∞), =0
sion coefficient [m2/s]. Then, the equation (1) can be written as follows- ∂r r=0
F0 D s After simplifying equation (6), it can be expressed by the familiar
(dw/dt) = (w(∞) − w(t)) (3)
R2p expression-
( )
After simplifying equation (3), the fractional uptake (θ) can be pre­ w(t) − w(initial) 6 ∑∞
1 2 2 Ds t
θ= = 1− 2 exp − F n π (7)
sented as follows- w(∞) − w(initial) π n=1 n2 0
R2p
( )
w(t) − w(initial) Ds t
θ= = 1 − exp − F0 2 (4) where Rp is the adsorbent particle radius and w(t) is the average con­
w(∞) − w(initial) Rp
centration through the adsorbent particle, defined by-

5
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 2. Selection rate of (a) AQSOA-Z01/water, (b) RD silica gel/water, and (c) Al-Fum/water pairs using RMSD errors.

6
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 3. Adsorption kinetics of (a) M− AC/ethanol pair for 50 ◦ C; and (b) WPT-AC/ethanol pair for 30 ◦ C using RMSD errors.


3 r
the uptake at short times [42]. Based on this error, Rupa et al. [33]
w(t) = wr2 dr
r3 0 proposed to modify the LDF with a time-based correction factor equation
(at2 + bt + c) to the LDF model. As a result, the time-adapted LDF Model
The equation (7) gives faster convergence over the long time period,
[33] is presented, and it can be stated as-
whereas the higher terms of this summation go to zero; consequently,
⎛ ( )(at2 +bt+c) ⎞
the FD model can be presented as follows [21]-
w(t) − w(initial) Ds t
( ) θ= = 1 − exp⎝ − F0 ⎠ (9)
6 Ds t w(∞) − w(initial) R2p
θ = 1 − 2 exp − F0 π 2 2 (8)
π Rp
Here a, b, and c represent fitting parameters. The unit of a is s− 2 or
min− 2, the unit of b is s− 1 or min− 1, and c is a unit less. The correction
factor is chosen to be a quadratic function because the quadratic func­
2.3. Time-adapted LDF model
tion helps to find out the curves which are in a parabola or U-shape and
varies the curve in width or steepness. The values of a, b, and c are fixed
When the adsorbent particle’s temperature is uniform, and the heat
for the same working pair. These values are specified after fitting the
transfer rate is ignored, the LDF equation is commonly employed [45]. It
experimental data with various configurations (thickness and sizes of
has a wide range of applications to describe sorption kinetics at non-
adsorbents), different measurement methods (LTJ and LPJ), and
equilibrium conditions and aids in understanding the fundamental
different adsorption temperatures and pressures.
concept of the adsorption process that ranges from transient to cyclic
steady-state [3346,47]. However, the LDF model has several flaws; it
sometimes overestimates the uptake at long times and underestimates

7
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 4. Selection rate of (a) M− AC/ethanol, and (b) WPT-AC/ethanol adsorption pairs using RMSD errors.

Table 5
Results of all ICs, p-value of proportion tests for M− AC/ethanol, and WPT-AC/ethanol pairs of probable kinetic models employing n = 10000 bootstrap samples.
M¡AC/ethanol

Model Error mean AIC BIC AICc mAIC AIC3 CAIC ABIC

LDF 0.061 − 3055.93 − 3051.72 − 3055.92 − 3055.93 − 3054.93 − 3050.72 − 3054.90


FD 0.116 − 2737.14 − 2732.93 − 2737.13 − 2737.14 − 2736.14 − 2731.93 − 2736.11
Time-adapted LDF 0.015 – ¡3728.90 ¡3745.65 ¡3721.73 ¡3741.73 ¡3724.90 ¡3741.60
Overall and pairwise tests P-value 3745.73
Overall 1.8E-12 1.8E-13 1.2E-13 1.5E-14 3.1E-11 1.0E-13 1.6E-14
LDF vs time-adapted LDF 2.2E-03 2.6E-04 1.2E-07 1.9E-06 1.7E-07 1.4E-08 2.7E-05
FD vs time-adapted LDF 2.5E-07 1.1E-12 1.5E-11 1.6E-10 1.4E-9 1.4E-09 2.1E-12
WPT-AC/ethanol
LDF 0.061 − 1952.2 − 1948.37 − 1952.19 − 1952.2 − 1951.2 − 1947.37 − 1951.55
FD 0.127 − 1704.34 − 1700.51 − 1704.33 − 1704.34 − 1703.34 − 1699.51 − 1703.69
Time-adapted LDF 0.023 – ¡2260.58 ¡2275.76 ¡2251.88 ¡2271.88 ¡2256.58 ¡2273.27
Overall and pairwise tests P-value 2275.88
Overall 1.96E-13 2.35E-12 1.61E-14 1.5E-16 2.1E-15 1.5E-14 1.9E-13
LDF vs time-adapted LDF 2.6E-04 2.4E-07 1.7E-09 2.8E-08 1.9E-08 1.3E-07 2.7E-10
FD vs time-adapted LDF 1.8E-06 1.8E-11 1.4E-13 1.6E-15 1.4E-12 1.8E-08 2.6E-13

*Bold values represent the smallest ICs.

8
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 5. Adsorption kinetics of (a) MIL-101-GP/methanol and (b) MIL-101-T(5)/methanol pairs using RMSD errors.

2.4. Experimental data model selection criteria. Original experimental data, however, allow
computing a single measurement of RSS, and this, in turn, computes a
In this study, experimental data extracted from thirteen widely used single value of the information criteria for each model. Multiple esti­
gas–solid adsorption pairs are analyzed to find the best-fitted kinetic mates of RSS and IC can be obtained by replicating the experiment. Since
model for each pair. Among them, nine pairs (SL. No 1–9 of Table 2) are the actual experiment was not replicated due to time and cost con­
chosen from cooling systems with different refrigerants, while the straints, a simulation-based approach, i.e., bootstrapping (Section 3.3),
remaining four are from CO2 capture and storage applications. The was applied to generate samples from the existing data. Bootstrap
adsorption pair, data sources, and surface properties of adsorbents are samples allow to estimate the RSS of a given model multiple times under
all listed in Table 2. the same experimental setup. This enables us to compare the informa­
tion loss of the candidate models from each bootstrap sample. Since the
3. Statistical tools experiment was operated under regulated conditions, the outcome and
simulated data were expected to signify the true outcome under the
The statistical techniques used to choose the optimum kinetic model treatment combinations. An empirical approximation of the rate at
(s) for gas–solid physical adsorption are described in this section. A set of which a model is preferred over alternative models is provided by the
model selection criteria outlined in Section 3.2 are used to compare percentage of samples where a specific model meets the minimal value
these models. The residual sum of squares (RSS) for the candidate of a selection criterion. The probability of each model is compared
models, i.e., the sum of the squared discrepancies between the observed independently based on each criterion, and tests are conducted to find
and estimated outcome (Y = uptake), is required for computing the evidence that specific model(s) have a higher selection rate (more likely

9
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Table 6
Results of all ICs, p-value of proportion tests for MIL-101-GP/methanol, MIL-101-T (5)/methanol, MIL-101-T(10)/methanol pairs of probable kinetic models
employing n = 10000 bootstrap samples.
MIL-101-GP/methanol

Model Error mean AIC BIC AICc mAIC AIC3 CAIC ABIC

LDF 0.039 − 2299.45 − 2297.22 − 2299.39 − 2299.45 − 2298.45 − 2296.22 − 2300.37


FD 0.130 − 1702.28 − 1700.05 − 1702.22 − 1702.28 − 1701.28 − 1699.05 − 1703.20
Time-adapted LDF 0.035 – ¡2338.19 ¡2346.50 –2323.12 ¡2343.12 ¡2334.19 ¡2350.79
Overall and pairwise tests P-value 2347.12
Overall 2.8E-15 2.1E-13 1.9E-13 1.8E-15 3.2E-11 1.8E-16 1.9E-14
LDF vs time-adapted LDF 2.1E-04 2.8E-04 1.5E-05 2.9E-06 3.7E-06 2.4E-09 1.7E-08
FD vs time-adapted LDF 1.9E-08 1.1E-11 1.5E-13 1.6E-11 1.4E-12 1.4E-11 2.1E-12
MIL-101-T (5)/methanol
LDF 0.643 1.85 4.04 1.91 1.85 2.85 5.04 0.89
FD 0.149 − 1634.62 − 1632.43 − 1634.56 − 1634.62 − 1633.62 − 1631.43 − 1635.58
Time-adapted LDF 0.033 – ¡2367.55 ¡2375.65 ¡2352.31 ¡2372.31 ¡2363.55 ¡2380.14
Overall and pairwise tests P-value 2376.31
Overall 1.7E-14 2.3E-15 1.7E-15 2.5E-16 2.9E-15 3.7E-14 1.9E-16
LDF vs time-adapted LDF 2.6E-05 1.4E-04 1.7E-06 2.8E-05 2.9E-05 2.3E-06 2.2E-09
FD vs time-adapted LDF 2.8E-08 2.8E-09 2.6E-11 2.4E-12 2.6E-11 2.8E-12 2.9E-12
MIL-101-T (10)/methanol
LDF 0.069 − 268.68 − 266.48 − 268.62 − 268.68 − 267.68 − 265.48 − 269.63
FD 0.1334 − 224.51 –222.31 − 224.45 − 224.51 –223.51 − 221.31 − 225.46
Time-adapted LDF 0.039 ¡292.09 ¡300.27 ¡276.91 ¡296.91 ¡288.09 ¡304.69
Overall and pairwise tests P-value ¡300.91
Overall 2.9E-15 2.3E-16 2.8E-14 2.5E-15 1.8E-16 1.7E-14 2.6E-15
LDF vs time-adapted LDF 2.1E-05 2.1E-06 1.8E-05 2.2E-06 1.5E-08 2.3E-05 2.8E-09
FD vs time-adapted LDF 2.7E-10 3.6E-11 2.5E-12 2.4E-13 2.8E-10 2.6E-09 2.6E-12

*Bold values represent the smallest ICs.

to be optimal) than others. The overall proportion test, as well as pair­


AIC : − 2l + 2p
wise comparisons, are covered in Section 3.4.
BIC : − 2l + pln(n)
AICc : − 2l + 2p(n/(n − p − 1))
3.1. Root mean square deviation (RMSD) AIC3 : − 2l + 3p (11)
mAIC : − 2l + 2p2
The error evaluation function RMSD is also commonly used [55], and CAIC : − 2l + p(ln(n) + 1)
its mathematical form is- ABIC : − 2l + pln((n + 2)/24)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
√n ( )2 Where 2l = − n(log2π + log(error) − logn + 1)
√ W − Wcal i
√i=1 exp
RMSD = (10)
n
3.3. Bootstrap approach
Here, Wexp and Wcal measure the experimental and projected/
calculated (using model) uptake, respectively. n presents the total Efron and Tibshirani (1993) [59] proposed the bootstrap sampling
number of observations. If the difference between the experimental and method for performing data-driven (non-parametric) statistical analysis
projected values is large, the square offers a large value in this error [60]. This method assumes that data (experimental) represent the un­
evaluation function. RMSD primarily advises against using models that derlying population. The bootstapping sampling gives a sampling dis­
produce substantial errors on a regular basis. It is based on the normal tribution of the estimates like proportion and average. Using the same
distribution, which is used to fit ordinary least square regression models. condition of replacement sampling method, the repeated samples are
drawn randomly from the original sample taking with same size of
3.2. Information based criterion sample in each repeated sample. Therefore, each repeated sample gives
an estimated value of the certain parameter (like adsorption uptake) so
In general, the penalized-likelihood information criteria (IC), − 2l + f that sampling distribution of the estimate is observed. Hence, the RMSD,
(n, p), contains a goodness of fit term (-2l) and a penalty function, f(n, p) standard errors (SE), and confidence intervals (CI) of the parameter
[56-58]. Here, n is the number of observations, and p indicates the (adsorption uptake) are easily estimated using the sampling distribu­
number of parameters in the kinetic model. The purpose of these criteria tion. These estimates are usually used for testing a parameter or
is to find the (optimal) model that best defines the data relationship. comparing models.
These criteria are founded on the idea that the best model is one that
sacrifices the least amount of information when estimating the true
3.4. Proportion tests
relationship. Thus, the information-based model selection criteria aim to
select the model which incurs the minimum discrepancy with the un­
For each candidate model, the bootstrap samples give 10,000 esti­
derlying model relative to all other models examined in the study. The
mates of each IC. At each bootstrap sample, the IC values of the candi­
following equation shows the list of IC’s used for selecting the optimal
date models are compared, and the smallest IC value is treated as
kinetic models in this study.
minimum IC value. For each sample, only one minimum IC value is
observed. The proportion of bootstrap samples when a model achieves
the minimum value of an IC estimates the selection rate of the model by
the specific IC, i.e.,
Here, i represents the candiadate model: LDF, FD, and time-adapted
LDF. Given information criterion, the selection rate of a specific model

10
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 6. Selection rate of (a) MIL-101-GP/methanol, and (b) MIL-101-T(5)/methanol, adsorption pairs using RMSD.

estimates the expected probability or likelihood that the criterion will pairwise comparisons, respectively. Thus, the selection of a single or set
declare the model to be the best kinetic model. Since the proportions of optimum models is validated separately for each criterion considered
obtained from bootstrap approximations of the true selection probabil­ in this study.
ities of the candidate models, a statistical test is conducted to examine
the equality of the likelihoods. Here the null hypothesis is, 3.4.1. Overall proportion tests

Number of bootstrap samples with minimum value of IC for model i


Selection rate for model i =
Number of bootstrap samples (10000)

H0; The three kinetic models are equally likely based on the IC’s vs. The Chi-squared (χ 2) test [61] is an overall test which can be used to
H1; At least two models have different selection probabilities for the examine whether a kinetic model is selected by chance, i.e., each of the
IC’s. three kinetic models are equally likely to be selected by a given IC. The
test procedure can be described using the following example. Let us say a
In case we find evidence against H0, a multiple comparison test gambler wants to see if the six numbers on a six-sided dice are equally
further examines the pairs of models that have unequal selection rates. likely to show on top when the dice is rolled. The gambler conducts an
The chi-square test and t-tests are used to conduct the overall and experiment by rolling the dice a large number of times (say 10000) and

11
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 7. Adsorption kinetics of (a) Zeolite tuff/CO2, (b) CS-CO2/CO2, and (c) Calgon BPL/CO2 pairs using RMSD errors.

observing the frequency Oi of i (i = 1,2,…,6) out of the 10,000 trials. The using the test–statistic is, χ 2 =
∑6 (Oi − Ei )2
where Oi denotes observed
i=1 Ei
proportion, Oi/10000 estimates the true probability (Pi) that the number
frequency and Ei = (1/6)*10000 estimates the expected frequency of i
i appears on top of the dice. The null hypothesis H0: P1 = P2 = …….. = P6.
under H0. The test statistic follows χ 2 distribution with five degrees of
Against the alternatives H1: At least two are not equal, can be tested
freedom (df.). If the p-value is small (<0.05), the null hypothesis H0 is

12
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Table 7
Optimal kinetic model using model selection criteria for n = 10000 bootstrap samples for zeolite tuff/CO2, CS-H2O/CO2, CS-CO2/CO2, and Calgon BPL/CO2 pairs using
RMSD errors.
Natural zeolite tuff/CO2

Model Error mean AIC BIC AICc mAIC AIC3 CAIC ABIC

LDF 0.0881 − 159.72 − 157.879 − 159.64 − 159.729 − 158.729 − 156.879 − 161.015


FD 0.1541 − 133.48 − 131.635 − 133.397 − 133.486 − 132.486 − 130.635 − 134.772
Time-adapted LDF 0.04339 ¡179.621 ¡186.069 ¡163.022 ¡183.022 ¡175.621 ¡192.166
Overall and pairwise tests P-value ¡187.02
Overall 2.4E-15 2.8E-16 2.4E-14 1.5E-16 2.5E-12 1.3E-16 2.1E-15
LDF vs time-adapted LFD 1.9E-03 1.6E-04 2.4E-04 1.9E-05 1.9E-04 2.1E-07 1.7E-08
FD vs time-adapted LFD 2.6E-07 1.5E-08 1.8E-11 2.8E-11 1.6E-09 1.2E-11 1.9E-10
CS-H2O/CO2
LDF 0.039 − 40.80 − 40.16 − 40.46 − 40.80 − 39.80 − 39.16 − 43.20
FD 0.098 − 27.72 − 27.08 − 27.38 − 27.72 − 26.72 − 26.08 − 30.12
Time-adapted LDF 0.005 ¡61.07 ¡59.18 ¡39.62 ¡59.62 ¡57.07 ¡73.24
Overall and pairwise tests P-value ¡63.62
Overall 2.3E-12 1.9E-13 3.1E-16 3.2E-15 2.7E-15 1.9E-13 1.8E-15
LDF vs time-adapted LDF 1.8E-05 1.9E-04 1.7E-04 2.4E-05 2.9E-04 2.3E-07 2.7E-08
FD vs time-adapted LDF 2.1E-07 2.3E-10 1.6E-13 1.4E-11 2.4E-13 2.4E-10 1.8E-12
CS-CO2/CO2
LDF 0.050 − 37.03 − 36.39 − 36.70 − 37.03 − 36.03 − 35.39 − 39.44
FD 0.132 –23.52 –22.88 –23.18 –23.52 –22.52 − 21.88 − 25.92
Time-adapted LDF 0.013 ¡49.80 ¡47.25 ¡45.36 ¡25.80 ¡45.80 ¡43.25 ¡59.43
Overall and pairwise tests P-value
Overall 1.5E-16 2.6E-15 1.9E-13 1.2E-15 1.7E-15 1.3E-13 2.2E-15
LDF vs time-adapted LDF 2.1E-07 2.1E-08 1.8E-06 2.2E-09 1.5E-04 2.3E-06 2.8E-10
FD vs time-adapted LDF 1.8E-11 1.2E-12 1.5E-11 1.4E-12 1.8E-10 2.4E-11 2.1E-10
Calgon BPL/CO2
LDF 0.032 − 43.36 − 42.72 − 43.03 ¡43.36 − 42.36 − 41.72 − 45.77
FD 0.117 − 25.24 − 24.60 − 24.91 − 25.24 − 24.24 –23.60 − 27.65
Time-adapted LDF 0.014 ¡46.54 ¡44.65 − 25.10 ¡45.10 ¡42.54 ¡58.72
Overall and pairwise tests P-value ¡49.10
Overall 1.8E-14 1.9E-15 1.8E-16 1.7E-14 1.8E-15 1.9E-15 2.5E-16
LDF vs time-adapted LDF 1.1E-08 2.1E-05 3.8E-07 3.2E-09 1.5E-08 4.3E-04 1.8E-06
FD vs time-adapted LDF 1.7E-12 1.6E-13 1.5E-12 2.5E-12 2.6E-11 2.4E-13 2.4E-12

*Bold values represent the smallest ICs.

rejected in favor of H1. In this study, we have three candidate models to different IC’s, the overall and pairwise comparison tests are performed
test the selection probabilities for different IC. Using the bootstrap independently for each criterion.
samples, the errors are obtained separately for the candidate models, ICs
are computed, and the models with the minimum value of each IC are 4. Results and discussion
recorded for each sample. Thus, given IC, the 10,000 bootstrap samples
provide Oi and Ei = (1/3)*10000 for the ith model (i = 1,2,3). Finally, the The GRG non-linear optimization technique was used to fit the
p-value is obtained using the observed value of the test statistic and χ 2 experimental adsorption kinetic data of thirteen gas–solid adsorption
distribution with 2 df. A small p-value (<0.05) provides sufficient evi­ pairs shown in Table 2. Table 3 presents the fitting parameters of the
dence that the three kinetic models are not equally probable to be studied kinetic models for different pairs using GRG non-linear optimi­
preferred by the given IC. This means at least one model may have a zation methods.
higher chance of being chosen than other candidates to be selected by All kinetic models were fitted to the experimental data, but only the
the criteria using the experimental data. In this study, the overall test is probable optimum models, i.e., FD, LDF, and time-adapted LDF models,
performed independently for each criterion to investigate if the specific are selected for further examination. Here, it can be said that only these
model(s) are selected to be the optimum by most of the criteria. three kinetic models fit comparatively better than other models.
Therefore, these three models were picked for statistical analysis. In this
3.4.2. Pairwise test: Multiple comparisons study, the RMSD error is calculated for each model. This section ex­
If the overall proportion test suggests that the selection probabilities amines statistical approaches like overall and pairwise proportion tests
of the three kinetic models are unequal, it is required to identify the pair and selection probabilities using 10,000 bootstrap samples.
(s) of models with different selection probabilities for the given IC. This
is accomplished by a multiple comparisons test [61], which compares 4.1. Water adsorption
the selection probabilities of all the 3C2 = 3 pairs of models under
investigation. For instance, the null hypothesis that two models (1 & 2) Teo et al. [48] studied the water adsorption on AQSOA-Z01 and
are equally probable, i.e., achieve minimum IC value, is given by H0: P1 AQSOA-Z02 at different temperatures. They evaluated adsorption
= P2, which is tested against the alternative hypothesis H1: They are not isotherm models for characterization. Sun et al. [7] addressed the
̂p 1 − ̂p 2
equal. The test statistic is given by- Z = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ p = xn11 +x
where ̂ 2
+n2
thermodynamics of water adsorption kinetics onto RD silica gel. Teo
̂p (1− ̂p )(1/n1 +1/n2 )
et al. [51] also focused on the kinetics of water adsorption onto
indicates pooled sample proportion, n1 and n2 (10000) denote the size of Aluminium Fumarate. Fig. 1 depicts fractional adsorption uptake at
the bootstrap samples. A small p-value gives evidence against the various times for three possible optimum kinetic model fittings of water
equality of selection probabilities of the two models, implying that one adsorption onto AQSOA-Z01, RD silica gel, and aluminium fumarate
model is more likely than the others. Thus, the multiple comparisons test adsorbents. The kinetic model fittings of the AQSOA-Z02 adsorbent are
validates the selection of the best model based on the selection rates that presented in SF1 in the supplementary file.
are significantly higher than others. As the study considers seven For four water adsorption pairs, Table 4 shows the RMSD error

13
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Fig. 8. Selection rate of (a) Natural zeolite tuff/CO2, (b) CS-CO2/CO2, and (c) Calgon BPL/CO2 adsorption pairs using RMSD.

means (10000 bootstrap sample) and the values of information criteria mean for all four pairs. This means that, when all IC is considered, the
(IC’s) of three possible optimum kinetic models. It also includes the p- information loss of the time-adapted LDF model is lower than that of the
value for both the overall and pairwise proportion tests. It is found that other two models. As a result, the time-adapted LDF model has a smaller
the values of all IC’s for the time-adapted LDF model are small and amount of information loss than other models for these three studied
associated with other possible models when considering the RMSD error water adsorption pairs. Table 4 shows the p-values of proportion tests for

14
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

the equality of the minimum IC’s. It demonstrates that the p-value for models for two methanol adsorption pairs is graphically depicted in
these pairs is very low, close to zero, and less than 0.01. As a result, at a Fig. 6. The selection rate diagram is presented in SF5 in the supple­
1% level of significance, both tests are highly significant. It claims that mentary documents. When all criteria are considered, the time-adapted
when compared to other possible models, the time-adapted LDF model LDF model clearly has the highest selection rate among the three kinetic
has much-reduced information loss than other models, hence selected as models.
the optimal model by each selection criterion. Fig. 2 illustrates a bar
diagram showing the selection probability that was constructed using 4.4. CO2 adsorption
10,000 bootstrap samples. The selection rate of the possible kinetic
models for three adsorption pairs was graphically depicted. The selec­ Ammendola et al. [53] studied CO2 adsorption onto natural zeolite
tion rate graph for AQSOA-Z02/water pair has been presented in SF2 in tuff at 25 ◦ C. Further to our research, we investigated CO2 adsorption
the supplementary files. When all criteria are considered, it is clear that [54] in the context of its capture and storage (CCS). Two AC’s made from
the time-adapted LDF model has the highest selection rate/probability biomass, namely, CS-H2O and CS-CO2; and Calgon BPL (bituminous
among the three kinetic models. According to AIC, the time-adapted LDF coal-based) are used. Fig. 7 shows the kinetic model fittings for CO2
model has an estimated selection probability of 75 percent for AQSOA- adsorption onto natural zeolite tuff, CS-CO2, and Calgon BPL are pre­
Z01/water pair, 78 percent for RD silica gel/water pair, 65 percent for sented at various times. The kinetic model fittings for CS-H2O/CO2 are
Al-Fum/water pair. As a result, the time-adapted LDF model is the op­ presented in SF6 in supplementary documents.
timum kinetic model. Table 7 presents the IC values and statistical test results for CO2
adsorption. The values of IC’s for the time-adapted LDF model are
4.2. Ethanol adsorption minimal when compared to LDF and FD models for all CO2 adsorption
pairs presented in Table 7. As a result, for these adsorption pairs, the
Pal et al. [26] studied ethanol vapor adsorption uptake and kinetics time-adapted LDF model has less information loss than LDF models.
onto biomass-based activated carbon. The author developed two ad­ Overall and pairwise proportion test confirms that the time-adapted LDF
sorbents: M− AC, which was prepared from mangroves, and WPT-AC, model is significantly more optimum than the other two models. The
which was prepared using waste palm trunks. The experimental ki­ selection rate depicted in Fig. 8 demonstrates that the time-adapted LDF
netic data are extracted from this published article [26]. Fig. 3(a) and (b) model obviously has the highest selection rate among the three kinetic
show the kinetic model fittings for ethanol vapor adsorption onto models. The selection rate diagram for CS-H2O/CO2 pair is presented in
M− AC/ethanol at 50 ◦ C and WPT-AC/ethanol at 30 ◦ C, respectively. The SF7 in the supplementary file.
kinetic model fittings for M− AC/ethanol at 70 ◦ C and WPT-AC/ethanol
at 50 ◦ C are also presented in SF3 in the supplementary file (see Fig. 4). 5. Conclusions
Table 5 displays the RMSD error means (10000 bootstrap sample)
and the values of IC’s of the possible kinetic models for two ethanol For a gas–solid adsorption system, the equilibrium fractional
adsorption pairs. Also, it presents the p-value of the overall and pairwise adsorption uptake of thirteen pairs has been modeled using popularly
proportion tests. It is found that all of the values of ICs for the time- used kinetic models employing generalized reduced gradient (GRG)
adapted LDF model are minimal when compared to LDF and FD non-linear optimization techniques considering RMSD error evaluation
models for M− AC/ethanol and WPT-AC/ethanol adsorption pairs. As a function. The findings of this investigation demonstrate which kinetic
result, for these two ethanol adsorption pairs, the time-adapted LDF model is most suited for a gas–solid adsorption process. The present
model has less information loss than LDF models. It demonstrates that study indicates that the time-adapted LDF model is the best suitable than
the p-value of overall and pairwise proportion tests for these pairs is very conventional LDF and FD models for gas–solid adsorption systems in
low, close to zero, and less than 0.01. It claims that when compared to terms of information losses based on a comprehensive assessment of IC’s
other possible models, the time-adapted LDF model has much-reduced employing bootstrap error means. The time-adapted LDF model is
information loss than other candidates, hence selected as the optimal significantly optimum from a statistical standpoint, as evidenced by the
model by each selection criterion. The selection rate of the possible ki­ small p-values (close to zero) for two proportion tests subsequent to
netic models for (a) M− AC/ethanol, and (b) WPT-AC/ethanol pairs have adsorption pairs, and the identified optimal model could also be perti­
been presented in Fig. 5. It is found that the time-adapted LDF model has nent for other pairs for the gas–solid adsorption process. According to
the highest selection rate/probability among the three kinetic models. AIC, it is found that the time-adapted LDF model has an estimated se­
So, it can be stated that the time-adapted LDF model is an optimum lection probability of 75% for AQSOA-Z01/water pair, 78% for RD silica
kinetic model. gel/water pair, 65% for Al-Fum/water pair, whereas for all of these pairs
LDF model shows 23%, 7.5%, and 32%, respectively. Finally, among all
4.3. Methanol adsorption possible kinetic models, selection probability based on bootstrapping
confirms that the time-adapted LDF model has the highest selection rate.
We extended our work to MIL-101(Cr) [52], a porous crystalline This study will help the researcher in the adsorption field in their
solid that belongs to a family of porous crystalline solids. The large rigorous investigation of adsorption kinetic analysis as well as adsorp­
pressure jump (LPJ) method was used to measure the dynamics of tion characteristics.
methanol sorption onto MIL-101(Cr). Fig. 5 shows the experimental
methanol uptakes onto MIL-101-GP and MIL-101-T (5), as well as the CRediT authorship contribution statement
three most likely optimum kinetics models. The kinetic model fittings of
MIL-101-T (10)/methanol pairs are presented in SF4 in the supple­ Md. Matiar Rahman: Conceptualization, Formal analysis, Method­
mentary file. Table 6 presents the IC values and statistical test results for ology, Writing – original draft. Abu Zar Shafiullah: Conceptualization,
methanol adsorption. It is found that the values of IC’s for the time- Formal analysis, Methodology. Mahua Jahan Rupa: Conceptualization,
adapted LDF model are minimal when compared to LDF and FD Formal analysis. Shamal Chandra Karmaker: Conceptualization,
models for all pairs. So, for these adsorption pairs, the time-adapted LDF Methodology. Shahadat Hosan: Formal analysis, Methodology. Ani­
model has less information loss than LDF and FD models. Both test re­ mesh Pal: Formal analysis, Writing – original draft. Bidyut Baran Saha:
sults in Table 6 confirm that the time-adapted LDF model is significantly Conceptualization, Methodology, Writing – review & editing,
better than the other two models. The selection rate of these three Supervision.

15
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Declaration of Competing Interest [22] I.I. El-Sharkawy, B.B. Saha, S. Koyama, K.C. Ng, A study on the kinetics of ethanol-
activated carbon fiber: Theory and experiments, Int. J. Heat Mass Transf. 49 (2006)
3104–3110, https://doi.org/10.1016/j.ijheatmasstransfer.2006.02.029.
The authors declare that they have no known competing financial [23] T. Yan, T.X. Li, R.Z. Wang, R. Jia, Experimental investigation on the ammonia
interests or personal relationships that could have appeared to influence adsorption and heat transfer characteristics of the packed multi-walled carbon
the work reported in this paper. nanotubes, Appl. Therm. Eng. 77 (2015) 20–29, https://doi.org/10.1016/j.
applthermaleng.2014.12.001.
[24] I.I. El-Sharkawy, K. Uddin, T. Miyazaki, B.B. Saha, S. Koyama, J. Miyawaki, S.-
Data availability H. Yoon, Adsorption of ethanol onto parent and surface treated activated carbon
powders, Int. J. Heat Mass Transf. 73 (2014) 445–455, https://doi.org/10.1016/J.
IJHEATMASSTRANSFER.2014.02.046.
Data will be made available on request. [25] M. Sultan, I.I. El-Sharkawy, T. Miyazaki, B.B. Saha, S. Koyama, T. Maruyama,
S. Maeda, T. Nakamura, Water vapor sorption kinetics of polymer based sorbents:
Appendix A. Supplementary data Theory and experiments, Appl. Therm. Eng. 106 (2016) 192–202, https://doi.org/
10.1016/j.applthermaleng.2016.05.192.
[26] A. Pal, H.S. Kil, S. Mitra, K. Thu, B.B. Saha, S.H. Yoon, J. Miyawaki, T. Miyazaki,
Supplementary data to this article can be found online at https://doi. S. Koyama, Ethanol adsorption uptake and kinetics onto waste palm trunk and
org/10.1016/j.applthermaleng.2023.120581. mangrove based activated carbons, Appl. Therm. Eng. 122 (2017) 389–397,
https://doi.org/10.1016/j.applthermaleng.2017.04.099.
[27] I.I. El-Sharkawy, K. Uddin, T. Miyazaki, B.B. Saha, S. Koyama, H.-S. Kil, S.-H. Yoon,
References J. Miyawaki, Adsorption of ethanol onto phenol resin based adsorbents for
developing next generation cooling systems, Int. J. Heat Mass Transf. 81 (2015)
[1] M.M. Rahman, M. Muttakin, A. Pal, A.Z. Shafiullah, B.B. Saha, A statistical 171–178, https://doi.org/10.1016/j.ijheatmasstransfer.2014.10.012.
approach to determine optimal models for IUPAC-classified adsorption isotherms, [28] S. Sircar, J.R. Hufton, Why Does the Linear Driving Force Model for Adsorption
Energies. 12 (2019), https://doi.org/10.3390/en12234565. Kinetics Work? Adsorption. 6 (2000) 137–147, https://doi.org/10.1023/A:
[2] B. Sakintuna, F. Lamari-Darkrim, M. Hirscher, Metal hydride materials for solid 1008965317983.
hydrogen storage: A review, Int. J. Hydrogen Energy. 32 (2007) 1121–1140, [29] T.D.J. Kaerger, D.M. Ruthven, Diffusion in nanoporous materials, Wiley-VCH,
https://doi.org/10.1016/j.ijhydene.2006.11.022. Weinheim, 2012.
[3] K. Kosaka, M. Asami, N. Kobashigawa, K. Ohkubo, H. Terada, N. Kishida, M. Akiba, [30] Y.I. Aristov, Optimal adsorbent for adsorptive heat transformers: Dynamic
Removal of radioactive iodine and cesium in water purification processes after an considerations, Int. J. Refrig. 32 (2009) 675–686, https://doi.org/10.1016/j.
explosion at a nuclear power plant due to the Great East Japan Earthquake, Water ijrefrig.2009.01.022.
Res. 46 (2012) 4397–4404, https://doi.org/10.1016/j.watres.2012.05.055. [31] A. Raymond, S. Garimella, Intraparticle mass transfer in adsorption heat pumps:
[4] R.T. Yang, Gas separation by adsorption processes, Butterworths, 1987. Limitations of the linear driving force approximation, J. Heat Transf. (2011),
[5] K. Thu, H. Yanagi, B.B. Saha, K.C. Ng, Performance investigation on a 4-bed https://doi.org/10.1115/1.4001310.
adsorption desalination cycle with internal heat recovery scheme, Desalination. [32] I.I. El-Sharkawy, On the linear driving force approximation for adsorption cooling
402 (2017) 88–96, https://doi.org/10.1016/j.desal.2016.09.027. applications, Int. J. Refrig. 34 (2011) 667–673, https://doi.org/10.1016/j.
[6] M.M. Rahman, S.C. Karmaker, A. Pal, O. Eljamal, B.B. Saha, Statistical techniques ijrefrig.2010.12.006.
for the optimization of cesium removal from aqueous solutions onto iron-based [33] M.J. Rupa, A. Pal, S. Mitra, B.B. Saha, Time adapted linear driving force model for
nanoparticle-zeolite composites, Environ. Sci. Pollut. Res. 28 (2021) 12918–12931, gas adsorption onto solids, Chem. Eng. J. 420 (2021) 129785, https://doi.org/
https://doi.org/10.1007/s11356-020-11258-1. 10.1016/j.cej.2021.129785.
[7] B. Sun, A. Chakraborty, Thermodynamic frameworks of adsorption kinetics [34] B. Likozar, D. Senica, A. Pavko, Equilibrium and Kinetics of Vancomycin
modeling: Dynamic water uptakes on silica gel for adsorption cooling applications, Adsorption on Polymeric Adsorbent, AIChE J. 59 (2012) 215–228, https://doi.org/
Energy. 84 (2015) 296–302, https://doi.org/10.1016/j.energy.2015.02.101. 10.1002/aic.
[8] L. Largitte, R. Pasquier, New models for kinetics and equilibrium homogeneous [35] B. Likozar, D. Senica, A. Pavko, Interpretation of experimental results for
adsorption, Chem. Eng. Res. Des. 112 (2016) 289–297, https://doi.org/10.1016/j. vancomycin adsorption on polymeric resins in a fixed bed column by mathematical
cherd.2016.06.021. modeling with independently estimated parameters, Ind. Eng. Chem. Res. 52
[9] J.P. Simonin, On the comparison of pseudo-first order and pseudo-second order (2013) 9247–9258, https://doi.org/10.1021/ie400832p.
rate laws in the modeling of adsorption kinetics, Chem. Eng. J. 300 (2016) [36] B. Likozar, D. Senica, A. Pavko, Comparison of adsorption equilibrium and kinetic
254–263, https://doi.org/10.1016/j.cej.2016.04.079. models for a case study of pharmaceutical active ingredient adsorption from
[10] Y.S. Ho, Second-order kinetic model for the sorption of cadmium onto tree fern: A fermentation broths: Parameter determination, simulation, sensitivity analysis and
comparison of linear and non-linear methods, Water Res. 40 (2006) 119–125, optimization, Brazilian, J Chem. Eng. 29 (2012) 635–652, https://doi.org/
https://doi.org/10.1016/j.watres.2005.10.040. 10.1590/S0104-66322012000300020.
[11] S. Azizian, M. Haerifar, H. Bashiri, Adsorption of methyl violet onto granular [37] E. Glueckauf, Theory of chromatography. Part 10.—Formulæ for diffusion into
activated carbon: Equilibrium, kinetics and modeling, Chem. Eng. J. 146 (2009) spheres and their application to chromatography, Trans. Faraday Soc. 51 (1955)
36–41, https://doi.org/10.1016/j.cej.2008.05.024. 1540–1551, https://doi.org/10.1039/TF9555101540.
[12] S.W. Rutherford, J.E. Coons, Equilibrium and kinetics of water adsorption in [38] E. Glueckauf, J.I. Coates, The Influence of Incomplete Equilibrium on the Front
carbon molecular sieve: Theory and experiment, Langmuir. 20 (2004) 8681–8687, Boundary of Chrornatogram and on the Effectiveness of SepuTation, Theory
https://doi.org/10.1021/la049330d. Chromatogr. (1947) 1315–1321.
[13] S. Tripathi, R.F. Tabor, Modeling two-rate adsorption kinetics: Two-site, two- [39] C.H. Liaw, J.S.P. Wang, R.A. Greenkorn, K.C. Chao, Kinetics of fixed-bed
species, bilayer and rearrangement adsorption processes, J. Colloid Interface Sci. adsorption: A new solution, AIChE J. 25 (1979) 376–381, https://doi.org/
476 (2016) 119–131, https://doi.org/10.1016/j.jcis.2016.05.007. 10.1002/aic.690250229.
[14] L. Largitte, R. Pasquier, A review of the kinetics adsorption models and their [40] S. Sircar, J.R. Hufton, Intraparticle adsorbate concentration profile for linear
application to the adsorption of lead by an activated carbon, Chem. Eng. Res. Des. driving force model, AIChE J. 46 (2000) 659–660, https://doi.org/10.1002/
109 (2016) 495–504, https://doi.org/10.1016/j.cherd.2016.02.006. aic.690460325.
[15] M. Sekar, V. Sakthi, S. Rengaraj, Kinetics and equilibrium adsorption study of lead [41] A.E. Rodrigues, C.M. Silva, What’s wrong with Lagergreen pseudo first order model
(II) onto activated carbon prepared from coconut shell, J. Colloid Interface Sci. 279 for adsorption kinetics? Chem. Eng. J. 306 (2016) 1138–1142, https://doi.org/
(2004) 307–313, https://doi.org/10.1016/j.jcis.2004.06.042. 10.1016/j.cej.2016.08.055.
[16] W. Dang, J. Zhang, H. Nie, F. Wang, X. Tang, N. Wu, Q. Chen, X. Wei, R. Wang, [42] Y.I. Aristov, Dynamics of adsorptive heat conversion systems: Review of basics and
Isotherms, thermodynamics and kinetics of methane-shale adsorption pair under recent advances, Energy. 205 (2020) 117998, https://doi.org/10.1016/j.
supercritical condition: Implications for understanding the nature of shale gas energy.2020.117998.
adsorption process, Chem. Eng. J. 383 (2020), 123191, https://doi.org/10.1016/j. [43] S. Lagergren, About the theory of so-called adsorption of soluble substances, Sven.
cej.2019.123191. Vetenskapsakad. Handl. 24 (1898) 1–39.
[17] Y. Liu, L. Shen, From Langmuir kinetics to first- and second-order rate equations for [44] M.J. Rupa, A. Pal, B.B. Saha, Activated carbon-graphene nanoplatelets based green
adsorption, Langmuir. 24 (2008) 11625–11630, https://doi.org/10.1021/ cooling system: Adsorption kinetics, heat of adsorption, and thermodynamic
la801839b. performance, Energy. 193 (2020) 116774, https://doi.org/10.1016/j.
[18] K. Chihara, M. Suzuki, Air drying by pressure swing adsorption, J. Chem. Eng. energy.2019.116774.
Japan. 16 (1983) 293–299. https://doi.org/S0960-1481 (97)00067-0. [45] B.B. Saha, I.I. El-Sharkawy, A. Chakraborty, S. Koyama, S.-H. Yoon, K.C. Ng,
[19] M. Haerifar, S. Azizian, An exponential kinetic model for adsorption at solid/ Adsorption rate of ethanol on activated carbon fiber, J. Chem. Eng. Data. 51 (2006)
solution interface, Chem. Eng. J. 215–216 (2013) 65–71, https://doi.org/10.1016/ 1587–1592, https://doi.org/10.1021/je060071z.
j.cej.2012.11.017. [46] J.H. Hills, An investigation of the linear driving force approximation to diffusion in
[20] M. Muttakin, A. Pal, M.J. Rupa, K. Ito, B.B. Saha, A critical overview of adsorption spherical particles, Chem. Eng. Sci. 41 (1986) 2779–2785, https://doi.org/
kinetics for cooling and refrigeration systems, Adv. Colloid Interface Sci. 294 10.1016/0009-2509(86)80009-4.
(2021), 102468, https://doi.org/10.1016/j.cis.2021.102468. [47] R.T. Yang, Rate Processes in Adsorbers, in: Gas Sep. by Adsorpt, Process., Elsevier,
[21] D.M. Ruthven, Principles of Adsorption and Adsorption Processes, 1st ed.,, John 1987, pp. 101–139, https://doi.org/10.1016/B978-0-409-90004-0.50007-6.
Wiley & Sons. Inc., 1984. [48] H.W.B. Teo, A. Chakraborty, W. Fan, Improved adsorption characteristics data for
AQSOA types zeolites and water systems under static and dynamic conditions,

16
Md.M. Rahman et al. Applied Thermal Engineering 229 (2023) 120581

Micropor. Mesopor. Mater. 242 (2017) 109–117, https://doi.org/10.1016/j. [55] M.M. Rahman, A. Pal, K. Uddin, K. Thu, B.B. Saha, Statistical analysis of optimized
micromeso.2017.01.015. isotherm model for maxsorb III/ethanol and silica gel/water pairs, Evergreen. 5
[49] M.A. Islam, A. Pal, B.B. Saha, Experimental study on thermophysical and porous (2018) 1–12, https://doi.org/10.5109/2174852.
properties of silica gels, Int. J. Refrig. 110 (2020) 277–285, https://doi.org/ [56] L.S. Sclove, Application of model selection criteria to some problems in
10.1016/j.ijrefrig.2019.10.027. multivariate analysis, Psychometrika. 52 (1987) 333–343.
[50] H.T. Chua, K.C. Ng, A. Chakraborty, N.M. Oo, M.A. Othman, Adsorption [57] C. Koukouvinos, K. Mylona, A. Skountzou, A variable selection method for
characteristics of silica gel + water systems, J. Chem. Eng. Data. 47 (2002) analyzing supersaturated designs, Commun. Stat. Simul. Comput. 40 (2011)
1177–1181, https://doi.org/10.1021/je0255067. 484–496, https://doi.org/10.1080/03610918.2010.546540.
[51] H.W.B. Teo, A. Chakraborty, Y. Kitagawa, S. Kayal, Experimental study of [58] J.J. Dziak, D.L. Coffman, S.T. Lanza, R. Li, Sensitivity and Specificity of
isotherms and kinetics for adsorption of water on Aluminium Fumarate, Int. J. Heat Information Criteria, Peer J Prepr. Corte Madera. (2015) 1–20.
Mass Transf. 114 (2017) 621–627, https://doi.org/10.1016/j. [59] B. Efron, R.J. Tibshirani, Introduction to the Bootstrap, Springer, Toronto, Canada,
ijheatmasstransfer.2017.06.086. 1993.
[52] M.V. Solovyeva, L.G. Gordeeva, Y.I. Aristov, “MIL-101(Cr)–methanol” as working [60] M. Joshi, A. Kremling, A. Seidel-Morgenstern, Model based statistical analysis of
pair for adsorption heat transformation cycles: Adsorbent shaping, adsorption adsorption equilibrium data, Chem. Eng. Sci. 61 (2006) 7805–7818, https://doi.
equilibrium and dynamics, Energy Convers. Manag. 182 (2019) 299–306, https:// org/10.1016/j.ces.2006.08.052.
doi.org/10.1016/j.enconman.2018.12.065. [61] S.K. Saha, M.E. Haque, D. Islam, M.M. Rahman, M.R. Islam, A. Parvin, S. Rahman,
[53] P. Ammendola, F. Raganati, R. Chirone, F. Miccio, Fixed bed adsorption as affected Comparative study between the effect of Momordica charantia (wild and hybrid
by thermodynamics and kinetics: Yellow tuff for CO2 capture, Powder Technol. variety) on hypoglycemic and hypolipidemic activity of alloxan induced type 2
(2020), https://doi.org/10.1016/j.powtec.2020.06.075. diabetic long-evans rats, J. Diabetes Mellit. 02 (2012) 131–137, https://doi.org/
[54] N. Álvarez-Gutiérrez, M.V. Gil, F. Rubiera, C. Pevida, Kinetics of CO2 adsorption on 10.4236/jdm.2012.21022.
cherry stone-based carbons in CO2/CH4 separations, Chem. Eng. J. 307 (2017)
249–257, https://doi.org/10.1016/j.cej.2016.08.077.

17

You might also like