Metal-organic frameworks (MOFs) and covalent organic frameworks (COFs) have great potential to be used as porous adsorbents and membranes to achieve high-performance methane purification. Although the continuous increase in the number and diversity of MOFs and COFs is a great opportunity for the discovery of novel adsorbents and membranes with superior performances, evaluating such a vast number of materials in the quickest and most effective manner requires the development of computational approaches. High-throughput computational screening based on molecular simulations has been extensively used to identify the most promising MOFs and COFs for methane purification. However, the enormous and ever-growing material space necessitates more efficient approaches in terms of time and effort. Combining data science with molecular simulations has recently accelerated the discovery of optimal MOF and COF materials for methane purification and revealed the hidden structure–performance relationships. In this perspective, we highlighted the recent developments in combining high-throughput molecular simulations and machine learning to accurately identify the most promising MOF and COF adsorbents and membranes among thousands of candidates for separating methane from other gases including acetylene, carbon dioxide, helium, hydrogen, and nitrogen. After providing a brief overview of the topic, we reviewed the pioneering contributions in the field and discussed the current opportunities and challenges that we need to direct our efforts for the design and discovery of adsorbent and membrane materials.

Methane (CH4) is the primary component of natural gas, and its separation from other impurity gases is important to use it as a clean energy source.1 Natural gas is extracted from various sources like shale gas, trapped within formations of fine-grained sedimentary rocks, and landfill gas, produced through the natural degradation of organic waste in landfills. Shale gas and landfill gas often contain a substantial amount of nitrogen (N2) and carbon dioxide (CO2), which should be efficiently separated from CH4 to increase the heating value of the natural gas.2 Separation of CH4 from hydrogen (H2) is important for pure H2 recovery since they exist together in the refinery off-gas mixtures.3 Conventional separation technologies, such as cryogenic distillation and chemical absorption, have been widely used for the separation of CH4 from other gases, but these techniques are known for their high energy consumption and associated high cost.4 

Adsorption and membrane-based gas separation using porous materials as adsorbents and membranes have emerged as strong alternatives thanks to their lower energy demands, environmentally friendly nature, and ease of operation. Adsorption-based gas separation is a process in which a gas mixture is separated based on the differences in the adsorption/desorption characteristics of the gases in the mixture, which are mainly driven by the physicochemical properties of the adsorbent, a solid porous material.5 Membrane-based gas separation utilizes porous materials as semi-permeable barriers to selectively allow certain gases to pass through the membrane, the permeate, while restricting the passage of others, the retentate.6 In this process, gases first adhere to the surface of the membrane and then diffuse where separation happens because of the differences in their transport rates through the pores. For both adsorption and membrane-based gas separations, selecting the porous material is the most important step to achieve high performance. A large number and variety of porous materials, such as zeolites, activated carbons, and porous polymers, have been used as adsorbents and membranes to date.7 

Metal-organic frameworks (MOFs) are a relatively new class of crystalline porous materials composed of metal nodes and organic linkers as represented in Fig. 1(a). The primary benefit of MOFs over other traditional porous materials is the ability to generate many structures by changing the combinations of metals and organic linkers during the synthesis to obtain materials with a wide variety of physical and chemical properties. MOFs offer exceptionally large surface areas (up to 10 000 m2/g),8 high porosities (up to 0.9), various pore sizes (3–100 Å),9 and tailorable physicochemical properties making them promising for many applications such as gas storage and separation,6,10,11 catalysis,12,13 chemical sensing,14 drug storage and delivery,15,16 and light harvesting.17 MOFs synthesized to date have been deposited into the Cambridge Structural Database (CSD),18 which is a data center for all the organic and metal-organic experimental crystal structures. Our search in the CSD using ConQuest19 software resulted in 123 457 different types of MOFs as of January 2024. The exploration of MOF structures goes beyond the realm of experimentally synthesized structures, as researchers have recently focused on generating hypothetical, computer-generated MOFs (hMOFs) by combining different types of metals, linkers, and topologies together and hundreds of thousands of hMOFs have been generated.20,21 Covalent-organic frameworks (COFs), on the other hand, are a class of materials consisting of covalently bonded linkers, and they have been reported as chemically robust structures possessing high thermal and mechanical stability.22 As of January 2024, 1242 synthesized COF structures23 in addition to several thousands of hypothetical COFs24 have been reported and studied for various gas separations.25 

FIG. 1.

(a) Descriptors of MOFs. (b) Representative workflow of harnessing molecular simulations with ML to study a large number of MOF adsorbents and membranes for CH4 purification.

FIG. 1.

(a) Descriptors of MOFs. (b) Representative workflow of harnessing molecular simulations with ML to study a large number of MOF adsorbents and membranes for CH4 purification.

Close modal

Several experimental studies have demonstrated the great promise of MOFs and COFs in CH4 purification applications and reported that they can be used as adsorbents and membranes for the separation of undesired gases such as CO2 and N2 from CH4 by outperforming traditional zeolites.26–30 However, experimental synthesis, characterization, and testing of every single MOF material for a target gas separation require extensive resources, time, and effort considering the vast spectrum of this material space. Computational studies have recently played a significant role in efficiently identifying the most promising MOF materials among thousands of candidates to direct the experimental efforts to these useful structures.31–33 High-throughput computational screening (HTCS) studies generally use Grand Canonical Monte Carlo (GCMC) simulations to mimic the experimental gas adsorption behavior of porous materials and Molecular Dynamics (MD) simulations to compute the diffusivity of gas molecules in the pores. These simulations provide molecular insights into the adsorption and membrane-based gas separation mechanisms of MOFs and help to establish structure–performance relations to guide the design of high-performing adsorbents and membranes. For example, adsorption-based separation of the CH4/N2 mixture was investigated by screening 5109 synthesized MOFs using GCMC simulations and results showed a significant correlation between the adsorption energy densities of CH4 molecules inside the MOFs and their CH4/N2 selectivities (up to 20.9),34 which are higher than those of zeolites such as zeolite 5 A (1.5).35 Membrane-based separation performances of 572 synthesized COFs were examined for CH4/N2 and CH4/H2 mixtures by using GCMC and MD simulations, and COF membranes were found to offer higher H2 permeabilities compared to MOFs and polymers.36 

With the continuous and exponential growth in the number of MOFs and COFs, even the HTCS approach faces limitations in comprehensively exploring the vast material spectrum to find promising adsorbent and membrane candidates for CH4 purification. Efficient and cost-effective material research now requires data-driven methods to reduce the need for conducting molecular simulations for every single material. Machine learning (ML), a discipline of artificial intelligence where machines automatically learn from the data to make accurate predictions with minimal human intervention,37 has been recently utilized in materials science to predict various properties of polymers,38 quartz-type materials with porous interface,39 and thermoelectric materials.40 The number of studies harnessing the data obtained from HTCS of MOFs with ML to make accurate and fast predictions for the properties and performances of MOFs in various gas adsorption and separation applications is increasing very rapidly.41,42

The common descriptors of MOFs used as input data in developing ML models are shown in Fig. 1(a), which include structural (pore size, surface area, pore volume, density, and porosity), chemical [element percentages, total degree of unsaturation (TDU), revised autocorrelation functions (RACs) that encode relationships of the atomic properties such as size, connectivity, and electronegativity on a molecular graph,43 and atomic property weighted radial distribution function (AP-RDF)], and energy-based [Henry's constant (KH), isosteric heat of adsorption (Qst), and energy histograms] features. Figure 1(b) represents the workflow generally used in ML-integrated HTCS studies on MOFs. A MOF database is first selected and then filtered based on the structural properties such as pore sizes to eliminate materials having very low surface areas or inappropriate pore sizes for the adsorption/diffusion of the molecules that exist in the mixture of separation of interest. This narrow-downed material set is then studied by molecular simulations where GCMC and/or MD simulations are performed to compute the adsorption and/or diffusion properties of gases in MOFs. The gas adsorption/diffusion properties obtained from simulations can be directly employed as the target data for ML, or these properties can be used to calculate adsorbent and membrane performance evaluation metrics (such as selectivity and permeability) and then used as the target data for ML. The developed ML models are finally tested and used to predict the gas adsorption and separation properties of a large number and variety of MOFs. For a more detailed explanation and discussion on the principles of ML, feature selection approaches and ML methods, the readers are directed to a comprehensive recent review article.44 

We see the topic of CH4 purification using MOFs and COFs as one of the fields that can have a great benefit from the ML approaches because (a) the number/variety of materials is too large to study each material individually by using brute-force molecular simulations and trial-and-error experiments; (b) even if a smaller number of materials is considered, making membranes from crystal materials is challenging and membrane simulations based on MD require significant computational time; and (c) CH4 purification involves several different separations including CH4/N2, CH4/H2, and CO2/CH4 and requires extensive studies on each gas mixture at various operating conditions. ML has the power to provide fast and accurate solutions to these challenges by extracting useful material property–performance relations that are not possible to acquire without the usage of data science. In this perspective, we focused on harnessing ML and HTCS to accelerate the design and discovery of novel MOF adsorbents and membranes for CH4 purification. After highlighting the importance of the specific gas separation, we discussed why and how ML can be useful in studying MOF adsorbents and membranes. The main contributions of the recent studies were discussed, and potential opportunities and challenges in this field were given from our perspective.

CH4/N2 separation is important in the process of upgrading natural gas and preventing the combustion of N2, which can produce harmful nitrogen oxides (NOx). CH4 is generally more strongly adsorbed into porous materials than N2. Performing molecular simulations for CH4/N2 separation is computationally costly since CH4 and N2 have similar physicochemical properties, resulting in competitive adsorption for available adsorption sites and subsequently extending the simulation time to reach equilibration. In the pursuit of quickly and accurately identifying the most promising MOF adsorbents for CH4/N2 separation, our group recently performed molecular simulations for 4612 synthesized MOFs and subsequently developed ML models to predict CH4/N2 mixture adsorption data based on the pore size, surface area, pore volume, porosity, element percentages, and simulated Henry's constants of gases in MOFs.45 As a result, adsorption-based CH4/N2 separation performance limits of a very large material spectrum, 4612 synthesized MOFs and 98 601 hMOFs, were revealed for the first time, and the existence of many promising MOF adsorbents having superior separation performance compared to zeolites and activated carbons was shown.

Structural descriptors of MOFs mentioned above have been generally used for the development of ML models. In contrast to the classical ML approach, deep learning can bypass the laborious and time-consuming process of descriptor selection. Motivated from this, a deep learning model that can predict CH4/N2 selectivities of more than 100 000 MOFs based on the string-based descriptors, building block representations of MOFs, that include both the chemical and structural properties was recently developed and shown to offer higher prediction accuracy than the classical ML models.46 To explore the effect of chemical features such as topology and metal types of MOFs on their CH4/N2 selectivities, T-SNE (t-distributed Stochastic Neighbor Embedding), which is an unsupervised non-linear dimensionality reduction technique for data exploration and visualizing, was used as shown in Fig. 2. The data points in the latent space showcase a gradient distribution of selectivity, which gradually rises from the upper left to the lower right [Fig. 2(a)]. Distinct clustering patterns of MOFs' topologies [Fig. 2(b)], in contrast to the dispersed distribution of metal types, particularly for [Cu][Cu] and [Fe][Fe] clusters [Fig. 2(c)], highlighted the critical importance of MOFs' topology for predicting their CH4/N2 selectivities.

FIG. 2.

Dimensionality reduction visualization of T-SNE for latent space colored based on (a) CH4/N2 selectivity, (b) topology, and (c) metal type of 113 160 MOFs. The most common topologies and metal nodes were depicted with different colors. Reproduced with permission from Li et al., ACS Appl. Mater. Interfaces 15(51), 59887–59894 (2023). Copyright 2023 American Chemical Society.46 

FIG. 2.

Dimensionality reduction visualization of T-SNE for latent space colored based on (a) CH4/N2 selectivity, (b) topology, and (c) metal type of 113 160 MOFs. The most common topologies and metal nodes were depicted with different colors. Reproduced with permission from Li et al., ACS Appl. Mater. Interfaces 15(51), 59887–59894 (2023). Copyright 2023 American Chemical Society.46 

Close modal

Studies we reviewed so far trained the ML models using the results of GCMC simulations to predict the separation performance of MOFs at specific operating conditions. An ML model can be more useful if it can be transferable to different gases, compositions, and operating conditions on which it was not trained. With this aim, a multipurpose model (MLP) was developed to predict single-component uptakes of 51 520 hMOFs for various gases including N2 and CH4, and coupled with the ideal adsorbed solution theory (IAST) to predict their binary mixture adsorption data.47 This approach was shown to efficiently identify the high-performing MOF adsorbents for N2 separation from CH4, which is important for natural gas upgrading particularly in cases where deposits contain N2 content >10%. Figure 3(a) shows that CH4 and N2 adsorption data obtained from simulations closely match the predictions of the MLP+IAST approach for 1000 hMOFs. Figure 3(b) demonstrates the reasonable agreement between GCMC simulations and MLP+IAST calculations for an adsorbent performance evaluation metric called separation potential, Δ Q N 2 / CH 4. The separation potential values of the hMOF database were predicted to identify the promising materials for N2/CH4 separation at a significantly lower computational time than simulations.

FIG. 3.

The comparison between MLP+IAST predictions and GCMC results of (a) CH4 and N2 loadings and (b) Δ Q N 2 / CH 4 for N2/CH4 mixture for ten pressure points between 1 and 10 bar, at 298 K. The plots are colored by point density. Reproduced from Anderson and Gómez-Gualdrón, J. Chem. Phys. 154(23), 234102 (2021), with the permission of AIP Publishing.47 

FIG. 3.

The comparison between MLP+IAST predictions and GCMC results of (a) CH4 and N2 loadings and (b) Δ Q N 2 / CH 4 for N2/CH4 mixture for ten pressure points between 1 and 10 bar, at 298 K. The plots are colored by point density. Reproduced from Anderson and Gómez-Gualdrón, J. Chem. Phys. 154(23), 234102 (2021), with the permission of AIP Publishing.47 

Close modal

Efficient CH4/H2 separation is vital in processes like steam reforming of methane or water-gas shift reactions for large-scale hydrogen production. In CH4/H2 separation, CH4 is the primarily adsorbed component because of its stronger van der Waals interactions with the porous material. Our group recently studied 597 synthesized COFs and 69 840 hCOFs for CH4/H2 separation by harnessing HTCS and ML.48 GCMC simulations were performed to compute mixture adsorption data for 597 synthesized COFs. The most promising COFs were identified based on the highest adsorbent performance score (APS) and regenerability (R%) >80%, and the hCOF database was filtered to identify 7737 hCOFs that have similar structural properties to the most promising synthesized COFs. ML models were then developed to predict CH4 and H2 mixture adsorption data and then transferred to the remaining unseen classes of hCOFs to accurately predict their CH4/H2 separation performances. Figure 4 demonstrates the CH4/H2 selectivity and CH4 working capacity of all 69 822 hCOFs at pressure swing adsorption (PSA) and vacuum swing adsorption (VSA) conditions. The selectivity ranges are comparable, but hCOFs exhibit higher working capacities under PSA condition. Performing molecular simulations for such a vast number of hCOFs would be impractical due to the extensive computational time requirement. The integration of molecular simulations with ML models enabled the rapid discovery of the hCOF materials space within seconds.

FIG. 4.

CH4/H2 separation performances of the hCOF materials space predicted by ML models for (a) PSA and (b) VSA processes. Reproduced with permission from Aksu and Keskin, J. Mater. Chem. A 11(27), 14788–14799 (2023). Copyright 2023 The Royal Society of Chemistry.48 

FIG. 4.

CH4/H2 separation performances of the hCOF materials space predicted by ML models for (a) PSA and (b) VSA processes. Reproduced with permission from Aksu and Keskin, J. Mater. Chem. A 11(27), 14788–14799 (2023). Copyright 2023 The Royal Society of Chemistry.48 

Close modal

Selective separation of CO2 from CH4 is crucial to mitigate greenhouse gas emissions and to increase the viability of natural gas as a clean energy source. CO2 is more strongly adsorbed into MOFs since it has a quadrupolar moment that leads to strong electrostatic interactions with the adsorbent in addition to the van der Waals interactions. HTCS of MOFs for CO2/CH4 mixture adsorption is very costly since (i) computing these electrostatic interactions requires the assignment of partial charges to MOF atoms through quantum chemical calculations that are highly expensive and (ii) there is a competitive adsorption between CO2 and CH4 molecules for the available adsorption sites and achieving thermodynamic equilibrium through simulations can be time-consuming. ML can act not only as a surrogate simulation to compute CO2/CH4 adsorption and separation properties of a large number of MOFs but also to do the inverse design of new materials with improved CO2/CH4 separation performance through generative algorithms. In a pioneer work,49 simulated CO2/CH4 mixture adsorption data of 45 000 hMOFs were fed to a deep learning algorithm along with the materials' structural features to generate novel hMOF structures. Generated MOFs with improved CO2 uptakes (7.5 mol/kg) and CO2/CH4 selectivities (16.0) were reported to have comparable or better performance than previously reported MOFs and zeolites.

Acetylene (C2H2) is often produced by the partial combustion of CH4, and C2H2/CH4 separation is important for the recovery and reuse of both gases, contributing to resource efficiency in the industrial processes. C2H2 is more selectively adsorbed than CH4, and molecular simulations require a significant amount of time due to the electrostatic interactions between C2H2 and MOF atoms. Our group recently combined HTCS and ML to study 936 anion-pillared MOFs to evaluate their adsorption-based C2H2/CH4 separation performances.50 Structural features (pore size, pore size ratio, surface area, porosity, pore volume, and/or linker length), chemical features (percentage of atom types), and energetic features (KH and/or Qst) of MOFs were used as the input data, whereas GCMC simulation results for C2H2/CH4 mixture adsorption were used as the target data to train ML models. The developed models were shown to accurately predict the adsorbent performance evaluation metrics of MOFs such as selectivity, working capacity, and regenerability, which are generally used to rank the adsorbent candidates to identify the most promising structures. ML-based rankings were shown to be in good agreement with the simulation-based rankings, indicating that our developed ML models can be used to efficiently and accurately identify the potentially viable materials for C2H2 separation from CH4 by bypassing the computationally demanding molecular simulations.

Computing molecular diffusivities of gases inside the pores of MOFs through MD simulations is required to assess their membrane-based separation performances. MD simulations require significant computational time and resources even when a few materials are considered. To address this challenge, an ML model was developed to predict the diffusivities of nine gases (including CH4, N2, H2, CO2, and He) in 6013 MOFs using the results of MD simulations where the structural features (pore size, density, surface area, and porosity) and physical properties of gas molecules (kinetic diameter, quadrupole moment, polarizability, and dipole moment) were used as inputs.51 A high accuracy in predicting diffusivities was achieved, and pore size and polarizability of the molecule were found to have a significant impact on the predicted gas diffusivity in MOFs.

A recent work52 combined multi-scale design of MOF-based membrane separation for CO2/CH4 mixture via integration of molecular simulation, ML, and process modeling. First, GCMC and MD simulations were performed to compute CO2/CH4 mixture adsorption and diffusion in IRMOF-1 at six different temperatures, a pressure range of 0–20 bar, and different feed CO2 mole fractions. Artificial neural network algorithms were used to develop models that accurately calculate gas permeability with a significantly lower computational time than the molecular simulations, and these models were integrated into a membrane process model. The results show that IRMOF-1 membranes are more efficient for CO2/CH4 separation than conventional polymeric membranes in terms of purity and recovery.

MOF/polymer mixed matrix membranes (MMMs), in which MOFs are used as filler particles in polymer matrices, have been shown to provide higher gas permeability and/or selectivity compared to pristine MOF membranes. Simulating every single MOF/polymer MMM is impractical due to the very large number of existing MOF-polymer combinations. ML was recently used to predict CO2 permeabilities and CO2/CH4 selectivities of 648 MOF/polymer MMMs using polymer type, MOF type, MOF loading, pore size, surface area, pressure, temperature, thickness, and aperture and filler sizes as the input data.53 Experimentally reported CO2 permeabilities and CO2/CH4 selectivities were collected from the literature and used to train ML models. Based on the results of ML, the polymer type, their intrinsic gas transport rates (control selectivity and control permeability), pore size, particle size, and surface area of the filler had high importance in predicting the gas separation performances of MOF/polymer MMMs. The optimum structural features of MOFs (pore size >10 Å and surface area ∼800 m2/g) for high-performance CO2/CH4 separation, were identified, and two MOFs, Cu-CAT-1 and Cu-THQ, were synthesized and combined with polymers, Pebax 2553 and PIM-1, to produce MMMs. Figure 5(a) shows that Cu-CAT-1/PIM-1 MMM surpassed Robeson's upper bound, which defines the separation performance limits of polymeric membranes, thanks to its high selectivity and permeability. In a follow-up study,54 ML models utilizing the same dataset were developed by employing filler size, surface area, cage size, aperture size, loading amount, thickness, polymer molecular descriptors, pressure, and temperature as input data. Feature importance analysis revealed that MOF cage size is the most important feature for the CO2 permeability of the MMMs. The types of polymer matrix and van der Waals surface area of polymers were also found to significantly affect the CO2 permeability, and temperature had a more pronounced influence on the CO2 permeability of MMMs compared to pressure. The loading amount of the filler was found as the most important descriptor for the membrane selectivity. This work highlighted the discovery of high-performance MMMs by integrating ML into molecular simulations.

FIG. 5.

(a) Experimentally reported CO2/CH4 selectivities as a function of CO2 permeabilities for the Cu-CAT-1/PIM-1 membrane. The selection and synthesis of this membrane were guided by insights derived from the ML model. Reproduced with permission from Guan et al., Cell Rep. Phys. Sci. 3(5), 100864 (2022). Copyright 2022 Elsevier.53 (b) Comparison of ML-predicted and simulated gas permeabilities with the available experimental data for MOF/polymer MMMs. Blue lines show the experimental gas permeabilities collected from the literature, and the number of the blue lines on each column represents the number of experimental data. The values in parentheses represent the volume fraction of MOF fillers. Reproduced with permission from Daglar and Keskin, ACS Appl. Mater. Interfaces 14(28), 32134–32148 (2022). Copyright 2022 American Chemical Society.55 

FIG. 5.

(a) Experimentally reported CO2/CH4 selectivities as a function of CO2 permeabilities for the Cu-CAT-1/PIM-1 membrane. The selection and synthesis of this membrane were guided by insights derived from the ML model. Reproduced with permission from Guan et al., Cell Rep. Phys. Sci. 3(5), 100864 (2022). Copyright 2022 Elsevier.53 (b) Comparison of ML-predicted and simulated gas permeabilities with the available experimental data for MOF/polymer MMMs. Blue lines show the experimental gas permeabilities collected from the literature, and the number of the blue lines on each column represents the number of experimental data. The values in parentheses represent the volume fraction of MOF fillers. Reproduced with permission from Daglar and Keskin, ACS Appl. Mater. Interfaces 14(28), 32134–32148 (2022). Copyright 2022 American Chemical Society.55 

Close modal

Separating He from CH4 is important not only to purify natural gas but also to obtain pure He for various industries, including healthcare and electronics. Our group recently integrated molecular simulations and ML to study the He/CH4 and N2/CH4 separation properties of MOF membranes and MOF/polymer MMMs.55 GCMC and MD simulations were performed to obtain gas adsorption and diffusion data of 5249 MOFs, and based on these simulated data, ML models that accurately predict N2, CH4, and He and permeabilities of 5224, 5215, 677 different types of MOFs and 31 344, 26 075, and 2031 types of MOF/polymer MMMs were developed, respectively. Figure 5(b) shows the good agreement between experimentally reported and ML-predicted He, CH4, and N2 permeabilities of five distinct MOF/polymer MMMs. The ML models developed to predict the gas uptake and diffusivity were tested on 1000 unseen hypothetical MOF-based MMMs, leading to accurate permeability and selectivity predictions, which will expedite the design of new MMMs for CH4 purification.

ML models were developed to accurately predict permeabilities, membrane selectivities, and the trade-off multiple selectivity and permeability (TMSP), a metric to demonstrate the trade-off between permeability and membrane selectivity, of synthesized MOFs for H2/CH4 separation when the structural descriptors (pore size, surface area, porosity, density, and pore size distribution) were used.56 A classification tree model was also developed by dividing the studied MOF membranes into three categories: top (top 5%), satisfactory (top 5%–10%), and poor (remaining) based on their structural features, which allowed rapid identification of promising MOFs for H2/CH4 separation. Instead of predicting the adsorption and diffusion data for every single material, this type of classification models could be used to rapidly identify promising membrane candidates for CH4 purification prior to extensive experimental studies.

Separating CH4 from the air can contribute to mitigating its hazardous impact on climate change by reducing its release into the atmosphere. Simulating multicomponent gas mixtures adds complexity to the MD simulations, and significant computational time is required to reach equilibrium. After performing GCMC and MD simulations to compute the adsorption and diffusion properties of 6013 synthesized MOFs, ML models were developed to predict adsorption selectivity, diffusion selectivity, permeability, and membrane selectivity based on the pore size, porosity, surface area, density, Qst, and Henry's constant of the MOFs.57 Several membrane design strategies were presented by comparing MOFs that differ solely in terms of metal, organic linker, or topology and exhibit promising and poor separation performance. This shows the importance of integrating molecular simulations and ML to rapidly design promising MOF membranes even only altering topology, organic linkers, or metal centers rather than synthesizing a completely new material.

Integration of ML to molecular simulations has recently played a significant role in identifying the most useful adsorbent and membrane materials among thousands of candidates for CH4 separation applications, and even in some examples, ML results were used to generate design strategies that lead to the development of novel materials outperforming the already existing one. We believe that there are still some challenges that can be seen as potential directions to advance this field.

Several studies we reviewed so far focused on the hypothetical MOF and COF materials since the number and variety of these computer-generated materials are significantly larger and wider than those of synthesized ones. The synthesis of hypothetical materials is not straightforward, and we expect that strong collaboration between synthetic chemists and computational researchers will accelerate the synthesis of the hypothetical MOFs tailored for CH4 separation in the upcoming years.

ML in chemistry and materials science needs accurate and accessible training data. Almost all the studies we discussed used molecular simulations that utilize generic force fields to train ML models for forecasting CH4 adsorption and separation properties of MOFs. These ML models are only as accurate as the simulation data used to train them. Development of new force fields to better describe the CH4-MOF interactions, especially for materials that have unusual properties such as open metal sites and defects, would be very useful to perform highly accurate molecular simulations and to achieve accurate ML models.

Studies using ML generally develop separate models for each gas of interest and each operating condition. Developing universal ML models that can be transferable/adaptable to different types of gas molecules available in a gas mixture and different operating conditions such as a large variety of pressures and temperatures, without losing accuracy, would be very useful to study the separation of CH4 from all other types of gases under various adsorption- and membrane-based separation processes. Finally, it is important to highlight that AI can reach its full potential in chemistry, physics, and materials science if the ML models and data are open accessed in a standardized and consistent format because researchers can only advance the field by quickly building on earlier work. Considering the rapid developments in the fields of porous materials, molecular simulation techniques, and ML applications, we expect that combining computational screening and data science will discover many more novel MOF adsorbents and membranes for CH4 purification.

S.K. acknowledges ERC-2017-Starting Grant. This research has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (ERC-2017-Starting Grant, Grant Agreement No. 756489-COSMOS).

The authors have no conflicts to disclose.

Hasan Can Gulbalkan: Conceptualization (equal); Investigation (equal); Writing – original draft (equal). Alper Uzun: Conceptualization (equal); Investigation (equal); Supervision (equal); Writing – original draft (equal); Writing – review & editing (equal). Seda Keskin: Conceptualization (equal); Funding acquisition (lead); Investigation (equal); Supervision (equal); Writing – original draft (equal); Writing – review & editing (equal).

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

1.
Q.
Wang
,
Y.
Yu
,
Y.
Li
,
X.
Min
,
J.
Zhang
, and
T.
Sun
,
Sep. Purif. Technol.
283
,
120206
(
2022
).
2.
D.
Saha
,
H. A.
Grappe
,
A.
Chakraborty
, and
G.
Orkoulas
,
Chem. Rev.
116
(
19
),
11436
11499
(
2016
).
3.
B.
Ma
,
C.
Deng
,
H.
Chen
,
M.
Zhu
,
M.
Yang
, and
X.
Feng
,
Ind. Eng. Chem. Res.
59
(
18
),
8715
8727
(
2020
).
4.
R. L.
Siegelman
,
P. J.
Milner
,
E. J.
Kim
,
S. C.
Weston
, and
J. R.
Long
,
Energy Environ. Sci.
12
(
7
),
2161
2173
(
2019
).
5.
M.
Taddei
and
C.
Petit
,
Mol. Syst. Des. Eng.
6
(
11
),
841
875
(
2021
).
6.
Q.
Qian
,
P. A.
Asinger
,
M. J.
Lee
,
G.
Han
,
K.
Mizrahi Rodriguez
,
S.
Lin
,
F. M.
Benedetti
,
A. X.
Wu
,
W. S.
Chi
, and
Z. P.
Smith
,
Chem. Rev.
120
(
16
),
8161
8266
(
2020
).
7.
S. U.
Nandanwar
,
D. R.
Corbin
, and
M. B.
Shiflett
,
Ind. Eng. Chem. Res.
59
(
30
),
13355
13369
(
2020
).
8.
O. K.
Farha
,
I.
Eryazici
,
N. C.
Jeong
,
B. G.
Hauser
,
C. E.
Wilmer
,
A. A.
Sarjeant
,
R. Q.
Snurr
,
S. T.
Nguyen
,
A.
Özgür Yazaydın
, and
J. T.
Hupp
,
J. Am. Chem. Soc.
134
(
36
),
15016
15021
(
2012
).
9.
H.
Furukawa
,
K. E.
Cordova
,
M.
O'Keeffe
, and
O. M.
Yaghi
,
Science
341
(
6149
),
1230444
(
2013
).
10.
W.
Fan
,
X.
Zhang
,
Z.
Kang
,
X.
Liu
, and
D.
Sun
,
Coord. Chem. Rev.
443
,
213968
(
2021
).
11.
T.
Wang
,
E.
Lin
,
Y.-L.
Peng
,
Y.
Chen
,
P.
Cheng
, and
Z.
Zhang
,
Coord. Chem. Rev.
423
,
213485
(
2020
).
12.
T. A.
Goetjen
,
J.
Liu
,
Y.
Wu
,
J.
Sui
,
X.
Zhang
,
J. T.
Hupp
, and
O. K.
Farha
,
Chem. Commun.
56
(
72
),
10409
10418
(
2020
).
13.
D.
Yang
,
M.
Babucci
,
W. H.
Casey
, and
B. C.
Gates
,
ACS Cent. Sci.
6
(
9
),
1523
1533
(
2020
).
14.
H.-Y.
Li
,
S.-N.
Zhao
,
S.-Q.
Zang
, and
J.
Li
,
Chem. Soc. Rev.
49
(
17
),
6364
6401
(
2020
).
15.
S.
Mallakpour
,
E.
Nikkhoo
, and
C. M.
Hussain
,
Coord. Chem. Rev.
451
,
214262
(
2022
).
16.
H. D.
Lawson
,
S. P.
Walton
, and
C.
Chan
,
ACS Appl. Mater. Interfaces
13
(
6
),
7004
7020
(
2021
).
17.
K.
Dashtian
,
S.
Shahbazi
,
M.
Tayebi
, and
Z.
Masoumi
,
Coord. Chem. Rev.
445
,
214097
(
2021
).
18.
C. R.
Groom
,
I. J.
Bruno
,
M. P.
Lightfoot
, and
S. C.
Ward
,
Acta Crystallogr. B
72
(
2
),
171
179
(
2016
).
19.
I. J.
Bruno
,
J. C.
Cole
,
P. R.
Edgington
,
M.
Kessler
,
C. F.
Macrae
,
P.
McCabe
,
J.
Pearson
, and
R.
Taylor
,
Acta Crystallogr. B
58
(
3
),
389
397
(
2002
).
20.
S.
Majumdar
,
S. M.
Moosavi
,
K. M.
Jablonka
,
D.
Ongari
, and
B.
Smit
,
ACS Appl. Mater. Interfaces
13
(
51
),
61004
61014
(
2021
).
21.
S.
Lee
,
B.
Kim
,
H.
Cho
,
H.
Lee
,
S. Y.
Lee
,
E. S.
Cho
, and
J.
Kim
,
ACS Appl. Mater. Interfaces
13
(
20
),
23647
23654
(
2021
).
22.
X.
Li
,
S.
Cai
,
B.
Sun
,
C.
Yang
,
J.
Zhang
, and
Y.
Liu
,
Matter
3
(
5
),
1507
1540
(
2020
).
23.
M.
Tong
,
Y.
Lan
,
Q.
Yang
, and
C.
Zhong
,
Chem. Eng. Sci.
168
,
456
464
(
2017
).
24.
R.
Mercado
,
R.-S.
Fu
,
A. V.
Yakutovich
,
L.
Talirz
,
M.
Haranczyk
, and
B.
Smit
,
Chem. Mater.
30
(
15
),
5069
5086
(
2018
).
25.
Z.
Wang
,
S.
Zhang
,
Y.
Chen
,
Z.
Zhang
, and
S.
Ma
,
Chem. Soc. Rev.
49
(
3
),
708
735
(
2020
).
26.
S. M.
Wang
,
M.
Shivanna
, and
Q. Y.
Yang
,
Angew. Chem., Int. Ed.
61
(
15
),
e202201017
(
2022
).
27.
Z.
Niu
,
X.
Cui
,
T.
Pham
,
P. C.
Lan
,
H.
Xing
,
K. A.
Forrest
,
L.
Wojtas
,
B.
Space
, and
S.
Ma
,
Angew. Chem., Int. Ed.
58
(
30
),
10138
10141
(
2019
).
28.
T.
Li
,
X.
Jia
,
H.
Chen
,
Z.
Chang
,
L.
Li
,
Y.
Wang
, and
J.
Li
,
ACS Appl. Mater. Interfaces
14
(
13
),
15830
15839
(
2022
).
29.
W.
Fan
,
Y.
Ying
,
S. B.
Peh
,
H.
Yuan
,
Z.
Yang
,
Y. D.
Yuan
,
D.
Shi
,
X.
Yu
,
C.
Kang
, and
D.
Zhao
,
J. Am. Chem. Soc.
143
(
42
),
17716
17723
(
2021
).
30.
H.
Fan
,
A.
Mundstock
,
J.
Gu
,
H.
Meng
, and
J.
Caro
,
J. Mater. Chem. A
6
(
35
),
16849
16853
(
2018
).
31.
H.
Daglar
and
S.
Keskin
,
Coord. Chem. Rev.
422
,
213470
(
2020
).
32.
E.
Ren
,
P.
Guilbaud
, and
F.-X.
Coudert
,
Digital Discovery
1
(
4
),
355
374
(
2022
).
33.
X.
Zhang
,
Z.
Xu
,
Z.
Wang
,
H.
Liu
,
Y.
Zhao
, and
S.
Jiang
,
APL Mater.
11
(
6
),
060901
(
2023
).
34.
T.
Yan
,
Y.
Lan
,
D.
Liu
,
Q.
Yang
, and
C.
Zhong
,
Chemistry
14
(
20
),
3688
3693
(
2019
).
35.
D.
Saha
,
Z.
Bao
,
F.
Jia
, and
S.
Deng
,
Environ. Sci. Technol.
44
(
5
),
1820
1826
(
2010
).
36.
O. F.
Altundal
,
Z. P.
Haslak
, and
S.
Keskin
,
Ind. Eng. Chem. Res.
60
(
35
),
12999
13012
(
2021
).
37.
J.
Alzubi
,
A.
Nayyar
, and
A.
Kumar
,
J. Phys. Conf. Ser.
1142
,
012012
(
2018
).
38.
A.
Mishra
,
P.
Rajak
,
A.
Irie
,
S.
Fukushima
,
R. K.
Kalia
,
A.
Nakano
,
K-i
Nomura
,
F.
Shimojo
, and
P.
Vashishta
,
Appl. Phys. Lett.
123
(
12
),
121901
(
2023
).
39.
F.
Najafi
,
H.
Sveinsson
,
C.
Dreierstad
,
H.
Glad
, and
A.
Sørenssen
,
Appl. Phys. Lett.
123
,
111601
(
2023
).
40.
X.
Jia
,
H.
Yao
,
Z.
Yang
,
J.
Shi
,
J.
Yu
,
R.
Shi
,
H.
Zhang
,
F.
Cao
,
X.
Lin
, and
J.
Mao
,
Appl. Phys. Lett.
123
(
20
),
203902
(
2023
).
41.
Z.
Shi
,
W.
Yang
,
X.
Deng
,
C.
Cai
,
Y.
Yan
,
H.
Liang
,
Z.
Liu
, and
Z.
Qiao
,
Mol. Syst. Des. Eng.
5
(
4
),
725
742
(
2020
).
42.
H.
Demir
,
H.
Daglar
,
H. C.
Gulbalkan
,
G. O.
Aksu
, and
S.
Keskin
,
Coord. Chem. Rev.
484
,
215112
(
2023
).
43.
J. P.
Janet
and
H. J.
Kulik
,
J. Phys. Chem. A
121
(
46
),
8939
8954
(
2017
).
44.
K. M.
Jablonka
,
D.
Ongari
,
S. M.
Moosavi
, and
B.
Smit
,
Chem. Rev.
120
(
16
),
8066
8129
(
2020
).
45.
H. C.
Gulbalkan
,
A.
Uzun
, and
S.
Keskin
,
ACS Appl. Mater. Interfaces
(
2023
).
46.
W.
Li
,
Y.
Situ
,
L.
Ding
,
Y.
Chen
, and
Q.
Yang
,
ACS Appl. Mater. Interfaces
15
(
51
),
59887
59894
(
2023
).
47.
R.
Anderson
and
D. A.
Gómez-Gualdrón
,
J. Chem. Phys.
154
(
23
),
234102
(
2021
).
48.
G. O.
Aksu
and
S.
Keskin
,
J. Mater. Chem. A
11
(
27
),
14788
14799
(
2023
).
49.
Z.
Yao
,
B.
Sánchez-Lengeling
,
N. S.
Bobbitt
,
B. J.
Bucior
,
S. G. H.
Kumar
,
S. P.
Collins
,
T.
Burns
,
T. K.
Woo
,
O. K.
Farha
, and
R. Q.
Snurr
,
Nat. Mach. Intell.
3
(
1
),
76
86
(
2021
).
50.
H.
Demir
and
S.
Keskin
,
Chem. Eng. J.
464
,
142731
(
2023
).
51.
S.
Guo
,
X.
Huang
,
Y.
Situ
,
Q.
Huang
,
K.
Guan
,
J.
Huang
,
W.
Wang
,
X.
Bai
,
Z.
Liu
, and
Y.
Wu
,
Adv. Sci.
10
,
2301461
(
2023
).
52.
X.
Cheng
,
Y.
Liao
,
Z.
Lei
,
J.
Li
,
X.
Fan
, and
X.
Xiao
,
J. Membr. Sci.
672
,
121430
(
2023
).
53.
J.
Guan
,
T.
Huang
,
W.
Liu
,
F.
Feng
,
S.
Japip
,
J.
Li
,
J.
Wu
,
X.
Wang
, and
S.
Zhang
,
Cell Rep. Phys. Sci.
3
(
5
),
100864
(
2022
).
54.
M.
Alizamir
,
A.
Keshavarz
,
F.
Abdollahi
,
A.
Khosravi
, and
S.
Karagöz
,
Sep. Purif. Technol.
325
,
124689
(
2023
).
55.
H.
Daglar
and
S.
Keskin
,
ACS Appl. Mater. Interfaces
14
(
28
),
32134
32148
(
2022
).
56.
X.
Bai
,
Z.
Shi
,
H.
Xia
,
S.
Li
,
Z.
Liu
,
H.
Liang
,
Z.
Liu
,
B.
Wang
, and
Z.
Qiao
,
Chem. Eng. J.
446
,
136783
(
2022
).
57.
H.
Li
,
C.
Wang
,
Y.
Zeng
,
D.
Li
,
Y.
Yan
,
X.
Zhu
, and
Z.
Qiao
,
Membranes
12
(
9
),
830
(
2022
).