pubs.acs.

org/JPCC Article

Metal−Organic Frameworks for Xylene Separation: From

Computational Screening to Machine Learning
Zhiwei Qiao, Yaling Yan, Yaxing Tang, Hong Liang, and Jianwen Jiang*
Cite This: https://doi.org/10.1021/acs.jpcc.0c10773 Read Online

ACCESS Metrics & More Article Recommendations *

sı Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: Separation of xylene isomers is an important process

in the chemical industry and there has been considerable interest in
developing advanced materials for xylene separation. In this study,
we synergize computational screening and machine learning to
Downloaded via UNIV OF MALAYA on April 5, 2021 at 02:16:38 (UTC).

explore the selective adsorption of p-xylene over o- and m-xylene in

metal−organic frameworks (MOFs). First, a large set (4764) of
computation-ready experimental MOFs is screened by geometric
analysis and molecular simulation. The relationships between MOF
structural descriptors (void fraction, volumetric surface area, and
largest cavity diameter) and separation performance metrics
(adsorption capacity of p-xylene Np‑xylene and selectivity of p-xylene
over o- and m-xylene Sp/(m+o)) are established. Then two machine-
learning methods (back-propagation neural network and decision
tree), as well as particle swarm optimization, are utilized to analyze and optimize Np‑xylene and Sp/(m+o). The importance of each
descriptor for separation is evaluated in six different MOF data sets. In the 100 top-performing MOFs, the pore limiting diameter
(PLD) and largest cavity diameter (LCD) are revealed to be key factors governing separation performance. On the basis of the
threshold values of Np‑xylene > 0.5 mol/kg and Sp/(m+o) > 5, seven top-performing MOFs are identified. By further incorporating
framework flexibility, JIVFUQ is predicted to be the best and superior to many reported MOFs in the literature.

1. INTRODUCTION MOFs have attracted considerable interest for many potential

Xylene is produced by the catalytic reforming of crude oil with applications such as separation, catalysis, sensing, etc. Xylene
an annual amount of several million tons.1 It has three isomers, separation has been examined in a number of MOFs with open
namely, p-, o-, and m-xylene; each isomer is an important raw metal sites,6,7 molecular sieving effects,8,9 or flexible frame-
material for the manufacturing of various high-value-added works.10−12 Using single-component adsorption isotherms and
chemicals and polymers. In particular, p-xylene is the most multicomponent breakthrough measurements, Gonzalez et al.
desirable for producing polyethylene terephthalate (PET), evaluated Co2(dobdc) and Co2(m-dobdc) for xylene separa-
polyester fiber, and synthetic resin. On the other hand, o- tion; through the varied extent of interaction for each isomer
xylene is a common precursor for phthalic anhydride and with two adjacent coordinatively unsaturated Co centers,
plasticizer; m-xylene is primarily used for the synthesis of Co2(dobdc) was found to distinguish among all of the isomers
isophthalic acid, which is further used for the blending of PET and maintain good selectivity over a wide range of
resin.2 The three isomers exist as a mixture after catalytic concentrations.7 Lannoeye et al. demonstrated the shape-
reforming, and it is thus important to separate them. selective adsorption of p-xylene by a highly stable aliphatic
Nevertheless, their separation is challenging because of their Cu−MOF with molecular sieving in its one-dimensional
identical molecular weights, similar molecular sizes, and channels.9 By embedding stable nitroxides in ZIF-8, Polyukhov
thermophysical properties (Table S1). At present, they are et al. precisely measured and controlled the window size of
separated by crystallization or simulated moving bed ZIF-8 from 298 to 333 K, reaching up to 92−95% efficiency
adsorption.3 for xylene separation.12 Meanwhile, a handful of simulation
There have been considerable efforts toward the develop-
ment of advanced materials for xylene separation.4 In the past
two decades, metal−organic frameworks (MOFs) have Received: December 2, 2020
emerged as a special family of porous materials.5 The degrees Revised: March 19, 2021
of diversity and multiplicity in MOFs are substantially more
extensive than any other types of porous materials.
Intriguingly, the crystalline structures, surface areas, and pore
sizes in MOFs can be tuned in a rational manner. Therefore,

© XXXX American Chemical Society https://doi.org/10.1021/acs.jpcc.0c10773

A J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

studies have been reported for the separation of xylene To the best of our knowledge, there is only one reported
isomers. Torres-Knoop et al. simulated the separation of xylene study on the HTCS of MOFs for xylene separation.
isomers in 14 MOFs as well as 12 zeolites, highlighted the Specifically, among ∼2500 MOFs, Gee et al. predicted two
crucial role of channel dimension in governing separation, and MOFs (MOF-48 and MIL-140B) with comparable perform-
found MAF-X8 allowing commensurate stacking as a ance to zeolite BaX, which is currently used in industrial p-
promising candidate for p-xylene adsorption.13 Four different xylene enrichment from a C8 aromatics mixture; then they
force fields were evaluated by Granato et al. for the adsorption synthesized the top-performing MOFs and evaluated them
of xylene isomers in UiO-66 and found to reproduce through breakthrough adsorption.28 Herein, we aim to
experimentally observed selective adsorption of o-xylene.14 synergize HTCS and ML for xylene separation in CoRE-
Agrawal et al. experimentally measured the adsorption and MOFs. The HTCS and ML methods are briefly outlined in
separation of xylene isomers in MIL-53 with different metals, section 2. In section 3, we first present univariate analysis for
and MIL-53(Al) was observed to possess high o-xylene xylene separation performance with structural descriptors and
selectivity; meanwhile, their simulation showed good overall establish structure−performance relationships. Next, multi-
agreement with experiment.15 By combining experimental and variate analysis is conducted using ML tools to reveal key
simulation techniques, we investigated xylene separation by a factors governing separation performance. Then the top-
highly selective and stable MIL-160 membrane with a suitable performing MOFs are identified, and the best one is examined.
pore size (0.5−0.6 nm), which could selectively allow p-xylene Finally, the concluding remarks are summarized in section 4.
to pass through while excluding o-xylene and m-xylene.16 From
2. METHODS
simulation, Kim et al. explored the origin of highly selective
adsorption of p-xylene in a DABCO pillar-layered MOF, which 2.1. Computational Screening. The screening consisted
was attributed to the preferential packing of p-xylene on the of three steps as illustrated in Figure 1. (1) The atomic
DABCO ligand.17
The above-mentioned experimental and computational
studies were focused on a handful of MOFs for xylene
separation. It is highly desirable to identify the best candidates
among tens of thousands of synthesized MOFs; however,
experimental testing alone is formidable. Over the past several
years, high-throughput computational screening (HTCS) and
machine learning (ML) have been increasingly employed to
screen and predict MOFs for gas separation, particularly for
CO2-containing mixtures. From 137 953 hypothetical MOFs
(hMOFs), Wilmer et al. established the correlations between
structural descriptors and adsorbent evaluation criteria for
CO2/CH4 and CO2/N2 separation.18 A portion of these
hMOFs was further examined for precombustion CO2/H2
separation using a genetic algorithm, and the top-performing Figure 1. Screening flowchart.
MOFs were experimentally synthesized.19 We screened the
computation-ready experimental (CoRE)-MOFs for CO2/CH4
structures of 4764 CoRE-MOFs29 were characterized in terms
and CO2/N2 separation and unravelled the optimal metal
of pore limiting diameter (PLD), largest cavity diameter
types.20 Furthermore, hydrophobic CoRE-MOFs were identi-
(LCD), volumetric surface area (VSA), free volume (Vfree),
fied for the separation of H2S and CO2 from a CH4/C2H6/ and void fraction ϕ. Specifically, the PLD and LCD were
C3H8/H2S/CO2/H2O mixture toward the upgrading of natural estimated using the Zeo++,30 while the VSA and ϕ were
gas.21 From a library of over 300 000 MOFs, Boyd et al. determined with N2 and He as probes (kinetic diameters =
identified different classes of strong CO2 binding sites 3.64 Å for N2 and 2.58 Å for He), respectively, using the
(adsorbaphores) and subsequently synthesized two water- RASPA.31 (2) On the basis of the kinetic diameters of the
stable MOFs containing the most hydrophobic adsorba- xylene isomers (Table S1), 1502 MOFs with PLD > 5.0 Å
phore.22 Using ML algorithms, Fernandez et al. developed were selected to simulate the adsorption of a xylene isomer
quantitative structure−property relationships to rapidly and mixture by grand-canonical Monte Carlo (GCMC) method.
accurately recognize high-performing MOFs for CO2 cap- Intuitively, one may expect that none of the three isomers can
ture.23 A ML method was developed by Moosavi et al. to enter into a MOF with PLD < 6.7 Å because the kinetic
quantify the similarities of MOFs and analyze their chemical diameter of p-xylene dk is 6.7 Å.32,33 Nevertheless, we should
diversity, which identifies possible bias in MOF databases.24 note that the kinetic diameter is an empirical measure of
Yao et al. proposed an automated nanoporous materials molecular size, and it is generally larger than the actual size.
discovery platform for the generative design of reticular When the size of an adsorbate molecule is slightly larger than
materials and demonstrated the automated design process for the PLD, the adsorbate molecule will experience a repulsive
CO2 separation in MOFs.25 Target-specific MOFs were interaction with the pore but may still be able to enter the
designed for CO2 capture by Zhang et al. based on Monte pore. This phenomenon was observed in our previous work on
Carlo tree search and recurrent neural networks according to the separation of CO2, N2, and CH4.34 In the GCMC
required objective functions.26 Recently, Jablonka et al. simulations, the 1502 MOFs were assumed to be rigid. From
comprehensively reviewed big-data and ML methods as a the adsorption capacity of p-xylene and the selectivity of p-
powerful tool to investigate porous materials and discover xylene over m- and o-xylene, the top-performing MOFs were
complex correlations.27 shortlisted. (3) To take framework flexibility into account, the
B https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Figure 2. Models of xylene isomers (CH3 as a united atom).

flexible snapshot method was adopted.35 For each top- bias potential.40 Each GCMC simulation was conducted with
performing MOF shortlisted from step 2, molecular dynamics 200 000 cycles, the first 100 000 cycles for equilibration and
(MD) simulation was conducted with framework flexibility. the remaining 100 000 cycles for ensemble averages. The
Ten different structures were generated from the MD number of cycles was found to be sufficient, and its further
simulation and used for GCMC simulations, from which the increase had an insignificant effect on simulation results. Each
adsorption capacity and selectivity were averaged. Finally, the cycle consisted of n trial moves (n = the number of adsorbate
best MOFs were identified from the top-performing ones. molecules) including translation, rotation, regrowth, and swap.
Xylene isomers were described by the optimized potentials The GCMC simulations were carried out using the RASPA
for liquid simulations (OPLS) force field, which was found to package.31 The MD simulations with framework flexibility
give a good prediction for the adsorption of xylene isomers in were conducted in a canonical ensemble at 373 K using the
UiO-66.14 As shown in Figure 2, the carbon and hydrogen Forcite module in Materials Studio. The same force field and
atoms on the phenyl ring of xylene were mimicked atomically, atomic charges as used in the GCMC simulations were
whereas the methyl groups were treated in a united-atom adopted. The system temperature was maintained by the
manner. The interaction between xylene molecules was Nosé−Hoover method. The MD duration was 2 ns, and 10

ÄÅ É
ÅÅi y12 i y6ÑÑÑ
described by electrostatic and Lennard−Jones (LJ) potentials structures were generated every 200 ps.

ÅÅjj σij zz j z ÑÑ
2.2. Machine Learning. To analyze structure−perform-

+ ∑ 4εijÅÅÅÅjjj zzz − jjjj zzzz ÑÑÑÑ

j z j z
qiqj ance relationships from the above GCMC simulations for the

ÅÅÅk rij { r Ñ
σ

k ij { ÑÑÑÖ
ij

ÅÇ
∑ adsorption in 1502 MOFs, two ML methods, namely, back-
4πε0rij propagation neural network (BPNN) and decision tree (DT),
(1)
from the nntool package in Matlab were used. The BPNN is a
where qi is the atomic charge of atom i, ε0 = 8.8542 × 10−12 C2 widely used multilayer feedforward neural network method.
N−1 m−2 is the permittivity of a vacuum, rij is the distance The main characteristic of BPNN is forward for signals and
between atoms i and j, and εij and σij are the well depth and back-propagation for errors. In the forward, input signals are
collision diameter. Table S2 lists the LJ potential parameters handled step by step from an input layer, through a hidden
and atomic charges of the xylene isomers. For MOFs, the LJ layer, to an output layer. Each layer of a neuron state affects
potential parameters as listed in Table S3 were adopted from only its subsequent layer. If the output layer does not achieve
the universal force field (UFF).36 The cross LJ interactions an expected precision, back-propagation is conducted
between MOFs and xylene molecules were calculated by the automatically. On the basis of prediction errors, the BPNN
Lorentz−Berthelot mixing rules. The atomic charges of the method can adjust network weights and thresholds, thus
MOFs were estimated using the MEPO-Qeq method, which gradually approaching the output layer. The key point in the
can rapidly and accurately evaluate electrostatic interactions.37 BPNN is that the output errors of the hidden layer propagate
To examine the effect of the charge estimation method, the gradually back to the input layer, and then the weight of each
density-derived electrostatic and chemical (DDEC) charges38 independent variable can be dynamically adjusted after the
were also used for a few top-performing MOFs shortlisted output errors spread to the input layer. In this study, the
from step 2. BPNN included three layers: an input layer, a hidden layer
The adsorption of an equimolar mixture of three xylene with 10 neurons, and an output layer, as shown in Figure S1.
isomers was simulated at 100 °C and a total pressure of 30 kPa. The activation and training functions in the BPNN used here
Experimentally, breakthrough experiments for xylene separa- were the hyperbolic tangent sigmoid transfer function41 and
tion are usually conduced at temperatures from 70 to 150 °C the Levenberg−Marquardt function,42 respectively. Further-
and xylene pressures from 0.6 to 60 kPa.39 For each MOF, the more, the DT method was also used. DT is a simple, fast
cell lengths in three dimensions were extended to at least 24 Å supervised learning method. It does not require the assumption
and the periodic boundary conditions were imposed. The LJ of parameters; moreover, the generalization ability and
interactions were calculated with a spherical cutoff radius of 12 accuracy of DT can be improved by pruning trees. On the
Å, and the long-range corrections were added. The electro- DT nodes, eigenvalues are selected from the most important to
static interactions were calculated using the Ewald summation the next important one. For each eigenvalue, calculations are
method. To improve sampling efficiency, we used the attempted, and then optimal classification is defined as the
continuous fractional component Monte Carlo (CFCMC) parent node after considering the attribution of each feature.
method, which gradually inserts and deletes molecules through The independent variable is divided into two or more groups
the use of a continuous coupling parameter and an adaptive according to a certain splitting criterion. As shown in Figure
C https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Figure 3. (a−c) Np‑xylene and (d−f) Sp/(m+o) versus structural descriptors ϕ, VSA, and LCD.

S2, the most commonly used binary DT was adopted in this was used for the PSO and the initial weight was 1. More details
study. The classification and regression tree theory43 was can be found in the Supporting Information.
employed to determine the splitting criterion. The minparent
was set to be 10, representing the minimum number of samples 3. RESULTS AND DISCUSSION
in the input layer upon the first splitting, while the minimum of First, univariate analysis is conducted for separation perform-
the leaf node was 1, and the maximum of each splitting was 1. ance with three structural descriptors (ϕ, VSA, and LCD).
More details of the ML methods are described in the Two metrics are used to quantify separation performance: the
Supporting Information. For each ML method, 70% of the adsorption capacity of p-xylene Np‑xylene and the selectivity of p-
simulation data was randomly selected as a training set and the xylene over m- and o-xylene Sp/(m+o) defined as Sp/(m+o) = [xp/
remaining 30% as a test set. The performance of ML was (xm + xo)]/[yp/(ym + yo)], where x and y are the
quantified by the determination coefficient (R2), root-mean- compositions in adsorbed and bulk phases, respectively.
square error (RMSE), and mean absolute error (MAE). Then two ML tools (BPNN and DT) are utilized for
In addition, to accelerate optimal prediction by the two ML multivariate analysis to unravel the importance of five
methods, a stochastic global optimization technique, namely, descriptors (ϕ, VSA, LCD, PLD, and framework density ρ)
in separation performance. Next, the top-performing MOFs are
particle swarm optimization (PSO), was applied to calculate
identified by considering both Np‑xylene and Sp/(m+o). Finally, the
the parameters for the trade-off between adsorption capacity
adsorption in the best MOF (JIVFUQ) is examined.
and selectivity. Originally used to examine the behavior of 3.1. Univariate Analysis. Figure 3a−c shows the relation-
birds foraging,44 the PSO is a cluster ingenuity optimization ships of Np‑xylene with ϕ, VSA, and LCD. The trends are
algorithm with a simple concept: easy to implement without generally similar to the previous HTCS studies for CO2/CH4
too many parameters to adjust and widely used in function and CO2/N2 separation.18,20 At the operating conditions of
optimization, fuzzy systems, and BPNN training. In particular, this study (100 °C, 30 kPa), the adsorption of each xylene
the multiobjective PSO based on the Pareto mechanism is the isomer is close to saturation as shall be seen below.
most representative.45 In this work, an initial population of 100 Consequently, Np‑xylene is largely proportional to Vfree, which
D https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Figure 4. Predicted Np‑xylene and Sp/(m+o) by BPNN (a and b) and by DT (c and d) versus simulated Np‑xylene and Sp/(m+o).

has a monotonic relationship with ϕ (Figure S4); thus, Np‑xylene exhibit low Sp/(m+o). For instance, BAZGAM has Sp/(m+o) of
rises with increasing ϕ as shown in Figure 3a. Among 1502 nearly 1, and it is not good for the selective adsorption.
MOFs, 4 MOFs (BAZGAM, XAHQAA, XAHPUT, and Instead, three different MOFs (UXABUR, CADPII, and
NIBJAK) with ϕ of 0.95 possess the highest Np‑xylene > 10 JIVFUQ) with Sp/(m+o) > 45 are highlighted in Figure 3d−f.
mol/kg. Nevertheless, VSA is not monotonic with ϕ (Figure Among the three, UXABUR and CADPII possess small ϕ,
S4) or Vfree; therefore, Np‑xylene and VSA do not have a VSA, and LCD; however, JIVFUQ has a small LCD and
monotonic relationship as shown in Figure 3b. Np‑xylene first relatively larger ϕ and VSA. This unique MOF (JIVFUQ)
rises with VSA, then exhibits a maximal value when VSA is turns out to be the best for selective adsorption of p-xylene by
between 800 and 1600 m2/cm3, and finally drops upon further considering the trade-off between Np‑xylene and Sp/(m+o), as
increasing VSA. The reason is that a MOF with a low VSA can further discussed below.
only accommodate a few adsorbate molecules; on the other 3.2. Multivariate Analysis. The above univariate analysis
hand, in a MOF with a high VSA, the interaction with reveals that either Np‑xylene or Sp/(m+o) is not simply determined
adsorbate is weak and causes low adsorption capacity. With a single structural descriptor. Thus, ML and new evaluation
moderate VSA, the four MOFs (BAZGAM, XAHQAA, metrics were adopted for further analysis. An ideal MOF for
XAHPUT, and NIBJAK) exhibit high Np‑xylene. The trend of selective adsorption should possess both large Np‑xylene and high
Np‑xylene with LCD is also similar to that with VSA. The Sp/(m+o). Two ML methods (BPNN and DT) were used to
maximal Np‑xylene is seen for LCD ranging from 21.0 to 32.5 Å. perform multivariate analysis for Np‑xylene and Sp/(m+o) with five
Among the four MOFs, XAHQAA, XAHPUT, and NIBJAK descriptors (ϕ, VSA, and LCD in addition to PLD and ρ). As
have LCD within this range; nevertheless, BAZGAM is an shown in Figure 4, BPNN and DT exhibit similar performance
exception with a LCD of 43 Å. Intuitively, Np‑xylene in to predict Np‑xylene and Sp/(m+o). The predicted Np‑xylene by both
BAZGAM with such a large LCD would be low. However, it ML methods match closely with the simulated values; R2 and
is found that the organic linkers in BAZGAM are exclusively MAE are around 0.90−0.92 and 0.26−0.28, respectively.
aromatic rings (Figure S5), which induces strong π−π Nevertheless, the predicted Sp/(m+o) are not in good agreement
interaction with xylene and hence favorable adsorption. This with the simulated values, as both ML methods give R2 < 0.65.
is the reason why Np‑xylene in BAZGAM is the highest (∼17 On this basis, it is not feasible to perform accurate prediction
mol/kg) among all of the MOFs in the current study. for Sp/(m+o) by the two ML methods.
The relationships of Sp/(m+o) with ϕ, VSA, and LCD are Subsequently, the PSO was applied to optimize the
illustrated in Figure 3d−f. A similar trend is observed for the combination of Np‑xylene and Sp/(m+o). By assuming the weights
three descriptors. At a small ϕ, VSA, or LCD, Sp/(m+o) has a of Np‑xylene and Sp/(m+o) are ωN and ωS (ωN + ωS = 1), a trade-
high value. With increasing ϕ, VSA, or LCD, Sp/(m+o) drops and off variable F was calculated from
gradually approaches to 1. Apparently, a MOF with high ϕ,
VSA, or LCD is not favorable for the selective adsorption F= (ωN f1′ )2 + (ωSf 2′ )2
because the MOF is too porous or contains a large pore and is (2)
unable to effectively distinguish the isomers. This trend also
exists for the separation of CO2/CH4 and CO2/N2.18,20 f1 − f1min
Consequently, the four MOFs (BAZGAM, XAHQAA, f1′ =
XAHPUT, and NIBJAK) with high Np‑xylene in Figure 3a−c f1max − f1min (3)

E https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Figure 5. Predicted F′ versus calculated F by (a) BPNN and (b) DT. Color indicates the density contour for the number of MOFs.

Table 1. Importance of Descriptors in Different MOF Datasets by BPNN

importance
MOF data set Np‑xylene (mol/kg) and Sp/(m+o) PLD LCD ϕ ρ VSA
all 0.231 0.252 0.241 0.237 0.039
top 1000 Np‑xylene > 0.20 and Sp/(m+o) > 0.90 0.248 0.271 0.233 0.211 0.037
top 800 Np‑xylene > 0.26 and Sp/(m+o) > 1.00 0.281 0.270 0.191 0.213 0.045
top 500 Np‑xylene > 0.35 and Sp/(m+o) > 1.15 0.303 0.281 0.176 0.183 0.057
top 300 Np‑xylene > 0.44 and Sp/(m+o) > 1.30 0.355 0.283 0.161 0.157 0.044
top 100 Np‑xylene > 0.50 and Sp/(m+o) > 1.70 0.371 0.281 0.155 0.124 0.069

Figure 6. Simulated Sp/(m+o) versus Np‑xylene for xylene mixtures: (a) linear and (b) logarithmic scale of Sp/(m+o).

f2 − f2min Figure 4. This suggests that the appropriate combination of

f 2′ = Np‑xylene and Sp/(m+o) is helpful to achieve better prediction by
f2max − f2min (4)
the ML methods.
where f1 and f 2 are the Np‑xylene and Sp/(m+o) for each MOF, f1min It is instructive to examine how the separation performance
and f 2min are the minimal Np‑xylene and Sp/(m+o) among all of the is affected by each descriptor, which would in principle allow
MOFs, while f1max and f 2max are the maximal Np‑xylene and one to design a MOF by tuning a certain descriptor. To do so,
Sp/(m+o). Under the conditions of ωN + ωS = 1, ωN > 0, ωS > 0, the BPNN method was applied to analyze the weight factor of
and a precision of 0.01, a unique solution from a set of Pareto each descriptor. The trade-off variable F, as a dependent
solutions for ωN and ωS was obtained by repeating the variable, was predicted using 4 descriptors among ρ, ϕ, LCD,
calculations from either the BPNN or the DT method. In the PLD, and VSA as independent variables in the absence of one
BPNN method, the characteristic vector X = (X1, X2, X3, X4, descriptor. As shown in Figure S6, the predicted F′ cannot
and X5) was the five MOF descriptors (VSA, PLD, LCD, ϕ, match with the calculated values in the absence of ρ, ϕ, LCD,
and ρ). After normalization, the vector was used as the input of or PLD. Only in the absence of VSA is the prediction
the neural network toolbox with a predicted output F′ after satisfactory. On the basis of the RMSE, the importance of each
1000 iterations. In the DT, the same descriptors (VSA, PLD, 5
LCD, ϕ, and ρ) were used for the classification of DT training descriptor was estimated by RMSEi /∑i = 1 RMSE in six
to predict F′. After the PSO, ωN = 0.6 and ωS = 0.4 were different MOF data sets (all, top 1000, top 800, top 500, top
obtained to achieve the best fitting, especially for the BPNN 300, and top 100 MOFs). The five data sets of the top MOFs
method. The predicted F′ by both BPNN in Figure 5a and DT were based on different threshold values of Np‑xylene and
in Figure 5b agree well with the calculated F. The R2 values are Sp/(m+o), as listed in Table 1. In all data set, the importance of
0.94 and 0.90, respectively, which are higher than those in VSA is negligibly small. Due to the high correlations among
F https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Table 2. Seven Top-Performing MOFs

no. CSD Np‑xylene (mol/kg) Sp/(m+o) PLD (Å) LCD (Å) ϕ ρ (g/cm3) VSA (m2/cm3)
1 JIVFUQ 2.78 46.9 5.27 6.23 0.58 1.145 1435.97
2 GAYFOD 1.55 5.4 5.52 5.98 0.57 1.152 1241.19
3 RACZEC 1.17 16.2 5.05 5.80 0.32 1.516 683.47
4 CAVSUP 1.21 6.8 5.47 6.03 0.35 1.378 900.32
5 RACZIG 0.98 6.1 5.08 5.84 0.33 1.540 690.96
6 MIFPUO 0.66 5.9 5.88 6.70 0.38 1.135 757.47
7 RIGWIP 0.53 6.2 5.17 5.40 0.33 2.253 649.17

geometric descriptors, similar importance is observed as listed

in Table 1 for all of the descriptors except VSA. To reduce the
similarity, we recalculated the importance using smaller data
sets. In all and the top 1000 data sets, PLD, LCD, ϕ, and ρ
possess similar importance (0.21−0.27). With increasing
Np‑xylene and Sp/(m+o) from the top 1000 to the top 100, i.e.,
the number of data points drops, PLD and LCD become more
important, whereas ϕ and ρ are less important. In the top 100
data set, the importance of PLD and LCD reaches 0.37 and
0.28, respectively. By contrast, the importance of ϕ and ρ is
0.12−0.16, suggesting their less important role in governing
separation performance.
3.3. Top-Performing MOFs. After examining the depend- Figure 7. Effect of framework flexibility in seven top-performing
ence of the adsorption capacity Np‑xylene and selectivity Sp/(m+o) MOFs.
on MOF descriptors and finding the important descriptors, it is
interesting to identify the top-performing MOFs. Figure 6 compares both Np‑xylene and Sp/(m+o) in the seven MOFs
shows Sp/(m+o) versus Np‑xylene for a xylene isomer mixture at (JIVFUQ, GAYFOD, RACZEC, CAVSUP, RACZIG, MIF-
100 °C and 30 kPa. Most of the Sp/(m+o) is indeed close to 1 PUO, and RIGWIP). Np‑xylene in all seven MOFs is weakly
because the three isomers have very similar properties, leading affected by flexibility. For Sp/(m+o), the effect depends on
to low selectivity. Due to the smallest kinetic diameter among MOFs. In GAYFOD, Sp/(m+o) remains nearly the same. In four
the three isotherms, however, p-xylene can be preferentially MOFs (CAVSUP, RACZIG, MIFPUO, and RIGWIP), Sp/(m+o)
adsorbed in certain MOFs. Sp/(m+o) exhibits a sharp peak at low values are slightly reduced in the presence of flexibility. In two
Np‑xylene and then approaches 1 upon increasing Np‑xylene. A MOFs (JIVFUQ and RACZEC), the reductions are significant
similar trend was observed for the separation of CO2/CH4.20 (from 46.9 to 7.3 in JIVFUQ and from 16.2 to 6.8 in
The reason is that a large Np‑xylene exists in a MOF with a large RACZEC). Such a reduction in selectivity was observed
pore, which is not beneficial to separation. previously from a simulation study for xylene separation in
Several MOFs predicted to be promising for xylene MOF-48 and MIL-47.35 After taking flexibility into account,
separation in previous simulation studies are highlighted in JIVFUQ still appears to be the best among the seven MOFs; it
Figure 6. From 14 MOFs, MAF-X8, MIL-125, and CoBDP is superior to MAF-X8, MIL-125, CoBDP, MIL-140B, and
were identified by Torres-Knoop et al.13 at 160 °C and 100 MOF-48 as highlighted in Figure 6.
kPa. At the current operating conditions (100 °C and 30 kPa), Furthermore, the effect of the charge estimation method was
these three MOFs exhibit Np‑xylene ≈ 2 mol/kg and Sp/(m+o) < 4. examined. Among the seven top-performing MOFs, four
In another study, MIL-140B and MOF-48 were identified by MOFs (JIVFUQ, GAYFOD, RACZEC, and RACZIG) were
Gee et al. from ∼2500 MOFs at 50 °C and 9 bar.28 Here, MIL- simulated using the DDEC chargers.38 As shown in Figure S7,
140B and MOF-48 possess Np‑xylene < 1 mol/kg and Sp/(m+o) ≤ both Np‑xylene and Sp/(m+o) are close using the MEPO-QEq and
6 at 100 °C and 30 kPa. A handful of MOFs from this study DDEC charges. This indicates that the charge estimation
are observed to have either larger Np‑xylene or higher Sp/(m+o) method has no strong effect on the simulation results.
than the five MOFs (MAF-X8, MIL-125, CoBDP, MIL-140B, 3.4. Best MOF − JIVFUQ. JIVFUQ is identified to be the
and MOF-48). If we set Np‑xylene > 0.5 mol/kg and Sp/(m+o) > 5, best with the largest Np‑xylene and the highest Sp/(m+o). It is thus
seven top-performing MOFs can be identified. As listed in instructive to specifically examine the structural and adsorption
Table 2, the seven MOFs exhibit Np‑xylene from 0.53 to 2.78 properties in JIVFUQ. Figure 8a and 8b illustrates the
mol/kg and Sp/(m+o) from 5.90 and 46.89. The PLD and LCD structure of JIVFUQ, which contains 1D quasi-ellipse-shaped
in the seven MOFs are in a narrow region 5.05−5.88 and pores along the x axis. The adsorption snapshot of p-xylene in
5.40−6.70 Å, respectively; the ϕ and ρ are in a relatively larger JIVFUQ in Figure 8c shows almost all of the p-xylene
range 0.32−0.58 and 1.378−2.253 g/cm3; nevertheless, the molecules have a similar adsorption orientation in the pore. As
VSA is in a wide range 649.17−1435.97 m2/cm3. Therefore, plotted in Figure S8, the pores are uniform with a diameter of
the pore size is a significant structural descriptor to determine about 6.23 Å. The pore shape and diameter are suited to
separation performance, whereas the VSA has a marginal role. accommodate p-xylene rather than m- and o-xylene. On the
This is consistent with earlier results in Table 1 on the other hand, the kinetic diameters of p-, m-, and o-xylene are
importance of different descriptors. about 6.7, 7.1, and 7.4 Å (Table S1). Intuitively, none of the
In screening step 3, framework flexibility was incorporated in three isomers can enter into JIVFUQ with a PLD of 5.27 Å. As
the top-performing MOFs shortlisted from step 2. Figure 7 pointed out earlier, the kinetic diameter of the adsorbate is an
G https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

Figure 8. JIVFUQ structure on the (a) yz plane (b) xy plane. (c) Snapshot of p-xylene adsorbed in JIVFUQ.

Figure 9. (a) Simulated adsorption isotherms of pure xylenes in JIVFUQ at 100 °C. (b) Simulated adsorption capacities and selectivities of xylene
mixture in JIVFUQ at 100 °C.

empirical measure and generally larger than the actual viously in simulation studies for the adsorption of water/
molecular size. To determine if p-xylene with dk = 6.7 Å can alcohol mixtures in various MOFs.47,48
diffuse in JIVFUQ, a MD simulation was conducted for a
single p-xylene molecule in JIVFUQ at 100 °C. The MD 4. CONCLUSIONS
simulation started with an initial configuration of 30 p-xylene
In this study, HTCS and ML have been synergized to
molecules in JIVFUQ, but the intermolecular interactions investigate xylene separation by adsorption in MOFs. First,
between p-xylene molecules were switched off. The MD univariate analysis is conducted for the relationships between
simulation duration was 10 ns. A video from the MD structural descriptors (ϕ, VSA, LCD) and separation perform-
simulation is provided in the Supporting Information, showing ance metrics (adsorption capacity Np‑xylene and selectivity
that p-xylene molecule (in yellow) can diffuse in the pore of Sp/(m+o)). A set of optimal ϕ, VSA, and LCD is found for large
JIVFUQ. Such a phenomenon was also observed for the Np‑xylene or high Sp/(m+o). However, Np‑xylene or Sp/(m+o) is not
diffusion of CH4 and CO2 in other MOFs with PLD less than determined solely by a single structural descriptor. Through
the kinetic diameters of CH4 and CO2.46 multivariate analysis using two ML methods (BPNN and DT),
Figure 9a shows the adsorption isotherms of pure xylenes in the combination of Np‑xylene and Sp/(m+o) is optimized and the
JIVFUQ at 100 °C from simulations without considering importance of each descriptor is evaluated in six MOF data
framework flexibility. All three isomers exhibit a type-I sets. It is revealed that PLD and LCD have greater importance,
isotherm. At a given pressure, p-xylene exhibits a higher while VSA, ρ, and ϕ are less important in governing the
capacity than m- and o-xylene. The preferential adsorption of separation performance. By setting Np‑xylene > 0.5 mol/kg and
p-xylene is also seen for the xylene mixture shown in Figure 9b. Sp/(m+o) > 5, seven top-performing MOFs are identified. Among
The difference is that the capacity of p-xylene is substantially them, JIVFUQ is predicted to be the best after taking
higher than that of the other two isomers due to competitive framework flexibility into account. This MOF consists of quasi-
adsorption present in the mixture. In the low-pressure range (< ellipse-shaped pores and appears to outperform many MOFs
reported in the literature.
7 kPa), the capacity of p-xylene rises with increasing pressure;
While this study provides useful insights into xylene
meanwhile, the capacity of m- or o-xylene drops. Thus, the
separation in MOFs, there are several points to be noted.
preadsorbed m-/o-xylene is displaced by p-xylene. Beyond 7 First, the initial screening (step 2) is based on rigid MOFs
kPa, the capacity of each isomer becomes nearly constant. without considering framework flexibility. In principle, this is
Overall, the adsorption of the xylene mixture in JIVFUQ is acceptable because no generic transferable force field currently
dominated by p-xylene with negligible m- and o-xylene, and exists to accurately describe the flexibility of various MOFs;
high Sp/(m+o) is achieved up to 46.9. As observed in Figure 8c, moreover, simulation with flexibility is prohibitively expensive
the pores in JIVFUQ are quasi-ellipse shaped, which match and not currently suitable for screening. As demonstrated in
well with the structure of p-xylene; consequently, p-xylene fits screening step 3, however, framework flexibility may strongly
better in the pores than m-/o-xylene and is more preferentially affect the separation performance, particularly selectivity, and
absorbed. Such competitive adsorption was observed pre- thus should be ultimately incorporated. Second, the separation
H https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C

■
pubs.acs.org/JPCC Article

performance metrics are quantified by adsorption capacity and ACKNOWLEDGMENTS

selectivity, which can be readily calculated. More realistically,
We gratefully acknowledge the A*STAR under its AME IRG
process- and even system-level performance metrics should be
Grant (A20E5c0092), the National University of Singapore,
considered.49−51 Third, geometric properties are used here as
and the Ministry of Education of Singapore (R-279-000-574-
the descriptors for ML, which are mutually correlated to a
114, R-279-000-598-114, C-261-000-207-532/C-261-000-777-
certain extent and have limited design capability. For the future
532) for financial support. Z.Q. also thanks the National
ML study, structure-based descriptors are more appealing to be
Natural Science Foundation of China (Nos. 21978058 and
explored. Fourth, JIVFUQ is predicted to be the best MOF for
21676094).

■
xylene separation; however, it does not simply suggest this
MOF can readily work in reality. For example, the predicted
surface area of JIVFUQ is 1253.60 m2/g, much higher than the REFERENCES
experimentally measured 208.99 m2/g value.52 The reason is (1) Fabri, J.; Graeser, U.; Simo, T. A. Xylenes: Ullmann’s Encyclopedia
that the MOF was only partially activated due to the presence of Industrial Chemistry; Wiley-VCH Verlag GmbH & Co: New York,
of solvent and ligand molecules. It is important to consider 2000.
these points highlighted here in future computational studies (2) Cannella, W. J. Xylenes and Ethylbenzene: Kirk-Othmer
for xylene separation or other separation in MOFs. Encyclopedia of Chemical Technology; John Wiley & Sons: New

■
York, 2000.
ASSOCIATED CONTENT (3) Sholl, D. S.; Lively, R. P. Seven Chemical Separations to Change
the World. Nature 2016, 532, 435−437.
*
sı Supporting Information
(4) Yang, Y.; Bai, P.; Guo, X. Separation of Xylene Isomers: A
The Supporting Information is available free of charge at Review of Recent Advances in Materials. Ind. Eng. Chem. Res. 2017,
https://pubs.acs.org/doi/10.1021/acs.jpcc.0c10773. 56, 14725−14753.
Physical properties of xylene isomers; Lennard−Jones (5) Furukawa, H.; Cordova, K. E.; O’Keeffe, M.; Yaghi, O. M. The
Chemistry and Applications of Metal-Organic Frameworks. Science
potential parameters; machine-learning methods; par- 2013, 341, 1230444.
ticle swarm optimization; relationships of Vfree ∼ ϕ and (6) Peralta, D.; Barthelet, K.; Pérez-Pellitero, J.; Chizallet, C.;
VSA ∼ ϕ; atomic structures of BAZGAM; predicted F’ Chaplais, G.; Simon-Masseron, A.; Pirngruber, G. D. Adsorption and
by BPNN versus calculated F; Np‑xylene and Sp/(m+o) in 4 Separation of Xylene Isomers: CPO-27-Ni vs HKUST-1 vs NaY. J.
top-performing MOFs using the MEPO-Qeq and DDEC Phys. Chem. C 2012, 116, 21844−21855.
charges; pore size distribution in JIVFUQ (PDF) (7) Gonzalez, M. I.; Kapelewski, M. T.; Bloch, E. D.; Milner, P. J.;
Simulation data in 1502 CoRE-MOFs (XLSX) Reed, D. A.; Hudson, M. R.; Mason, J. A.; Barin, G.; Brown, C. M.;
Video for the diffusion of p-xylene in JIVFUQ (AVI) Long, J. R. Separation of Xylene Isomers through Multiple Metal Site

■
Interactions in Metal-Organic Frameworks. J. Am. Chem. Soc. 2018,
140, 3412−3422.
AUTHOR INFORMATION (8) Jin, Z.; Zhao, H.-Y.; Zhao, X.-J.; Fang, Q.-R.; Long, J. R.; Zhu,
Corresponding Author G.-S. A Novel Microporous MOF with the Capability of Selective
Jianwen Jiang − Department of Chemical and Biomolecular Adsorption of Xylenes. Chem. Commun. 2010, 46, 8612−8614.
Engineering, National University of Singapore, Singapore (9) Lannoeye, J.; Van de Voorde, B.; Bozbiyik, B.; Reinsch, H.;
Denayer, J.; De Vos, D. An Aliphatic Copper Metal-Organic
117576, Singapore; orcid.org/0000-0003-1310-9024; Framework as Versatile Shape Selective Adsorbent in Liquid Phase
Email: [email protected] Separations. Microporous Mesoporous Mater. 2016, 226, 292−298.
(10) El Osta, R.; Carlin-Sinclair, A.; Guillou, N.; Walton, R. I.;
Authors
Vermoortele, F.; Maes, M.; de Vos, D.; Millange, F. Liquid-Phase
Zhiwei Qiao − Department of Chemical and Biomolecular Adsorption and Separation of Xylene Isomers by the Flexible Porous
Engineering, National University of Singapore, Singapore MIL-53(Fe). Chem. Mater. 2012, 24, 2781−2791.
117576, Singapore; Guangzhou Key Laboratory for New (11) Zhang, K.; Lively, R. P.; Zhang, C.; Chance, R. R.; Koros, W. J.;
Energy and Green Catalysis, School of Chemistry and Sholl, D. S.; Nair, S. Exploring the Framework Hydrophobicity and
Chemical Engineering, Guangzhou University, Guangzhou Flexibility of ZIF-8: From Biofuel Recovery to Hydrocarbon
510006, China; orcid.org/0000-0002-1264-3762 Separations. J. Phys. Chem. Lett. 2013, 4, 3618−3622.
Yaling Yan − Guangzhou Key Laboratory for New Energy and (12) Polyukhov, D. M.; Poryvaev, A. S.; Gromilov, S. A.; Fedin, M.
Green Catalysis, School of Chemistry and Chemical V. Precise Measurement and Controlled Tuning of Effective Window
Engineering, Guangzhou University, Guangzhou 510006, Sizes in ZIF-8 Framework for Efficient Separation of Xylenes. Nano
China Lett. 2019, 19, 6506−6510.
(13) Torres-Knoop, A.; Krishna, R.; Dubbeldam, D. Separating
Yaxing Tang − Guangzhou Key Laboratory for New Energy Xylene Isomers by Commensurate Stacking of p-Xylene within
and Green Catalysis, School of Chemistry and Chemical Channels of MAF-X8. Angew. Chem., Int. Ed. 2014, 53, 7774−7778.
Engineering, Guangzhou University, Guangzhou 510006, (14) Granato, M. A.; Martins, V. D.; Ferreira, A. F. P.; Rodrigues, A.
China E. Adsorption of Xylene Isomers in UiO-66 by Molecular Simulation.
Hong Liang − Guangzhou Key Laboratory for New Energy Microporous Mesoporous Mater. 2014, 190, 165−170.
and Green Catalysis, School of Chemistry and Chemical (15) Agrawal, M.; Bhattacharyya, S.; Huang, Y.; Jayachandrababu, K.
Engineering, Guangzhou University, Guangzhou 510006, C.; Murdock, C. R.; Bentley, J. A.; Rivas-Cardona, A.; Mertens, M.
China M.; Walton, K. S.; Sholl, D. S.; Nair, S. Liquid-Phase Multicomponent
Adsorption and Separation of Xylene Mixtures by Flexible MIL-53
Complete contact information is available at: Adsorbents. J. Phys. Chem. C 2018, 122, 386−397.
https://pubs.acs.org/10.1021/acs.jpcc.0c10773 (16) Wu, X.; Wei, W.; Jiang, J. W.; Caro, J.; Huang, A. High-Flux
High-Selectivity Metal-Organic Framework MIL-160 Membrane for
Notes Xylene Isomer Separation by Pervaporation. Angew. Chem., Int. Ed.
The authors declare no competing financial interest. 2018, 57, 15354−15358.

I https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX
The Journal of Physical Chemistry C pubs.acs.org/JPCC Article

(17) Kim, S.-I.; Lee, S.; Chung, Y. G.; Bae, Y.-S. The Origin of P- Membrane Separation of a CO2/N2/CH4 Mixture. J. Mater. Chem. A
Xylene Selectivity in a Dabco Pillar-Layered Metal-Organic Frame- 2016, 4, 15904−15912.
work: A Combined Experimental and Computational Investigation. (35) Gee, J. A.; Sholl, D. S. Effect of Framework Flexibility on C8
ACS Appl. Mater. Interfaces 2019, 11, 31227−31236. Aromatic Adsorption at High Loadings in Metal-Organic Frame-
(18) Wilmer, C. E.; Farha, O. K.; Bae, Y.-S.; Hupp, J. T.; Snurr, R. Q. works. J. Phys. Chem. C 2016, 120, 370−376.
Structure-Property Relationships of Porous Materials for Carbon (36) Rappe, A. K.; Casewit, C. J.; Colwell, K. S.; Goddard, W. A.;
Dioxide Separation and Capture. Energy Environ. Sci. 2012, 5, 9849− Skiff, W. M. UFF, a Full Periodic Table Force Field for Molecular
9856. Mechanics and Molecular Dynamics Simulations. J. Am. Chem. Soc.
(19) Chung, Y. G.; Gómez-Gualdrón, D. A.; Li, P.; Leperi, K. T.; 1992, 114, 10024−10035.
Deria, P.; Zhang, H.; Vermeulen, N. A.; Stoddart, J. F.; You, F.; Hupp, (37) Kadantsev, E. S.; Boyd, P. G.; Daff, T. D.; Woo, T. K. Fast and
J. T.; Farha, O. K.; Snurr, R. Q. In Silico Discovery of Metal-Organic Accurate Electrostatics in Metal Organic Frameworks with a Robust
Frameworks for Precombustion CO2 Capture Using a Genetic Charge Equilibration Parameterization for High-Throughput Virtual
Algorithm. Sci. Adv. 2016, 2, No. e1600909. Screening of Gas Adsorption. J. Phys. Chem. Lett. 2013, 4, 3056−3061.
(20) Qiao, Z.; Zhang, K.; Jiang, J. W. In Silico Screening of 4764 (38) Nazarian, D.; Camp, J. S.; Sholl, D. S. A Comprehensive Set of
Computation-Ready, Experimental Metal-Organic Frameworks for High-Quality Point Charges for Simulations of Metal-Organic
CO2 Separation. J. Mater. Chem. A 2016, 4, 2105−2114. Frameworks. Chem. Mater. 2016, 28, 785−793.
(21) Qiao, Z.; Xu, Q.; Jiang, J. W. Computational Screening of (39) Finsy, V.; Verelst, H.; Alaerts, L.; De Vos, D.; Jacobs, P. A.;
Hydrophobic Metal-Organic Frameworks for the Separation of H2S Baron, G. V.; Denayer, J. F. M. Pore-Filling-Dependent Selectivity
and CO2 from Natural Gas. J. Mater. Chem. A 2018, 6, 18898−18905. Effects in the Vapor-Phase Separation of Xylene Isomers on the
(22) Boyd, P. G.; Chidambaram, A.; Garcia-Diez, E.; Ireland, C. P.; Metal-Organic Framework MIL-47. J. Am. Chem. Soc. 2008, 130,
7110−7118.
Daff, T. D.; Bounds, R.; Gładysiak, A.; Schouwink, P.; Moosavi, S. M.;
(40) Shi, W.; Maginn, E. J. Continuous Fractional Component
Maroto-Valer, M. M.; Reimer, J. A.; Navarro, J. A. R.; Woo, T. K.;
Monte Carlo: An Adaptive Biasing Method for Open System
Garcia, S.; Stylianou, K. C.; Smit, B. Data-Driven Design of Metal-
Atomistic Simulations. J. Chem. Theory Comput. 2007, 3, 1451−1463.
Organic Frameworks for Wet Flue Gas CO2 Capture. Nature 2019, (41) Menon, A.; Mehrotra, K.; Mohan, C. K.; Ranka, S.
576, 253−256. Characterization of a Class of Sigmoid Functions with Applications
(23) Fernandez, M.; Boyd, P. G.; Daff, T. D.; Aghaji, M. Z.; Woo, T. to Neural Networks. Neural Networks 1996, 9, 819−835.
K. Rapid and Accurate Machine Learning Recognition of High (42) Moré, J. J. The Levenberg-Marquardt Algorithm: Implementa-
Performing Metal-Organic Frameworks for CO2 Capture. J. Phys. tion and Theory. In Numerical Analysis. Lecture Notes in Mathematics;
Chem. Lett. 2014, 5, 3056−3060. Watson, G. A., Ed.; Springer: Berlin, Heidelberg, 1978; Vol 630.
(24) Moosavi, S. M.; Nandy, A.; Jablonka, K. M.; Ongari, D.; Janet, (43) Yeh, C.-H. Classification and Regression Trees. Chemom. Intell.
J. P.; Boyd, P. G.; Lee, Y.; Smit, B.; Kulik, H. J. Understanding the Lab. Syst. 1991, 12, 95−96.
Diversity of the Metal-Organic Framework Ecosystem. Nat. Commun. (44) Kennedy, J.; Eberhart, R. C. Particle Swarm Optimization.
2020, 11, 4068. Proceedings of IEEE International Conference on Neural Networks,
(25) Yao, Z.; Sánchez-Lengeling, B.; Bobbitt, N. S.; Bucior, B. J.; Piscataway, NJ, USA; IEEE, 1995; pp 1942−1948.
Kumar, S. G. H.; Collins, S. P.; Burns, T.; Woo, T. K.; Farha, O. K.; (45) Wang, Y.; Cai, Z. Combining Multiobjective Optimization with
Snurr, R. Q.; Aspuru-Guzik, A. Inverse Design of Nanoporous Differential Evolution to Solve Constrained Optimization Problems.
Crystalline Reticular Materials with Deep Generative Models. Nat. IEEE Trans. Evol. Comput. 2012, 16, 117−134.
Machine Intelligence 2021, 3, 76−86. (46) Haldoupis, E.; Nair, S.; Sholl, D. S. Efficient Calculation of
(26) Zhang, X.; Zhang, K.; Yoo, H.; Lee, Y. Machine Learning- Diffusion Limitations in Metal Organic Framework Materials: A Tool
Driven Discovery of Metal-Organic Frameworks for Efficient CO2 for Identifying Materials for Kinetic Separations. J. Am. Chem. Soc.
Capture in Humid Condition. ACS Sustainable Chem. Eng. 2021, 9, 2010, 132, 7528−7539.
2872−2879. (47) Nalaparaju, A.; Zhao, X. S.; Jiang, J. W. Molecular
(27) Jablonka, K. M.; Ongari, D.; Moosavi, S. M.; Smit, B. Big-Data Understanding for the Adsorption of Water and Alcohols in
Science in Porous Materials: Materials Genomics and Machine Hydrophilic and Hydrophobic Zeolitic Metal-Organic Frameworks.
Learning. Chem. Rev. 2020, 120, 8066−8129. J. Phys. Chem. C 2010, 114, 11542−11550.
(28) Gee, J. A.; Zhang, K.; Bhattacharyya, S.; Bentley, J.; Rungta, M.; (48) Nalaparaju, A.; Zhao, X. S.; Jiang, J. W. Biofuel Purification by
Abichandani, J. S.; Sholl, D. S.; Nair, S. Computational Identification Pervaporation and Vapor Permeation in Metal-Organic Frameworks:
and Experimental Evaluation of Metal-Organic Frameworks for A Computational Study. Energy Environ. Sci. 2011, 4, 2107−2116.
Xylene Enrichment. J. Phys. Chem. C 2016, 120, 12075−12082. (49) Nalaparaju, A.; Khurana, M.; Farooq, S.; Karimi, I. A.; Jiang, J.
(29) Chung, Y. G.; Camp, J.; Haranczyk, M.; Sikora, B. J.; Bury, W.; W. CO2 Capture in Cation-Exchanged Metal-Organic Frameworks:
Krungleviciute, V.; Yildirim, T.; Farha, O. K.; Sholl, D. S.; Snurr, R. Q. Holistic Modeling from Molecular Simulation to Process Optimiza-
Computation-Ready, Experimental Metal-Organic Frameworks: A tion. Chem. Eng. Sci. 2015, 124, 70−78.
Tool to Enable High-Throughput Screening of Nanoporous Crystals. (50) Park, J.; Rubiera Landa, H. O.; Kawajiri, Y.; Realff, M. J.; Lively,
Chem. Mater. 2014, 26, 6185−6192. R. P.; Sholl, D. S. How Well Do Approximate Models of Adsorption-
(30) Willems, T. F.; Rycroft, C. H.; Kazi, M.; Meza, J. C.; Based CO2 Capture Processes Predict Results of Detailed Process
Haranczyk, M. Algorithms and Tools for High-Throughput Models? Ind. Eng. Chem. Res. 2020, 59, 7097−7108.
Geometry-Based Analysis of Crystalline Porous Materials. Micro- (51) Yancy-Caballero, D.; Leperi, K. T.; Bucior, B. J.; Richardson, R.
porous Mesoporous Mater. 2012, 149, 134−141. K.; Islamoglu, T.; Farha, O. K.; You, F.; Snurr, R. Q. Process-Level
(31) Dubbeldam, D.; Calero, S.; Ellis, D. E.; Snurr, R. Q. Raspa: Modelling and Optimization to Evaluate Metal-Organic Frameworks
Molecular Simulation Software for Adsorption and Diffusion in for Postcombustion Capture of CO2. Mol. Syst. Des. Eng. 2020, 5,
1205−1218.
Flexible Nanoporous Materials. Mol. Simul. 2016, 42, 81−101.
(52) Liang, J.; Shimizu, G. K. H. Crystalline Zinc Diphosphonate
(32) Bellat, J.-P.; Simonot-Grange, M.-H.; Jullian, S. Adsorption of
Metal-Organic Framework with Three-Dimensional Microporosity.
Gaseous p-Xylene and m-Xylene on NaY, KY, and Bay Zeolites.
Inorg. Chem. 2007, 46, 10449−10451.
Zeolites 1995, 15, 124−130.
(33) Liu, J.; Dong, M.; Sun, Z.; Qin, Z.; Wang, J. Sorption of Xylenes
in CoAlPO4-5 Molecular Sieves. Colloids Surf., A 2004, 247, 41−45.
(34) Qiao, Z.; Peng, C.; Zhou, J.; Jiang, J. W. High-Throughput
Computational Screening of 137953 Metal-Organic Frameworks for

J https://doi.org/10.1021/acs.jpcc.0c10773
J. Phys. Chem. C XXXX, XXX, XXX−XXX