Skip to main content
NCBI home page
ACS AuthorChoice logoLink to ACS AuthorChoice
. 2025 Jul 4;129(28):13089–13099. doi: 10.1021/acs.jpcc.5c02779

Integrating Molecular Simulations with Machine Learning to Discover Selective MOFs for CH4/H2 Separation

Pelin Sezgin 1, Seda Keskin 1,*
PMCID: PMC12278306  PMID: 40697403

Abstract

As the number of synthesized and hypothetical metal–organic frameworks (MOFs) continues to grow, identifying the most selective adsorbents for CH4/H2 separation through experimental or computational methods has become increasingly complex. This study integrates molecular simulations with machine learning (ML) to evaluate the CH4/H2 separation performance of 126605 distinct types of MOFs. Grand canonical Monte Carlo (GCMC) simulations were performed to produce CH4 and H2 adsorption data for synthesized MOFs at various pressures, which were then used to train ML models incorporating structural, chemical, and energetic features of the MOFs. These ML models were subsequently transferred to hypothetical MOFs, enabling the rapid and accurate screening of promising adsorbents for CH4/H2 separation. The top-performing MOFs were identified based on their CH4/H2 selectivities, and their key structural and chemical characteristics were analyzed. Synthesized (hypothetical) MOFs having narrow pores and pyridine-, histidine-, and imidazole-based (carboxylate-, benzoate-, and cubane-based) linkers demonstrated high selectivities up to 85 (115) at 1 bar and 298 K. Our findings highlight the potential of MOFs as superior alternatives to traditional adsorbent materials for CH4/H2 separation.

graphic file with name jp5c02779_0010.jpg

graphic file with name jp5c02779_0008.jpg

1. Introduction

Efficient separation of methane (CH4) from hydrogen (H2) is important to obtain high purity gases in several industrial applications, including refinery off-gas processing and steam methane reforming. Conventional separation methods, such as cryogenic distillation and chemical absorption, demand excessive energy and result in high costs. Adsorption-based gas separation utilizing porous materials is an energy-efficient and environmentally friendly technology. Over the past decades, various porous adsorbents, such as carbon-based materials and zeolites, have been used for CH4/H2 separation. , However, their low selectivities have driven the search for the design and discovery of more effective adsorbent materials.

Metal–organic frameworks (MOFs) have received significant attention as strong alternatives to traditional porous materials for adsorption-based gas separations thanks to their unique properties: exceptionally high porosities, large surface areas, low densities, and a large variety of pore sizes. − Constructed from metal clusters connected by organic linkers, MOFs exhibit a nearly limitless potential for structural diversity. The combination of various metals and linkers enables the precise tailoring of their geometry and high chemical functionality, making them ideal for the design of novel MOF adsorbents for specific gas separations. Alongside several thousands of experimentally synthesized MOFs, hundreds of thousands of structures have been computationally designed as hypothetical MOFs. Although the existence of a very large number of MOF structures is a great opportunity to find the optimal adsorbents for target gas adsorption and separation applications, it is practically impossible to experimentally synthesize and test every single MOF.

Computational methods, specifically molecular simulations, have been very useful to assess the gas adsorption and separation properties of very large numbers of MOFs with high accuracy, reducing the reliance on time, effort, and cost-intensive experimental approaches. , Grand Canonical Monte Carlo (GCMC) simulations have been widely used to assess the gas adsorption properties of thousands of MOFs and to guide the experimental efforts toward the most promising candidates while uncovering the materials’ molecular structure–performance relations. Most of the molecular simulation studies have focused on either CH4 or H2 storage in MOFs, − whereas a limited number of them have examined adsorption-based CH4/H2 separation of selected MOFs. For example, GCMC simulations were used to compute CH4/H2 selectivities of 105 synthesized MOFs for an equimolar mixture and the most promising MOFs were reported to have higher selectivities (∼126 at 10 bar, 298 K) than zeolites and carbonaceous materials. 250 synthesized MOFs were studied by GCMC simulations and CH4/H2 mixture selectivity was reported as 314 at 1 bar, 298 K for the top-performing material. 440 MOFs were screened using GCMC simulations and CH4/H2 mixture selectivities were reported to be between 15 and 22 for the top 20 MOFs at 10 bar, 300 K. 1109 MOFs with varying chemical compositions were examined using GCMC simulations to assess their ideal CH4/H2 selectivities across different pressures and the highest selectivity was reported as 1370 at infinite dilution. CH4/H2 separation performances of 2408 MOFs were examined by GCMC simulations and selectivities between 39 and 95 were reported for nine promising MOFs at 10 bar, 303 K. 2735 MOFs were studied by GCMC simulations for equimolar CH4/H2 mixture separation and promising MOFs were found to have high selectivities up to 773 at 1 bar, 298 K. The potential of 4350 synthesized MOFs for adsorption-based CH4/H2 separation was assessed by GCMC simulations and numerous MOFs were identified to have remarkably higher selectivities for equimolar CH4/H2 mixture separation compared to those of traditional adsorbents. 11115 MOFs were screened by GCMC simulations for an equimolar CH4/H2 mixture separation and the top 20 MOFs were reported to have selectivities between 33 and 46 at 10 bar, 298 K.

As this literature review shows, current GCMC studies generally assessed the CH4/H2 separation performances of a selected number of synthesized MOFs. Considering the ever-growing MOF space, studying the entire MOF spectrum composed of both synthesized and hypothetical materials is important to evaluate the full potential of this porous material family for CH4/H2 separation by comparing it with other porous adsorbents. However, performing GCMC simulations for hundreds of thousands of MOFs is computationally very expensive. Therefore, recent efforts focused on combining machine learning (ML) with GCMC simulations to develop models that can accurately predict the gas adsorption properties of MOFs in a very short time when the specific materials features are introduced to the models. − For example, ML algorithms were used to assess CH4 adsorption performances of hypothetical MOFs at 35 bar, 298 K; , H2 adsorption performances of synthesized MOFs at 2, 10, 50, and 100 bar, 77 and 298 K; and H2 adsorption capacity of both synthesized and hypothetical MOFs at 1 and 100 bar, 77 K. These studies showed that ML models can accurately and efficiently predict the CH4 and H2 adsorption capacities of different types of MOFs. However, a gap exists in the literature when it comes to predicting the CH4/H2 separation performance of both synthesized and hypothetical MOFs.

Motivated from this, in this work, we integrated GCMC simulations and ML approaches to assess CH4/H2 separation performances of both synthesized and hypothetical MOFs, composed of 126605 different types of materials at pressures of 0.1, 1, and 10 bar. We initially performed GCMC simulations for 4331 synthesized MOFs to compute their CH4 and H2 uptakes and then used these high-fidelity adsorption data to train ML models that can accurately estimate the gas uptake of any MOF only in seconds once the structural, chemical, and energetic features of the material are given. These ML models were then transferred to 122274 unseen hypothetical MOFs to evaluate their gas uptakes. ML-predicted gas uptakes were used to assess the CH4/H2 selectivities of all MOFs. The results were then compared with the selectivity predictions from two previously proposed mathematical models in the literature, providing a more comprehensive assessment of the reliability and generalizability of our ML models. Finally, the most promising adsorbents offering the highest selectivities were identified and analyzed in detail to reveal the desired structural and chemical properties of MOFs that lead to high selectivities. Our computational approach will be a very efficient alternative to brute-force molecular simulations for assessing the gas adsorption and separation performances of numerous MOFs while also directing the experimental efforts toward the most promising materials.

2. Computational Details

2.1. MOF Databases

Figure demonstrates the computational approach we followed to investigate CH4 and H2 adsorption and separation properties of three different MOF databases: Computation-Ready Experimental MOF (CoRE MOF) database, Wilmer’s hypothetical MOF database, and Smit’s hypothetical MOF database. Structural properties of all MOFs available in these three databases such as pore limiting diameter (PLD), the largest cavity diameter (LCD), accessible surface area (ASA), geometric pore volume (PV), and porosity (ϕ) were computed using Zeo++ software (version 0.3). The Monte Carlo sampling method was used to calculate ASA and PV, using probe radii of 1.82 Å (representing N2) and 0 Å, respectively. All solvent removed (ASR) and nondisordered MOF structures from the most recent version of the CoRE MOF 2019 database including 11793 experimentally reported MOFs were collected. These MOFs were filtered based on their PLDs (>3.8 Å) to ensure that both CH4 (3.80 Å) and H2 (2.96 Å) molecules can be adsorbed into the pores. Further elimination was done to obtain structures with ASA > 0 m2/g and to include only “computation-ready” structures resulting in 4331 synthesized MOFs. Wilmer’s database includes 137953 different types of hypothetical MOFs constructed from 102 building blocks varying in chemical compositions, five different metal nodes, 84 organic linkers, and six topologies. Smit’s database contains 23891 hypothetical MOFs having 14 different metal nodes, 95 organic linkers, and 52 topologies. Following the same structural filtering explained above, we ended up with 98601 MOFs from Wilmer’s database and 23673 MOFs from Smit’s database, which will be refereed as Hypo1 and Hypo2 databases, respectively, throughout our manuscript.

1.

1

Open in a new tab

Computational workflow: (1) Computing structural, chemical, and energetic features of materials in three MOF databases we studied. (2) Performing GCMC simulations for assessing CH4 and H2 adsorption data of CoRE MOFs; training and testing ML models using the results of molecular simulations and transferring these models to assess CH4 and H2 adsorption properties of Hypo1 and Hypo2MOF databases. (3) Evaluating CH4/H2 separation performances of all materials present in three databases to identify the most selective MOFs and analyzing their properties in detail.

Chemical features of all MOFs including C%, H%, N%, O%, Metal%, Halogen (Br, Cl, F, I) %, Ametal (Se, S, P) %, and Metalloid (As, B, Ge, Te, Sb, Si) % were obtained from the crystallographic information files existing in the corresponding MOF databases. The number of defined elements was divided by the number of all atoms that exist in the extended unit cell and multiplied by 100 to calculate the atomic percentages.

2.2. Molecular Simulations

For GCMC simulations, we employed the RASPA simulation code (version 2.0.5) to compute the adsorption of CH4 and H2 molecules in 4331 CoRE MOFs at 0.1, 1, and 10 bar at room temperature. The Lennard-Jones (LJ) 12-6 potentials were defined for gas–MOF and gas–gas interactions and LJ parameters were obtained from the Universal Force Field (UFF) for MOFs. All UFF parameters of framework atoms are provided on GitHub (https://github.com/pelinsezgin/CH4H2_ML). Single-site spherical potentials were utilized for modeling CH4 and H2 gases. , The Lorentz–Berthelot combining rules were used to determine the pair potentials between the distinct atoms. All short-range LJ interactions were calculated using a cutoff distance of 13.0 Å, and cell lengths were extended to have at least twice the designated cutoff distance in each dimension. We employed 104 cycles for initialization and 2 × 104 cycles for taking the ensemble averages. The isosteric heat of adsorption (Q st) and Henry’s constants (K H,i) of gas molecules were computed using the Widom particle insertion method at infinite dilution at 298 K. Adsorption density plots were generated using the iRASPA software for the most promising MOFs identified from each database to visualize the preferential adsorption sites for the gas molecules. These visualizations were supported by radial distribution function (RDF) analysis, which describes the probability of finding a particle at a distance r from a given particle, computed during GCMC simulations. RDF results were written every 104 cycles with an RDF histogram size of 300 Å and an RDF range of 13 Å. RASPA performs normalization based on the total system volume, and the resulting normalized RDF histograms were used to construct the RDF plots.

2.3. Machine Learning

We trained six different ML models to predict CH4 and H2 uptakes of 4331 CoRE MOFs at three different pressures using the simulated gas uptake data as the target data and the structural, chemical, and energetic properties of the MOFs as the input data. As listed in Table S1, five structural descriptors: PLD, LCD, ASA, ϕ, and PV, eight chemical descriptors: C%, H%, N%, O%, Metal%, Halogen%, Ametal%, and Metalloid%, and an energy-based descriptor: Henry’s constant of CH4 (KH,CH4) or H2 (KH,H2) was used as the input data to train the ML models. We examined the correlation between the MOFs’ features and their simulated CH4 and H2uptakes by calculating the Pearson correlation coefficients, as shown in Figures S1 and S2, respectively. Tree-based pipeline optimization tool (TPOT, version 0.12.2) was used in the automated ML framework to determine the best algorithms and to optimize the hyperparameters. Regression algorithms from the scikit-learn toolkit were employed for the TPOT model selection. With 80% of the data designated for training and 20% for testing, we employed the stratified sampling technique to maintain a constant feature distribution between test and training data sets. We used 5-fold cross-validation to prevent overfitting.

The coefficient of determination (R 2), mean absolute error (MAE), and root-mean-square error (RMSE) were calculated to check the accuracy of ML models using the equations given in Table S2. Leveraging these metrics, we used XGBoost algorithm for CH4 uptake predictions at all pressures; ElasticNetCV algorithm for H2 uptake predictions at 0.1 and 1 bar; and RidgeCV algorithm for H2 uptake predictions at 10 bar. The optimized hyperparameters of these models are given in Table S3. The impact of features on ML predictions was evaluated using the SHapley Additive exPlanations (SHAP) analysis. Positive or negative SHAP values were determined by subtracting the model’s output for simulated CH4 and H2 uptakes from the average predictions. ML models were then transferred to the Hypo1 and Hypo2MOF databases to assess their CH4 and H2 adsorption properties without performing GCMC simulations for thousands of different structures at three different pressures. To test the transferability of ML models developed for CoRE MOFs to hypothetical MOFs, we performed GCMC simulations on two representative subsets of hypothetical MOFs, 9868 MOFs from the Hypo1 database and 2788 MOFs from the Hypo2 database, which account for 10% of their respective databases. These subsets were selected to capture the key structural and energetic characteristics of the full Hypo1 and Hypo2MOF databases, including surface area, porosity, and K H,i, as we discuss in the following section.

2.4. Selectivity Calculations

The ratios of ML-predicted gas uptakes were used to calculate the CH4/H2 selectivities of all MOFs at the corresponding pressures. We also used two mathematical models previously proposed to assess selectivities of MOFs: The first model predicts the selectivity, S, at zero coverage using the difference in the heat of adsorption values of gases (ΔQ st°), lnS= (−0.8558) + ΔQ st°/(R × T), where R and T represent the gas constant and temperature, respectively. The second model estimates the selectivity at 0.01 bar using S = 0.0031 × (1/ΔAD)−3.1, where adsorbility (ΔAD) was defined as ΔQ st° divided by the porosity (ϕ) of the material. We used these two models to evaluate CH4/H2 selectivities of 16987 MOFs, for which we have selectivity data directly from the GCMC simulations, and compared the estimates of mathematical models, our ML predictions, and simulation results. After demonstrating the high accuracy of our ML models in predicting selectivities, we applied them to assess the selectivities of 4331 CoRE, 98601 Hypo1MOFs, and 23673 Hypo2MOFs. Selectivities of all 126605 MOFs computed at 1 bar were ranked to identify the most selective synthesized and hypothetical MOFs for CH4/H2 separation, and the top 10 MOFs from each database were then examined in detail to uncover their most important structural and chemical features that lead to high selectivities.

3. Results and Discussion

3.1. CH4 and H2 Adsorption Properties of CoRE MOFs

Figure displays the simulated CH4 and H2 uptakes of 4331 CoRE MOFs as a function of the materials’ three features: Henry’s constants, surface areas, and porosities. These features were strongly correlated with the gas uptakes at three different pressures, as represented by the Pearson correlation heat maps given in Figures S1 and S2. At 0.1 bar, both gas uptakes are strongly correlated with Henry’s constants since these constants represent the strength of MOF–gas interactions at low pressure as shown in Figure a,d. The Henry’s constants of CH4 and H2 were computed to be in the range of 4.4 × 10–7–5.8 × 10–4 mol/kg/Pa and 4.3 × 10–8–2.7 × 10–6 mol/kg/Pa, respectively. These values indicate that CH4 is more strongly adsorbed, with uptakes ranging from 4.4 × 10–3 to 2.4 mol/kg, compared to H2, which has uptakes ranging from 4.2 × 10–4 to 2.7 × 10–2 mol/kg.

2.

2

Open in a new tab

Simulated CH4 and H2 uptakes of 4331 CoRE MOFs as a function of their (a,d) Henry’s constants, (b,e) accessible surface areas, and (c,f) porosities at 0.1, 1, and 10 bar.

At 1 bar, CH4 uptakes increase from 4.2 × 10–2 to 5.0 mol/kg as the ASA rises to ∼1600 m2/g, while H2 uptakes almost linearly increase from 4.3 × 10–3 to 0.3 mol/kg with increasing ASA as displayed in Figure b,e. Figure c,f highlights that higher porosities lead to higher gas uptakes at 10 bar since MOFs with larger pore volumes provide more adsorption space. Specifically, CH4 uptakes (0.3–12.5 mol/kg) increase with increasing porosity (0.3–0.9), whereas H2 uptakes rise exponentially from 0.4 to 2.6 mol/kg. These findings are consistent with the earlier molecular simulation studies on MOFs, which concluded that surface area and porosity are the key determinants of the adsorbed gas amounts at moderate/high pressures. ,

We then used these simulated gas uptakes of CoRE MOFs to train ML models that can rapidly and accurately predict the CH4 and H2 adsorption data of any given MOF in just minutes as a very efficient alternative to brute-force GCMC simulations. Figure shows the strong predictive power of our ML models, which accurately assess both gas uptakes at different pressures for 4331 CoRE MOFs. The R 2 values for the test set were calculated as 0.986, 0.968, and 0.946 for CH4 adsorption and as 0.999, 1.000, and 0.999 for H2 adsorption at 0.1, 1, and 10 bar, respectively. Our models can capture and generalize the patterns of the training data to the corresponding test data, as evidenced by the distributions in training and test sets being similar to the simulated ones.

3.

3

Open in a new tab

Comparison of ML-predicted and simulated (a–c) CH4 and (d–f) H2 uptakes of 4331 CoRE MOFs at 0.1, 1, and 10 bar, respectively. Green and orange points represent the data of materials in training and test sets, respectively.

Figure a,d shows that ML-predicted CH4 and H2 uptakes at 0.1 bar (1.2 × 10–2–2.4 and 4.4 × 10–4–2.7 × 10–2 mol/kg, respectively) closely align with the GCMC simulation results (4.4 × 10–3–2.4 and 4.2 × 10–4–2.7 × 10–2 mol/kg, respectively) except 13 MOFs that have the lowest simulated uptake values (4.4 × 10–3–9.9 × 10–3 mol/kg) among 4331 CoRE MOFs. ML-predicted CH4 and H2 uptakes at 1 bar range from 5.9 × 10–2 to 4.7 mol/kg and 4.3 × 10–3 to 2.7 × 10–1 mol/kg, respectively, while simulated uptakes are between 4.2 × 10–2 and 5.0 and 4.3 × 10–3 and 2.7 × 10–1 mol/kg, showing a strong agreement as illustrated in Figure b,e. Figure c,f demonstrates that ML-predicted CH4 and H2 uptakes at 10 bar (0.4–11.9 and 4.0 × 10–2–2.6 mol/kg, respectively) also agree well with the simulated uptakes (0.3–12.5 and 4.2 × 10–2–2.6 mol/kg, respectively).

To investigate how different MOF features affect the model’s predictions, we performed SHAP analysis. Figure S3a,b shows that at low pressures, K H,CH4 and K H,H2 affect the uptake predictions positively, meaning that as Henry’s constants increase, gas uptakes also increase as expected. The influence of pore size and porosity becomes more noticeable for CH4 uptake predictions compared to H2 at 1 bar as presented in Figure S3c,d where narrower pore sizes and higher porosities lead to higher CH4 uptakes. At 10 bar, structural features such as pore volume and surface area play an important role for precisely predicting CH4 and H2 uptakes as shown in Figure S3e,f. Pore volume is the most important (second most important) feature for CH4 (H2) uptake predictions, and uptakes increase as PV increases.

3.2. CH4 and H2 Adsorption Properties of HypoMOFs

We then tested the transferability of the ML models that we developed for CoRE MOFs to hypoMOFs since the three databases have structures with varying properties. For example, CoRE MOFs that we studied in this work were computed to have the following PLD, LCD, ASA, PV, and ϕ values of 3.8–71.5 Å, 4.0–71.6 Å, 45.4–6842 m2/g, 0.13–6.02 cm3/g, and 0.32–0.94, respectively. These properties were calculated to be between 3.8 and 25.5 Å, 4.0 and 37.0 Å, 88.6 and 7037 m2/g, 0.12 and 8.05 cm3/g, and 0.33 and 0.95 for Hypo1MOFs and 3.8 and 54.2 Å, 4.7 and 56.6 Å, 86.1 and 7958 m2/g, 0.26 and 18.01 cm3/g, and 0.45 and 0.98 for Hypo2MOFs, indicating that MOFs in the Hypo2 database generally have larger pore sizes and pore volumes than those in the Hypo1 database.

Before using the ML models for the entire databases, we selected 9868 and 2788 hypoMOFs as the subsets representing 10% of the entire Hypo1 and Hypo2 databases, respectively, to make a comparison with the GCMC simulation results. To ensure the accurate transfer of ML models, the features of the training data set must capture the features of the selected Hypo1 and Hypo2MOF subsets. Figure shows the analysis of the structural and energetic features of 3464 CoRE MOFs used for training ML models (CoRE Train), 867 CoRE MOFs used for testing ML models (CoRE Test), 9868 selected Hypo1MOFs (Hypo1subset), 2788 selected Hypo2MOFs (Hypo2subset), and all Hypo1 and Hypo2MOF databases composed of 98601 and 23673 structures, respectively (Hypo1all and Hypo2all). At low pressures, the most important feature was Henry’s constant as we discussed before and Figure a displays that K H,CH4 values of the training CoRE MOF set (5.4 × 10–7–5.8 × 10–4 mol/kg/Pa) cover those of Hypo1 (1.8 × 10–6–5.3 × 10–4 mol/kg/Pa) and Hypo2 (1.6 × 10–6–3.1 × 10–5 mol/kg/Pa) subsets. Similarly, K H,H2 values of the training set, ranging between 4.4 × 10–8 and 2.7 × 10–6 mol/kg/Pa, cover the values observed in Hypo1 and Hypo2MOF subsets, which are in the range of 1.7 × 10–7–2.6 × 10–6 and 1.4 × 10–7–2.7 × 10–6 mol/kg/Pa as shown in Figure b. At high pressures, the most important features were structural properties of MOFs. As shown in Figure c–f, accessible surface area, porosity, LCD, and pore volume ranges of the materials in the training set (45.5–6842 m2/g, 0.33–0.93, 4.0–71.6 Å, and 0.1–5.7 cm3/g) mostly resemble those of Hypo1 (158–6472 m2/g, 0.39–0.93, 4.3–29.6 Å, and 0.2–5.6 cm3/g) and Hypo2MOF subsets (115–7169 m2/g, 0.50–0.94, 5.1–46.6 Å, and 0.3–6.3 cm3/g). Since both the structural and energetic features of CoRE MOFs used in training ML models cover the properties of Hypo1 and Hypo2MOF subsets, ML models are expected to be successfully transferred to hypoMOFs.

4.

4

Open in a new tab

Energetic and structural feature distribution of MOFs in different databases: (a) Henry’s constants for CH4, (b) Henry’s constants for H2, (c) accessible surface areas, (d) porosities, (e) the largest cavity diameters, and (f) pore volumes.

Figure S4 shows that ML models accurately predict the CH4 uptakes of 9868 Hypo1MOFs with high R2 values of 0.990, 0.953, and 0.856 and low RMSE values of 0.01, 0.14, and 0.51 mol/kg at 0.1, 1, and 10 bar, respectively. For H2 uptakes, linear models provided extrapolation capability, and highly accurate predictions were obtained, resulting in R 2 values of 0.998, 1.000, and 0.999, at 0.1, 1, and 10 bar, respectively. Similarly, ML models achieved highly accurate CH4 and H2 uptake predictions for 2788 Hypo2MOFs, resulting in R 2 values > 0.949 at all pressures as shown in Figure S5.

Generating adsorption data for all 122274 hypoMOFs at various pressures using brute-force GCMC simulations would take several months. Using the ML models we developed, this data was generated in minutes through easily computable MOF features. ML predictions for CH4 and H2 uptakes for the entire Hypo1 and Hypo2 databases are given in Figure . Figure a,d shows that most of the Hypo1MOFs have significantly higher CH4 uptakes (1.7 × 10–2–2.3 mol/kg) than Hypo2MOFs (1.7 × 10–2–2.8 × 10–1 mol/kg) at 0.1 bar, since Hypo1MOFs have higher KH,CH4 values. Furthermore, there are two CoRE MOFs having higher CH4 uptake (2.4 mol/kg) at 0.1 bar compared to Hypo1MOFs. Figure b,e shows that ML-predicted CH4 and H2 uptakes of Hypo1MOFs range between 0.1 and 4.4 mol/kg and 1 × 10–2 and 0.3 mol/kg, respectively, at 1 bar while Hypo2MOFs have lower CH4 uptakes (0.1–1.9 mol/kg) and slightly higher H2 uptakes (1.1 × 10–2–0.8 mol/kg). Seven CoRE MOFs achieve CH4 uptakes (4.5–5.0 mol/kg) that are higher than those of Hypo1MOFs at 1 bar. Here, we note that while all three MOF databases contain materials with similar surface areas, CoRE MOFs stand out for their greater diversity in metals and linkers, leading to a wider range of CH4 uptakes. Although the Hypo1 and Hypo2 databases are significantly larger than the CoRE MOF database in terms of the number of structures, the recurring structural patterns within these hypothetical databases limit their chemical diversity.

5.

5

Open in a new tab

ML-predicted CH4 (blue) and H2 (green) uptakes of 98601 Hypo1MOFs and 23673 Hypo2MOFs as a function of their Henry’s constants, surface areas, and porosities at (a,d) 0.1, (b,e) 1, and (c,f) 10 bar.

Figure c,f shows that Hypo1MOFs have similar CH4 uptakes (1.0–10.8 mol/kg) with Hypo2MOFs (1.0–8.1 mol/kg), at 10 bar, while H2 uptakes of Hypo1MOFs (0.1–3.4 mol/kg) are lower than Hypo2MOFs (0.1–7.3 mol/kg). Two CoRE MOFs achieve higher CH4 uptakes (12.1 and 12.5 mol/kg) than Hypo1MOFs at 10 bar; however, H2 uptakes (4.2 × 10–2–2.6 mol/kg) of CoRE MOFs are lower than both hypothetical databases. For the H2 molecule, available space is more important than interactions at high pressures, thus, larger porosities of materials in the Hypo2MOF database compared to others lead to higher H2 uptakes. Overall, CH4 uptakes of CoRE MOFs are similar to those of Hypo1MOFs and higher than those of Hypo2MOFs, while H2 uptakes of Hypo2MOFs surpass those of both CoRE and Hypo1MOFs across all three pressures.

3.3. CH4/H2 Separation Performances of All MOFs

Since ML models provide accurate predictions for the gas uptakes of unseen MOFs, we further used these predictions to compute CH4/H2 selectivities of CoRE, Hypo1, and Hypo2MOF databases. We then compared ML-predicted selectivities to two mathematical models previously proposed in the literature. The first model predicts the selectivities using ΔQ st° of gases, and the second model uses the ratio of ΔQ st° to materials’ porosity. Since we had already computed porosities and the heat of adsorption values of gases at infinite dilution for 16987 MOFs (4331 CORE, 9868 Hypo1, and 2788 Hypo2MOFs), we applied both models. Since the mathematical models were mainly proposed for infinite dilution, we compared the selectivities at 0.1 bar in Figure S6. Results showed that our ML-predicted selectivities (1.7–510) closely matched the simulated selectivities (1.7–500), whereas the first (second) mathematical model significantly overestimated the simulated selectivities, in the range of 1.2–2.9 × 104 (1.1–737). The superior accuracy of our ML models can be attributed to the diverse set of features we used in training, which include chemical, structural, and energetic properties of MOFs. Furthermore, the two correlations were designed for zero coverage, thus they may tend to overestimate selectivities at 0.1 bar.

Figure represents CH4/H2 selectivities of 4331 CoRE MOFs, 98601 Hypo1MOFs, and 23673 Hypo2MOFs at 1 bar. Selectivities were computed as 1.6–85 for CoRE MOFs, 1.5–115 for Hypo1MOFs, and 1.3–48 for Hypo2MOFs. There are three Hypo1MOFs having higher selectivities (87–115) than the maximum selectivity observed for CoRE MOFs. 1048 CoRE, 4088 Hypo1, and 12 Hypo2MOFs have high selectivities, in the range of 40–115.

6.

6

Open in a new tab

Relation between porosities, pore sizes, and CH4/H2 selectivities computed at 1 bar for (a) 4331 CoRE MOF, (b) 98601 Hypo1, and (c) 23673 Hypo2MOFs. The numbers in selectivity bars represent the number of MOFs having selectivities in the given range. Large purple spheres represent the top 10 MOFs having the highest selectivities in each database.

Figure clearly shows that as the pore size and porosity of MOFs decrease, higher selectivities are observed. Figure a,b shows that CoRE and Hypo1MOFs having narrow LCDs (4.0–6.0 Å) and low porosities (0.2–0.4) exhibit higher CH4/H2 selectivities because of the stronger confinement of CH4 molecules in these narrow pores. Since Hypo2MOFs have significantly larger LCDs and porosities compared to the other two databases, they exhibit much lower selectivities (1.3–48) as shown in Figure c. We note that 91% of the Hypo2MOFs have porosities larger than 0.7 and 90% of them have LCDs more than 10 Å. Among the databases that we studied, Hypo1MOFs exhibit the highest selectivities at 1 bar. Besides, 24%, 4%, and 0.05% of CoRE, Hypo1, and Hypo2MOFs, respectively, have selectivities greater than 40, highlighting the significant potential of CoRE MOFs for CH4/H2 separation.

Selectivities of MOFs at 0.1 and 10 bar as a function of their LCD, ASA, and PV values are also given in Figure S7. At 0.1 bar, CH4/H2 selectivities were computed to be in the ranges of 1.7–510, 1.5–796, and 1.4–72 for CoRE, Hypo1, and Hypo2MOF databases, respectively. Selectivities generally decrease as the pressure is increased, since the entropic effects favor H2 adsorption. At 10 bar, CoRE, Hypo1, and Hypo2MOFs have the selectivities varying between 1.6 and 21, 1.4 and 22, and 1.0 and 15, respectively. Overall, CoRE and Hypo1MOFs exhibit higher selectivities compared to Hypo2MOFs at all pressures we studied.

The top performing 10 MOFs with the highest selectivities at 1 bar in each database were analyzed in detail to determine which structural and chemical characteristics of the materials lead to high CH4/H2 selectivities. The large purple spheres shown in Figure are the top 10 CoRE, Hypo1, and Hypo2MOFs having selectivities of 78–85, 71–115, and 42–48, respectively. Table S4 presents their metal types, most frequently occurring linker subunits, and topologies, in addition to their structural properties. The top MOFs identified from each database contain a variety of transition metals (Ag, Cd, Co, Cu, Ni, and Zn). The most common linker subunits found in the top 10 CoRE MOFs are heterocyclic linkers, such as pyridine, histidine, triazole, and imidazole, whereas the top Hypo1 and Hypo2MOFs have carboxylate-, benzoate-, and cubane-based linkers in common. Various types of topologies were detected for the top materials. The top 10 CoRE, Hypo1, and Hypo2MOFs have narrow PLDs (3.8–4.3, 3.8–4.8, and 3.8–4.7 Å), low to medium surface areas (114–982, 113–472, and 115–1069 m2/g), and mediocre porosities (0.36–0.54, 0.33–0.55, and 0.45–0.57). We observed that most MOFs in the Hypo1 (82%), Hypo2 (99%), and CoRE (77%) databases have larger PLDs than the maximum PLDs found in the top 10 materials from each respective database. Similarly, 73% (64%) of CoRE, 97% (99%) of Hypo1, and 99% (99%) of Hypo2 databases have porosities (surface areas) greater than the highest porosity (surface area) values observed in the top 10 MOFs.

The most promising MOFs, NELVAC from the CoRE database, hMOF-6002975 from the Hypo1 database, and ddmof_7440 from the Hypo2MOF database, were further analyzed by generating the adsorption density plots and RDF analysis to illustrate the preferential adsorption sites of gas molecules in the structures. We specifically focused on the interactions between H, C, Zn, V, and Ni and the gas molecules to support the observations from the adsorption density plots, which reveal that gas molecules tend to localize near the organic ligands, metal clusters, or both. Figure a–c shows that both CH4 and H2 molecules are mostly close to the center of pores in addition to interacting with piperazine-, tetrabromocubane-, and pyrene-based linkers for NELVAC, hMOF-6002975, and ddmof_7440, respectively. NELVAC and hMOF-6002975 have rectangular channels, and ddmof_7440 has hexagonal-shaped channels. Normalized RDF plots demonstrate that there is almost a 2 Å difference in the nearest (H–gas) and farthest interactions (Zn–gas) of NELVAC, which means that gas molecules are concentrated in one location, closer to linkers than metals. Similarly, for hMOF-6002975, gas molecules are located closer to linkers (H–gas) rather than metal (V–gas). For ddmof_7440, there is a small difference (∼1 Å) between the nearest (H–gas) and farthest (C–gas) interactions. In addition to linkers, gas molecules also prefer metal parts (Ni) in ddmof_7440. The enhanced CH4/H2 selectivity is associated with the stronger interactions of CH4 molecules with the framework compared to H2. For instance, in NELVAC, the first RDF peak for H–CH4 appears at 3.6 Å with a height of 1.3, indicating a significantly stronger interaction compared to H2, whose corresponding peak occurs at 3.7 Å with a much lower height of 0.3. Similarly, although the trend remains consistent, Zn–CH4 peaks appear at 6.6 Å with a height of 1.7, which is notably higher than the Zn–H2 peak observed at 6.5 Å with a height of 0.48. We note that the structural and chemical integrity of the top-performing CoRE MOF (NELVAC) was verified using the CoRE MOF Web App as reported in the recently updated CoRE MOF 2025 database. For the two hypothetical MOFs, we note that structures are computer-generated, and their synthesizability and stability must be further confirmed with experiments.

7.

7

Open in a new tab

Adsorption density profiles for CH4 and H2 in the top-performing (a) CoRE, (b) Hypo1, and (c) Hypo2MOFs and their normalized RDF plots obtained at 1 bar. Purple and dark blue colors represent the preferentially adsorbed regions by CH4 and H2 molecules, respectively. Carbon, hydrogen, oxygen, zinc, phosphorus, vanadium, bromine, and nickel atoms are represented by gray, white, red, blue, orange, light gray, brown, and green colors, respectively.

We finally compared the CH4/H2 separation performance of MOFs studied in this work to the selectivity of various adsorbents reported in the literature. CH4/H2 selectivities of traditional adsorbents, activated alumina, silica gel, coal carbon, 5A zeolite, and coconut carbon were experimentally reported as 13.0, 14.3, 34.1, 37.9, and 58.3, respectively, using the ratio of Henry’s constants. We identified 1048 (1001) CoRE and 4100 (4727) Hypo MOFs having much larger selectivities, >40 at 1 bar (>60 at 0.1 bar) than all these materials. This comparison highlights the potential of MOFs to replace the conventional porous materials. CH4/H2 selectivities of CAU-17, ZIF-8, and IRMOF-1 were calculated as 43.5, 11, and 8.1, respectively, using the adsorption data obtained from molecular simulations at 1 bar, 298 K, while experiments reported CH4/H2 selectivity of the widely studied Cu-BTC as 19.7 at 1 bar, 298 K. Molecular simulations performed for a larger number of MOFs listed in Table S5 showed that reported selectivities (1.9–75.7, 1.4–76.8, and 1.4–84.7) are similar to the selectivities (1.6–85.4) of CoRE MOFs computed in this work and Hypo1MOFs that we studied exhibit much higher selectivities (1.5–115 at 1 bar) than the previously studied MOFs. This result remarks that a very large number of MOFs with high CH4/H2 separation potentials have not yet been synthesized or synthesized but not tested for this separation. A recent study examined both synthesized and hypothetical covalent organic frameworks (COFs) and reported CH4/H2 selectivities up to 91.9 at 1 bar. The top-performing COF identified for CH4/H2 separation has a narrow PLD (5.3 Å), low surface area (589 m2/g), and low porosity (0.48), similar to the characteristics observed in the top-performing MOFs in our study. The adamantane linker identified in the top-performing COF exhibits structural and functional similarity to the cubane-based linkers found in our top-performing hypothetical MOFs, suggesting that rigid, nonplanar aliphatic linkers may play a key role in enhancing CH4/H2 selectivity.

We note that we focused on ideal CH4/H2 selectivities throughout our study; however, mixture selectivities may differ from ideal selectivities, especially at high pressures. , Therefore, we computed the equimolar CH4/H2 mixture selectivities of the top 10 CoRE, Hypo1, and Hypo2MOFs. Figure S8 shows that ideal and mixture CH4/H2 selectivities are close to each other at 0.1 bar. For example, mixture selectivities of the top 10 CoRE MOFs range from 126 to 683 at 0.1 bar, similar to their ideal selectivities, varying from 128 to 477. At 1 bar, mixture selectivities are higher than the ideal selectivities for both CoRE and hypoMOFs due to the competition between the two gas molecules. This also highlights that the materials identified as promising based on their ideal selectivity are also promising for mixture separation.

We also compared ML-predicted and simulated CH4/H2 selectivities of 4331 CoRE MOFs at 0.1, 1, and 10 bar in addition to the selectivities calculated from the ratios of Henry’s constants in Figure S9. ML-predicted selectivities agree very well with the simulated selectivities at all pressures. While selectivities calculated from the ratio of Henry’s constants can offer a preliminary estimate at low pressure (0.1 bar), they become unreliable for accurately assessing materials’ performance at higher pressures (10 bar). At 0.1 bar, the ML-predicted selectivities and the Henry’s constant ratios for 4331 CoRE MOFs align well (Figure S9a). However, at 1 and 10 bar, more representative pressures for practical conditions, Figure S9b,c, demonstrate that the Henry’s constant ratios substantially overestimate the actual selectivities obtained from simulations. To quantitatively assess the effect of using different selectivity definitions to identify the most selective MOFs, we ranked 4331 MOFs based on the ML-predicted selectivities and selectivities calculated from the ratios of Henry’s constants. We then computed the Spearman rank correlation coefficient (SRCC). At 0.1 bar, the MOF rankings derived from ML-predicted selectivities and Henry’s ratios are highly correlated (SRCC: 0.998). In contrast, at 10 bar, the MOF rankings show considerable differences, with an SRCC value of 0.449. These findings emphasize the importance of computing selectivities based on gas uptakes at relevant operating pressures rather than relying solely on the ratio of Henry’s constants.

4. Conclusion

This study presents a comprehensive computational framework that integrates molecular simulations and ML to evaluate the CH4/H2 separation performances of 126605 different types of synthesized and hypothetical MOFs. Developing ML models utilizing only 14 features enabled accurate estimations for CH4 and H2 uptakes of 122274 hypothetical MOFs at three different pressures in just six minutes on a computer with 8 CPUs. We made all structural, chemical, and energetic features of 126605 different types of synthesized and hypothetical MOFs available on GitHub for public use, so that gas adsorption and separation performance of MOFs can be predicted by ML in just minutes. This is a significant improvement compared to direct molecular simulations of hundreds of thousands of materials, which would typically require months of computation on hundreds of CPUs. Thus, our work highlights the impact of combining molecular simulations with ML to accelerate the exploration of MOFs for target gas adsorption and separation applications. Our results showed that Hypo1MOFs, specifically the ones with narrow pores, transition metals, and linkers incorporating carboxylate, benzoate, and cubane, have high CH4/H2 selectivities up to 115 at 1 bar, outperforming traditional adsorbents and some synthesized MOFs. These results will be useful for directing experimental efforts in designing new MOF materials with high CH4/H2 selectivity.

Supplementary Material

jp5c02779_si_001.pdf (2.6MB, pdf)

Acknowledgments

Dedicated to Prof. Karl Johnson on his retirement. Prof. Johnson’s remarkable contributions to the field of molecular simulations of materials for energy and environmental applications have left an important mark on our research. S.K. acknowledges funding by the European Union (ERC, STARLET, 101124002). Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them.

Force field parameters, all structural, chemical, and energetic features of 126605 MOFs, raw adsorption data from molecular simulations, ML scripts developed in this work, and ML-predicted data are available on GitHub: https://github.com/pelinsezgin/CH4H2_ML.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jpcc.5c02779.

  • Correlation heat map of physical, chemical, and energetic features of MOFs and their simulated CH4 and H2 uptakes; SHAP feature importance distributions used in predicting CH4 and H2 uptakes; comparison of ML-predicted and simulated uptakes for 9868 Hypo1MOFs and 2788 Hypo2MOFs; comparison of model-predicted and simulated CH4/H2 selectivities for 16987 MOFs; relation between CH4/H2 selectivities and structural properties of 4331 CoRE, 98601 Hypo1, and 23673 Hypo2MOFs at three different pressures; comparison of equimolar CH4/H2 mixture selectivities and ideal selectivities of top 10 CoRE, Hypo1, and Hypo2MOFs; comparison of ML-predicted and simulated CH4/H2 selectivities with selectivities calculated from the ratios of Henry’s constants for 4331 CoRE MOFs; 14 descriptors used for ML model development; quantities calculated to evaluate the ML models’ accuracy; ML pipelines and parameters; structural and chemical properties of the top 10 MOFs for CH4/H2 separation; and comparison of selectivities reported in the literature (PDF)

The authors declare no competing financial interest.

Published as part of The Journal of Physical Chemistry C special issue “J. Karl Johnson Festschrift”.

References

  1. Herm Z. R., Krishna R., Long J. R.. CO2/CH4, CH4/H2 and CO2/CH4/H2 separations at high pressures using Mg2 (dobdc) Microporous Mesoporous Mater. 2012;151:481–487. doi: 10.1016/j.micromeso.2011.09.004. [DOI] [Google Scholar]
  2. Sholl D. S., Lively R. P.. Seven chemical separations to change the world. Nature. 2016;532(7600):435–437. doi: 10.1038/532435a. [DOI] [PubMed] [Google Scholar]
  3. Chen H., Sholl D. S.. Predictions of selectivity and flux for CH4/H2 separations using single walled carbon nanotubes as membranes. J. Membr. Sci. 2006;269(1–2):152–160. doi: 10.1016/j.memsci.2005.06.030. [DOI] [Google Scholar]
  4. Poshusta J. C., Tuan V. A., Pape E. A., Noble R. D., Falconer J. L.. Separation of light gas mixtures using SAPO-34 membranes. AIChE J. 2000;46(4):779–789. doi: 10.1002/aic.690460412. [DOI] [Google Scholar]
  5. Daglar H., Keskin S.. Recent advances, opportunities, and challenges in high-throughput computational screening of MOFs for gas separations. Coord. Chem. Rev. 2020;422:213470. doi: 10.1016/j.ccr.2020.213470. [DOI] [Google Scholar]
  6. Farha O. K., Eryazici I., Jeong N. C., Hauser B. G., Wilmer C. E., Sarjeant A. A., Snurr R. Q., Nguyen S. T., Yazaydın A. O. z. r., Hupp J. T.. Metal–organic framework materials with ultrahigh surface areas: is the sky the limit? J. Am. Chem. Soc. 2012;134(36):15016–15021. doi: 10.1021/ja3055639. [DOI] [PubMed] [Google Scholar]
  7. Xuan W., Zhu C., Liu Y., Cui Y.. Mesoporous metal–organic framework materials. Chem. Soc. Rev. 2012;41(5):1677–1695. doi: 10.1039/C1CS15196G. [DOI] [PubMed] [Google Scholar]
  8. Furukawa H., Cordova K. E., O’Keeffe M., Yaghi O. M.. The chemistry and applications of metal-organic frameworks. Science. 2013;341(6149):1230444. doi: 10.1126/science.1230444. [DOI] [PubMed] [Google Scholar]
  9. Liu J., Keskin S., Sholl D. S., Johnson J. K.. Molecular simulations and theoretical predictions for adsorption and diffusion of CH4/H2 and CO2/CH4 mixtures in ZIFs. J. Phys. Chem. C. 2011;115(25):12560–12566. doi: 10.1021/jp203053h. [DOI] [Google Scholar]
  10. Ruffley J. P., Goodenough I., Luo T.-Y., Richard M., Borguet E., Rosi N. L., Johnson J. K.. Design, synthesis, and characterization of metal–organic frameworks for enhanced sorption of chemical warfare agent simulants. J. Phys. Chem. C. 2019;123(32):19748–19758. doi: 10.1021/acs.jpcc.9b05574. [DOI] [Google Scholar]
  11. Gómez-Gualdrón D. A., Wilmer C. E., Farha O. K., Hupp J. T., Snurr R. Q.. Exploring the limits of methane storage and delivery in nanoporous materials. J. Phys. Chem. C. 2014;118(13):6941–6951. doi: 10.1021/jp502359q. [DOI] [Google Scholar]
  12. Wilmer C. E., Leaf M., Lee C. Y., Farha O. K., Hauser B. G., Hupp J. T., Snurr R. Q.. Large-scale screening of hypothetical metal–organic frameworks. Nature Chem. 2012;4(2):83–89. doi: 10.1038/nchem.1192. [DOI] [PubMed] [Google Scholar]
  13. Kim S. Y., Han S., Lee S., Kang J. H., Yoon S., Park W., Shin M. W., Kim J., Chung Y. G., Bae Y. S.. Discovery of High-Performing Metal–Organic Frameworks for On-Board Methane Storage and Delivery via LNG–ANG Coupling: High-Throughput Screening, Machine Learning, and Experimental Validation. Advanced Science. 2022;9(21):2201559. doi: 10.1002/advs.202201559. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Colón Y. J., Fairen-Jimenez D., Wilmer C. E., Snurr R. Q.. High-throughput screening of porous crystalline materials for hydrogen storage capacity near room temperature. J. Phys. Chem. C. 2014;118(10):5383–5389. doi: 10.1021/jp4122326. [DOI] [Google Scholar]
  15. Bucior B. J., Bobbitt N. S., Islamoglu T., Goswami S., Gopalan A., Yildirim T., Farha O. K., Bagheri N., Snurr R. Q.. Energy-based descriptors to rapidly predict hydrogen storage in metal–organic frameworks. Molecular Systems Design & Engineering. 2019;4(1):162–174. doi: 10.1039/C8ME00050F. [DOI] [Google Scholar]
  16. Bobbitt N. S., Chen J., Snurr R. Q.. High-throughput screening of metal–organic frameworks for hydrogen storage at cryogenic temperature. J. Phys. Chem. C. 2016;120(48):27328–27341. doi: 10.1021/acs.jpcc.6b08729. [DOI] [Google Scholar]
  17. Zhou W., Wu H., Hartman M. R., Yildirim T.. Hydrogen and methane adsorption in metal– organic frameworks: a high-pressure volumetric study. J. Phys. Chem. C. 2007;111(44):16131–16137. doi: 10.1021/jp074889i. [DOI] [Google Scholar]
  18. Hou X.-J., He P., Li H., Wang X.. Understanding the adsorption mechanism of C2H2, CO2, and CH4 in isostructural metal–organic frameworks with coordinatively unsaturated metal sites. J. Phys. Chem. C. 2013;117(6):2824–2834. doi: 10.1021/jp310517r. [DOI] [Google Scholar]
  19. Chen H., Snurr R. Q.. Computational screening of metal-catecholate-functionalized metal–organic frameworks for room-temperature hydrogen storage. J. Phys. Chem. C. 2021;125(39):21701–21708. doi: 10.1021/acs.jpcc.1c06549. [DOI] [Google Scholar]
  20. Wu D., Wang C., Liu B., Liu D., Yang Q., Zhong C.. Large-scale computational screening of metal-organic frameworks for CH4/H2 separation. AIChE journal. 2012;58(7):2078–2084. doi: 10.1002/aic.12744. [DOI] [Google Scholar]
  21. Basdogan Y., Sezginel K. B., Keskin S.. Identifying highly selective metal organic frameworks for CH4/H2 separations using computational tools. Ind. Eng. Chem. Res. 2015;54(34):8479–8491. doi: 10.1021/acs.iecr.5b01901. [DOI] [Google Scholar]
  22. Junior M. J. C., Wang Y., Wu X., Cai W.. Computational screening of metal-organic frameworks with open copper sites for hydrogen purification. Int. J. Hydrogen Energy. 2020;45(51):27320–27330. doi: 10.1016/j.ijhydene.2020.07.041. [DOI] [Google Scholar]
  23. Daglar H., Gulbalkan H. C., Avci G., Aksu G. O., Altundal O. F., Altintas C., Erucar I., Keskin S.. Effect of metal–organic framework (MOF) database selection on the assessment of gas storage and separation potentials of MOFs. Angew. Chem., Int. Ed. 2021;60(14):7828–7837. doi: 10.1002/anie.202015250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Wang S., Cheng M., Luo L., Ji X., Liu C., Bi K., Zhou L.. High-throughput screening of metal–organic frameworks for hydrogen purification. Chem. Eng. J. 2023;451:138436. doi: 10.1016/j.cej.2022.138436. [DOI] [Google Scholar]
  25. Liu X., Chen H., Yuan J., Li X., Yu J.. Evaluation and screening of metal-organic frameworks for the adsorption and separation of methane and hydrogen. Mater. Today Commun. 2022;33:104602. doi: 10.1016/j.mtcomm.2022.104602. [DOI] [Google Scholar]
  26. Altintas C., Erucar I., Keskin S.. High-throughput computational screening of the metal organic framework database for CH4/H2 separations. ACS Appl. Mater. Interfaces. 2018;10(4):3668–3679. doi: 10.1021/acsami.7b18037. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Demir H., Keskin S.. Computational investigation of multifunctional MOFs for adsorption and membrane-based separation of CF4/CH4, CH4/H2, CH4/N2, and N2/H2 mixtures. Molecular Systems Design & Engineering. 2022;7(12):1707–1721. doi: 10.1039/D2ME00130F. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Pardakhti M., Moharreri E., Wanik D., Suib S. L., Srivastava R.. Machine learning using combined structural and chemical descriptors for prediction of methane adsorption performance of metal organic frameworks (MOFs) ACS Comb. Sci. 2017;19(10):640–645. doi: 10.1021/acscombsci.7b00056. [DOI] [PubMed] [Google Scholar]
  29. Wu X., Xiang S., Su J., Cai W.. Understanding quantitative relationship between methane storage capacities and characteristic properties of metal–organic frameworks based on machine learning. J. Phys. Chem. C. 2019;123(14):8550–8559. doi: 10.1021/acs.jpcc.8b11793. [DOI] [Google Scholar]
  30. Sarikas A. P., Fanourgakis G. S., Tylianakis E., Gkagkas K., Froudakis G. E.. Comparison of Energy-Based Machine Learning Descriptors for Gas Adsorption. J. Phys. Chem. C. 2023;127(43):20995–21005. doi: 10.1021/acs.jpcc.3c04223. [DOI] [Google Scholar]
  31. Thornton A. W., Simon C. M., Kim J., Kwon O., Deeg K. S., Konstas K., Pas S. J., Hill M. R., Winkler D. A., Haranczyk M.. et al. Materials genome in action: identifying the performance limits of physical hydrogen storage. Chem. Mater. 2017;29(7):2844–2854. doi: 10.1021/acs.chemmater.6b04933. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Majumdar S., Moosavi S. M., Jablonka K. M., Ongari D., Smit B.. Diversifying databases of metal organic frameworks for high-throughput computational screening. ACS Appl. Mater. Interfaces. 2021;13(51):61004–61014. doi: 10.1021/acsami.1c16220. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Willems T. F., Rycroft C. H., Kazi M., Meza J. C., Haranczyk M.. Algorithms and tools for high-throughput geometry-based analysis of crystalline porous materials. Microporous Mesoporous Mater. 2012;149(1):134–141. doi: 10.1016/j.micromeso.2011.08.020. [DOI] [Google Scholar]
  34. Chung Y. G., Haldoupis E., Bucior B. J., Haranczyk M., Lee S., Zhang H., Vogiatzis K. D., Milisavljevic M., Ling S., Camp J. S.. et al. Advances, updates, and analytics for the computation-ready, experimental metal–organic framework database: CoRE MOF 2019. J. Chem. Eng. Data. 2019;64(12):5985–5998. doi: 10.1021/acs.jced.9b00835. [DOI] [Google Scholar]
  35. Chung, Y. G. , Haldoupis, E. , Bucior, B. J. , Haranczyk, M. , Lee, S. , Zhang, H. , Vogiatzis, K. D. , Milisavljevic, M. , Ling, S. , Camp, J. S. . et al. , Computation-Ready Experimental Metal-Organic Framework (CoRE MOF) 2019 Dataset, 2019. https://zenodo.org/records/14184621.
  36. Dubbeldam D., Calero S., Ellis D. E., Snurr R. Q.. RASPA: molecular simulation software for adsorption and diffusion in flexible nanoporous materials. Mol. Simul. 2016;42(2):81–101. doi: 10.1080/08927022.2015.1010082. [DOI] [Google Scholar]
  37. Rappé A. K., Casewit C. J., Colwell K., Goddard W. A., Skiff W. M.. UFF, a full periodic table force field for molecular mechanics and molecular dynamics simulations. J. Am. Chem. Soc. 1992;114(25):10024–10035. doi: 10.1021/ja00051a040. [DOI] [Google Scholar]
  38. Martin M. G., Siepmann J. I.. Transferable potentials for phase equilibria. 1. United-atom description of n-alkanes. J. Phys. Chem. B. 1998;102(14):2569–2577. doi: 10.1021/jp972543+. [DOI] [Google Scholar]
  39. Buch V.. Path integral simulations of mixed para-D2 and ortho-D2 clusters: The orientational effects. J. Chem. Phys. 1994;100(10):7610–7629. doi: 10.1063/1.466854. [DOI] [Google Scholar]
  40. Frenkel, D. ; Smit, B. . Understanding molecular simulation: from algorithms to applications; Elsevier, 2023. [Google Scholar]
  41. Dubbeldam D., Calero S., Vlugt T. J.. iRASPA: GPU-accelerated visualization software for materials scientists. Mol. Simul. 2018;44(8):653–676. doi: 10.1080/08927022.2018.1426855. [DOI] [Google Scholar]
  42. Le T. T., Fu W., Moore J. H.. Scaling tree-based automated machine learning to biomedical big data with a feature set selector. Bioinformatics. 2020;36(1):250–256. doi: 10.1093/bioinformatics/btz470. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Martinsson P.-G., Rokhlin V., Tygert M.. A randomized algorithm for the decomposition of matrices. Appl. Comput. Harmon. Anal. 2011;30(1):47–68. doi: 10.1016/j.acha.2010.02.003. [DOI] [Google Scholar]
  44. Liang H., Jiang K., Yan T.-A., Chen G.-H.. XGBoost: an optimal machine learning model with just structural features to discover MOF adsorbents of Xe/Kr. ACS Omega. 2021;6(13):9066–9076. doi: 10.1021/acsomega.1c00100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Ogutu J. O., Schulz-Streeck T., Piepho H. P.. Genomic selection using regularized linear regression models: ridge regression, lasso, elastic net and their extensions. BMC Proc. 2012;6:S10. doi: 10.1186/1753-6561-6-S2-S10. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Calvetti D., Morigi S., Reichel L., Sgallari F.. Tikhonov regularization and the L-curve for large discrete ill-posed problems. J. Comput. Appl. Math. 2000;123(1–2):423–446. doi: 10.1016/S0377-0427(00)00414-3. [DOI] [Google Scholar]
  47. Lundberg S. M., Erion G., Chen H., DeGrave A., Prutkin J. M., Nair B., Katz R., Himmelfarb J., Bansal N., Lee S.-I.. From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020;2(1):56–67. doi: 10.1038/s42256-019-0138-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Yang Z., Peng X., Cao D.. Carbon dioxide capture by PAFs and an efficient strategy to fast screen porous materials for gas separation. J. Phys. Chem. C. 2013;117(16):8353–8364. doi: 10.1021/jp402488r. [DOI] [Google Scholar]
  49. Frost H., Düren T., Snurr R. Q.. Effects of surface area, free volume, and heat of adsorption on hydrogen uptake in metal– organic frameworks. J. Phys. Chem. B. 2006;110(19):9565–9570. doi: 10.1021/jp060433+. [DOI] [PubMed] [Google Scholar]
  50. Düren T., Sarkisov L., Yaghi O. M., Snurr R. Q.. Design of new materials for methane storage. Langmuir. 2004;20(7):2683–2689. doi: 10.1021/la0355500. [DOI] [PubMed] [Google Scholar]
  51. Zhao G., Brabson L. M., Chheda S., Huang J., Kim H., Liu K., Mochida K., Pham T. D., Terrones G. G., Terrones G. G.. et al. CoRE MOF DB: A curated experimental metal-organic framework database with machine-learned properties for integrated material-process screening. Matter. 2025;8:102140. doi: 10.1016/j.matt.2025.102140. [DOI] [Google Scholar]
  52. Ludwig, K. Final ReportDevelopment of New Pressure Swing Adsorption (PSA) Technology to Recover High Valued Products from Chemical Plant and Refinery Waste Systems; Air Products and Chemicals Inc., 2004. 10.2172/878902. [DOI] [Google Scholar]
  53. Zhou B., Li W., Zhang J.. Theoretical simulation of CH4 separation from H2 in CAU-17 materials. J. Phys. Chem. C. 2017;121(37):20197–20204. doi: 10.1021/acs.jpcc.7b07108. [DOI] [Google Scholar]
  54. Sezginel K. B., Keskin S., Uzun A.. Tuning the gas separation performance of CuBTC by ionic liquid incorporation. Langmuir. 2016;32(4):1139–1147. doi: 10.1021/acs.langmuir.5b04123. [DOI] [PubMed] [Google Scholar]
  55. Aksu G. O., Keskin S.. The COF Space: Materials Features, Gas Adsorption, and Separation Performances Assessed by Machine Learning. ACS Mater. Lett. 2025;7:954–960. doi: 10.1021/acsmaterialslett.4c02594. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

jp5c02779_si_001.pdf (2.6MB, pdf)

Data Availability Statement

Force field parameters, all structural, chemical, and energetic features of 126605 MOFs, raw adsorption data from molecular simulations, ML scripts developed in this work, and ML-predicted data are available on GitHub: https://github.com/pelinsezgin/CH4H2_ML.

Articles from The Journal of Physical Chemistry. C, Nanomaterials and Interfaces are provided here courtesy of American Chemical Society

RESOURCES