Discovering metal-organic framework/polymer mixed-matrix membranes via machine learning for CO2 separation Authors: Feride Neva Yüngül, Seda Keskin Published: 2026 ABSTRACT AbstractMixed-matrix membranes (MMMs) that incorporate metal-organic frameworks (MOFs) as fillers surpass the permeability-selectivity trade-off of polymeric membranes. However, the enormous chemical diversity of MOFs and high computational cost of molecular simulations have hindered the systematic evaluation of large numbers of MOF/polymer MMMs. In this work, we present a data-driven discovery framework that integrates molecular simulations and machine learning (ML) to predict the CO2, CH4, N2, and H2 permeabilities of 104,196 different types of MOF/polymer MMMs. Molecular simulations were performed to obtain gas adsorption and diffusion properties for 3,982 synthesized MOFs and 4,701 hypothetical MOFs. Leveraging this dataset, we developed three ML approaches to enable rapid prediction of MMMs’ permeabilities and benchmarked their performance against both simulation results and experimental measurements. Our results revealed that many MOF/polymer MMMs exceed the upper bounds by achieving very high CO2 permeabilities and MOFs’ structural and chemical features play a decisive role in fine-tuning MMM performance.Subject terms: Metal-organic frameworks, Metal-organic frameworks Introduction Gas separation membranes have become an increasingly important alternative to energy-intensive separation technologies thanks to their continuous operation, low energy consumption, and scalability1,2. Polymeric membranes currently dominate the industrial applications such as natural gas purification (CO2/CH4), flue gas separation (CO2/N2), and hydrogen recovery (H2/CO2) because they are easy to process and economical3. However, the separation performance of polymeric membranes is constrained by the well-known permeability-selectivity trade-off as described by the Robeson upper bound4. Despite continuous efforts to design new polymers, most polymers remain confined within this limitation, highlighting the need for alternative strategies to achieve high permeability and high selectivity simultaneously. Incorporating porous fillers into polymers has long been considered as an effective strategy to overcome this intrinsic trade-off5. Mixed-matrix membranes (MMMs), in which porous organic or inorganic fillers are dispersed within a continuous polymer phase, have enhanced gas transport properties of pure polymers. Early MMM studies focused on activated carbons and zeolites as fillers, and these membranes generally exhibited improved separation performance compared to pristine polymers6,7. Among the wide range of filler materials, metal-organic frameworks (MOFs) and covalent organic frameworks (COFs) have recently emerged as promising candidates thanks to their exceptionally high surface areas, large porosities, and tunable pore structures8. MOF/polymer MMMs combine the mechanical robustness and easy processability of polymers with the structural diversity of MOFs, creating hybrid membranes that surpass the performance of either component alone9. Although several different types of MOF/polymer MMMs have been fabricated and tested for different gas separations10,11, the majority of these studies have relied on a limited number of MOFs. With over 150,000 structures currently reported12, the MOF library continues to expand, making it practically impossible to experimentally test all possible MOFs for MMM applications. Molecular simulations play a critical role in screening large numbers of MOF/polymer MMMs by providing molecular-level insight into gas permeation while avoiding costly and time-consuming experiments9,13. In this approach, gas adsorption data obtained from Grand Canonical Monte Carlo (GCMC) simulations and gas diffusivities computed by Molecular Dynamics (MD) are used to quantify the gas permeabilities through MOFs14. However, conducting computationally intensive MD simulations for large numbers of MOFs is prohibitively expensive, underscoring the need for faster and more efficient predictive methodologies. Machine learning (ML) has been recently integrated with molecular simulations in the field of MOFs not only to accelerate the simulations but also to analyze the complex relationships between structures and their performances15,16. For example, Daglar and Keskin17 combined molecular simulations with ML to evaluate the gas separation potential of 31,494 different types of MOF/polymer MMMs composed of 5249 MOFs and 6 polymers. They used ML models to predict uptake (Ni) and diffusivity (Di) of He, H2, N2, and CH4 gases in MOFs and then computed permeabilities. This data was combined with the polymers’ permeability through Maxwell model to assess MOF/polymer MMMs. Predicted permeability of MMMs were shown to agree with the experimental reports. Sholl’s group14 combined ML-predicted gas adsorption and diffusion properties of MOFs with polymer permeabilities obtained from Polymer Genome18 through the Maxwell model. They screened 131,722 types of MOF/polymer MMMs for CO2/CH4, O2/N2, and He/H2 separations and identified the promising MMMs that can surpass the upper bounds. The CO2 separation performance of 54,117 MOF/polymer MMMs, constructed from 9 polymers and 6,013 MOFs, was examined by using ML models19. GCMC and MD simulations were performed to obtain gas uptakes and diffusivities of MOFs and that simulation data was merged with experimental polymer permeabilities via the Maxwell model to assess MOF/polymer MMMs’ separation performances. ML models were then trained to predict the CO2 permeability, CO2/N2 selectivity, and the trade-off metric (TSP, the ratio of selectivity to permeability combined into a single trade-off indicator) of MOF/polymer MMMs. Using a similar approach, molecular simulation data of 10,143 MOFs were combined with experimentally reported data of 26 polymers to construct a dataset of 456,872 MOF/polymer MMMs for He/CH4 separation20. ML models were then trained to estimate He permeability and He/CH4 selectivity of MOF/polymer MMMs. The studies we reviewed above used the molecular simulation results of MOFs to train ML models while some recent works trained their models using the experimentally reported MOF/polymer MMM datasets. Yao et al.21 developed an artificial neural network model to predict CO2 permeability and CO2/N2 selectivity of MOF/polymer MMMs by compiling experimentally reported data for 291 MOF/polymer MMMs comprising 43 MOFs and 19 polymers. Guan et al.22 trained an ML model using CO2 permeability and CO2/CH4 selectivity data for MOF/polymer MMMs, belonging to 36 MOFs and 41 polymers, to predict the relative CO2 permeability and CO2/CH4 selectivity, defined as the ratio of the MMM performance to that of the pure polymer. These models guided the synthesis of two new MOFs that were subsequently used to fabricate MMMs with polymers, Pebax 2533 and PIM-1 (Polymer of Intrinsic Microporosity-1). ML predictions showed a good agreement with the experimentally measured CO2 permeabilities. A subset of the Guan’s dataset, comprising 376 MOF/polymer MMMs’ CO2 permeability and CO2/CH4 selectivity, was used in a following work to train an ML model for predicting CO2 permeability and CO2/CH4 selectivity of MOF/polymer MMMs23. As summarized in this literature review, two main strategies currently exist for predicting the gas permeability and selectivity of large numbers of MOF/polymer MMMs: either training ML models to estimate the fundamental gas adsorption and transport properties of MOFs (Ni and Di), or training ML models to directly predict gas permeability of MOF/polymer MMMs. A systematic, quantitative assessment of which target variable yields the most accurate, generalizable, and physically meaningful predictions for MOF/polymer MMMs is still lacking. Furthermore, previous studies have typically focused on a limited set of MOFs (a small variety of synthesized MOFs), polymers, and gas molecules, which restricts the ability to derive generalizable guidelines for designing novel MOF/polymer MMMs suitable for various types of gas separations. In this work, we addressed these key gaps by providing a comprehensive, comparative benchmark of three distinct computational approaches for evaluating the CO2, CH4, N2, and H2 permeability of 104,196 different types of MOF/polymer MMMs, composed of 8683 MOFs (both experimentally synthesized and hypothetical structures) and 12 polymers. We performed molecular simulations to compute adsorption and diffusion data of CO2, CH4, N2, and H2 gases in 3,982 experimentally synthesized and 4701 hypothetical MOF (hMOF) structures. These simulation results were then used to train ML models following three approaches: (1) Developing ML models for predicting MOFs’ gas uptakes and diffusivities, calculating MOFs’ gas permeabilities, and subsequently integrating them into the Maxwell model to assess MOF/polymer MMMs; (2) developing ML models for directly predicting MOFs’ gas permeabilities followed by Maxwell model integration to assess MOF/polymer MMMs; (3) developing ML models for the direct prediction of MOF/polymer MMMs’ gas permeabilities. We compared the outcomes of these three approaches against a very large body of experimentally reported gas permeabilities of MOF/polymer MMMs. By systematically and quantitatively evaluating CO2/N2, CO2/CH4, and H2/CO2 separation performances of MOF/polymer MMMs using these approaches, we identified the most promising MOF/polymer MMMs for each separation. We then extracted the key structural features of the best MOFs for a target polymer and provided material-specific design guidelines that enable the rational design of next-generation MOF/polymer MMMs. Results and Discussion Figure 1 represents the computational methodology combining molecular simulations and ML to study 104,196 different types of MOF/polymer MMMs for CO2/CH4, CO2/N2, and H2/CO2 separations. Computational workflow of this study. (1) GCMC and MD simulations to obtain gas uptake, diffusivity, and permeability of MOFs; (2) collection of experimental permeability data of polymers; (3) calculation of permeability of MOF/polymer MMMs using the Maxwell model; (4) analysis and construction of the feature space for MOFs and polymers; (5) development of ML models using three approaches; (6) comparison of ML-predicted, simulated, and experimentally reported permeability data for MMMs; (7) analysis of structure-performance relations for MMMs. To examine how the choice of target feature in ML models affects both predictive accuracy and practical applicability, we compared three approaches that differ in the level at which gas permeability properties are learned, either at the MOF level (Approach 1 and 2) or at the MMM level (Approach 3). The MMM permeability (PMMM,i) predictions of three approaches were compared with the simulation results for 104,196 MOF/polymer MMMs and finally validated against experimentally reported permeability data of 30 different types of MMMs. This comparison allows us to assess which approach yields more generalizable predictions across different gas species, MOFs, and polymers. Approach (1) Fig. 2 shows the performance of gas-specific ML models that were trained to predict uptakes and diffusivities of CO2, CH4, N2, and H2 gases in MOFs. Since PMOF,i (gas permeability of the MOFs) was directly computed as the product of Ni (gas uptake) and Di (gas diffusivity), accurate prediction of both is essential for obtaining reliable permeability estimates. CO2 and N2 models were each trained on 8,683 MOFs, including 3982 computation-ready, experimental MOF database (CoRE MOF) and 4701 hMOFs; while CH4 and H2 models were trained on 8,128 and 8,139 MOFs (3444 CoRE MOFs and 4684 hMOFs; 3,444 CoRE MOFs and 4,695 hMOFs, respectively). The uptake models exhibit strong predictive accuracy across all gases, although the level of agreement depends on the gas type. For CO2, CH4, N2, and H2, the models achieved a test set R2 of 0.911, 0.951, 0.884, and 0.999, respectively. To further improve the interpretability of these models, Shapley Additive exPlanations (SHAP)24 analysis was performed and feature importance bar charts presented in the left side of Fig. 2 were generated based on the mean absolute SHAP values. Results revealed that heat of adsorption, Qst, is the dominant descriptor for CO2, CH4, and N2. Qst was followed by ASA (accessible surface area) and LCD (the largest cavity diameter) for CO2 and CH4, whereas for N2, ASA and PV (pore volume) emerged as the next important descriptors. Features related to available space in MOFs such as PV and ρ (density) were the influential descriptors for H2 since it has the weakest interactions with MOFs. Machine learning predictions of gas adsorption and diffusion properties of MOFs. Comparison of ML with simulation data to predict uptakes (left) and diffusivities (right) of (a) CO2, (b) CH4, (c) N2, and (d) H2 in MOFs. The bar figures represent the importance of the three most influential features used in the ML models. The diffusivity models also achieved high predictive accuracy for CO2 (test set R2 of 0.959), CH4 (R2 of 0.971), N2 (R2 of 0.983), and H2 (R2 of 0.954) as shown in Fig. 2. A logarithmic scale was used in these ML models since diffusion rates span several orders of magnitude across different MOFs, making log-transformed values more suitable for regression and physical interpretation. The feature importance values of these models, presented in the right side of Fig. 2, revealed that the most influential descriptor was logD0 (infinite dilution diffusivity) followed by ϕ (porosity) and Qst for CO2, CH4, N2; by PV and Qst for H2. SHAP plots for the gas-specific ML models that were trained to predict uptakes and diffusivities of CO2, CH4, N2, and H2 gases in MOFs are presented in Figs. S1, S2. Once gas uptakes and diffusivities were predicted by ML models, PMOF,i values were computed and fed into the Maxwell model to predict the gas permeabilities of MMMs across thousands of MOF-polymer combinations. Figure 3 represents the gas permeability results of MOF/polymer MMMs: 104,196 MMMs for CO2, 73,251 MMMs for CH4, 78,147 MMMs for N2, and 97,536 MMMs for H2. There is a good agreement between ML-predicted and simulated permeabilities across all gases as evidenced by high R2 values and low mean absolute percentage error (MAPE) values, indicating that precise prediction of gas adsorption and diffusion properties of MOFs enables an accurate transfer of this information to MMM performance. Since MMM permeability values span several orders of magnitude depending on the gas, MOF, and polymer type, the performance of the ML models was evaluated using the R2 and MAPE by categorizing MMM permeabilities into regions as summarized in Table S1. For a given gas, each permeability region was constructed to contain an equal number of MOF/polymer MMMs, ensuring a balanced comparison across all approaches. As shown in Fig. 3a, the ML predictions for CO2 show good agreement with the simulated data over a broad range of permeability values (0.001–89,894 Barrer). The MAPE remains below 0.5% in the low to medium permeability regions (0.001–4020 Barrer). Even in the high-permeability region (10,438–89,783 Barrer), dominated by highly permeable polymers such as PTMGP, PTMSP, PIM-TMN-Trip, the average MAPE remains low at 1.1% and the correlation remains high (R2 = 0.997). We note that out of the total 104,196 MMMs evaluated for CO2 permeability, the deviation between ML predictions and simulations exceeds 30% for only 45 MMMs. ML-predicted and simulated gas permeabilities of MOF/polymer MMMs using Approach (1). Comparison of ML-predicted (Approach (1)) and simulated gas permeabilities of MOF/polymer MMMs for (a) CO2, (b) CH4, (c) N2, and (d) H2. In Fig. 3b, ML predictions show excellent agreement with the simulated data across the entire permeability range for CH4 (0.4-12,650 Barrer). In the low and medium permeability regions (0.4-219 Barrer), perfect correlations (R2 = 1.0) and negligibly small MAPE values (< 0.005%) indicate highly accurate predictions. In the high-permeability region (2718–12,653 Barrer), the MAPE increases 0.3%, with only 9 MMMs exhibiting errors exceeding 30%, while the correlation remains very high (R2 = 0.999). For N2 and H2 shown in Fig. 3c, d, ML predictions show excellent agreement with simulated permeabilities across the entire permeability ranges (0.4–6,196 Barrer for N2 and 0.1–32,751 Barrer for H2), with perfect correlations (R2 = 1.0) and very low MAPE values (< 0.1%) in the low to medium permeability regimes. In high-permeability region, the MAPE increases to 0.3% and 0.8% for N2 and H2, respectively. Overall, Approach (1) gives MMM permeability predictions with maximum MAPE deviations of 1.1%, 0.3%, 0.3%, and 0.8% from the simulation results for CO2, CH4, N2, and H2, respectively. Approach (2) Fig. 4 shows the performance of gas-specific ML models that were trained to predict permeabilities of CO2, CH4, N2, and H2 gases in MOFs using the same MOF datasets as in Approach (1). Since PMOF,i span several orders of magnitude, the ML models were trained using logarithmic permeability values. As shown in Fig. 4a–d, the predictive performance of the models varies across gases. PMOF,i models for CO2 and CH4 achieve moderate test-set R2 values of 0.826 and 0.793, respectively. Substantially higher predictive accuracy was obtained for N2 and H2, with test-set R2 values of 0.951 and 0.955, respectively. For all gases, the logarithm of the infinite-dilution gas permeability (logPMOF,i0) emerges as the dominant feature, reflecting its role in directly capturing the combined contributions of gas adsorption and transport within the MOFs. Here, we note that computation of PMOF,i0 requires significantly less simulation time than that of PMOF,i since the former only requires the Henry’s coefficient and infinite-dilution diffusivities, neglecting the computationally demanding calculation of guest-guest interactions. For CO2, Qst and ASA emerged as the next most important features, showing that adsorption strength and accessible pore area remain important contributors to gas permeability. In the case of CH4, ρ and Qst emerge as secondary descriptors, indicating the influence of both framework compactness and adsorption energetics. For N2, ρ and ϕ were identified as the next most influential features, whereas for H2, PV and ρ played an important role. Machine learning predictions of gas permeabilities of MOFs. Comparison of ML with simulation data for permeabilities of (a) CO2, (b) CH4, (c) N2, and (d) H2 in MOFs. The bar figures represent the importance of the three most influential features used in the ML models. ML-predicted PMOF,i values were then integrated into the Maxwell model to obtain gas permeabilities of MOF/polymer MMMs. Figure 5 compares ML-predicted and simulated PMMM,i values obtained using Approach (2). The same MMM datasets used in the Approach (1) were considered here, 104,196 MMMs for CO2, 73,251 MMMs for CH4, 78,147 MMMs for N2, and 97,536 MMMs for H2. A strong agreement was observed between ML and simulations for all gases, as evidenced by high R2 values and low MAPE values across the entire permeability regions as given in Table S2. These results demonstrate that directly predicting PMOF,i provides an effective and robust pathway for transferring MOF-level transport properties to MMM performance within the Maxwell framework. ML-predicted and simulated gas permeabilities of MOF/polymer MMMs using Approach (2). Comparison of ML-predicted (Approach (2)) and simulated gas permeabilities of MOF/polymer MMMs for (a) CO2, (b) CH4, (c) N2, and (d) H2. As shown in Fig. 5a, ML predictions for CO2 are in a very good agreement with the simulated permeabilities over a wide range (0.001–89,559 Barrer). Correlations in the low to medium permeability regions (0.001–4020 Barrer) were accompanied by very low MAPE values (< 0.5%). In the high-permeability region (10,438–89,783 Barrer), the MAPE increases to 1.1%, while the correlation remains very high (R2 = 0.997). In Fig. 5b, ML-predicted CH4 permeabilities show excellent agreement with the simulated values across all permeability regions. In the high-permeability region (2,718–12,653 Barrer), the average MAPE remains very low as 0.3% and the correlation remains very high (R2 = 0.999). Similarly, for N2 and H2 shown in Fig. 5c, d, in the high-permeability regions (1143–6197 Barrer for N2 and 3440–32,775 Barrer for H2), the MAPE values increase to 0.2% and 0.8%, respectively. Across the high-permeability regions, only 47 out of 104,196, 11 out of 73,251, and 9 out of 97,536 MMMs exhibit errors exceeding 30% for CO2, CH4, and H2, respectively, while no such cases were observed for N2, indicating that such large deviations are confined to a very limited number of cases. Overall, Approach (2) gives permeability predictions with maximum MAPE deviations of 1.1%, 0.3%, 0.2%, and 0.8% from the direct simulations for CO2, CH4, N2, and H2 permeabilities, respectively. Approach (3) Fig. 6 represents the gas-specific ML models developed to directly predict CO2, CH4, N2, and H2 permeabilities of MOF/polymer MMMs. Each model was trained on a large and compositionally diverse MMM dataset constructed by combining simulated MOF permeabilities with experimental polymer permeabilities. The datasets comprised 104,196 MMMs for CO2 (3982 CoRE MOFs and 4701 hMOFs; 12 polymers), 73,251 MMMs for CH4 (3444 CoRE MOFs and 4695 hMOFs; 9 polymers), 78,147 MMMs for N2 (3982 CoRE MOFs and 4,701 hMOFs; 9 polymers), and 97,536 MMMs for H2 (3444 CoRE MOFs and 4684 hMOFs; 12 polymers). Owing to the very wide permeability ranges arising primarily from the diversity of polymer permeabilities, all models were trained on a logarithmic scale. Very good agreement was observed between the ML-predicted and simulated logPMMM,i values for all gases, with very low RMSE and MAE values. Feature importance analysis revealed a strong and consistent dominance of polymer-related features across all gases. Polymer density (ρpolymer) emerges as the most influential feature, followed by fractional free volume (FFV), underscoring that polymer packing efficiency and available free volume primarily control the gas transport in MMMs. This is expected as the polymer forms the continuous phase in the Maxwell equation. Machine learning prediction of gas permeabilities in MOF/polymer MMMs using Approach (3). Comparison of ML with simulation data to predict permeabilities of (a) CO2, (b) CH4, (c) N2, and (d) H2 in MOF/polymer MMMs. The bar figures represent the importance of the three most influential features used in the ML models. Figure 7 represents the performance of Approach (3), in which ML predictions trained on logarithmic permeability values were transformed back to the absolute scale to enable a direct and consistent comparison with the other approaches. Across all gases, in the low to medium permeability regions, the models yield R2 values of 1 and extremely low MAPE values, as summarized by the region-based error metrics given in Table S3. In Fig. 7a, excellent agreement was observed in the low to medium permeability regions of CO2 (0.001–4020 Barrer) dominated by conventional polymers such as liquid polyesters, Matrimid, CA, and PI-5. In the high-permeability region governed by ultra-high permeable polymers (10,438–89,783 Barrer), the model maintains a high R2 value, while the MAPE increases markedly to 2.5%, suggesting that prediction errors become more pronounced in this region for Approach (3). A closer inspection of the individual data points indicates that the largest deviations are observed for CO2 in this region, where 25 MMMs exhibit errors exceeding 30%, with the maximum deviation reaching 55% for a MOF/PIM-1 MMM. As shown in Fig. 7b–d for CH4, N2, and H2, near-unity R2 values were obtained across all permeability regimes and MAPE values were low, increasing only in the high-permeability region. The corresponding maximum deviations are 26% for CH4, 22% for N2, and 32% for H2. Overall, Approach (3) was shown to yield permeability predictions with maximum MAPE deviations of 2.5%, 0.5%, 0.5%, and 0.7% from the direct simulations for CO2, CH4, N2, and H2, respectively, with higher deviations compared to Approaches (1) and (2). ML-predicted and simulated gas permeabilities of MOF/polymer MMMs using Approach (3). Comparison of ML-predicted (Approach (3)) and simulated MOF/polymer MMM permeabilities for (a) CO2, (b) CH4, (c) N2, and (d) H2. Overall, the three approaches perform similarly in the low-to-medium permeability regime, while clear differences emerge at high permeabilities. While Approaches (1) and (2) maintain comparable accuracy for all gases, Approach (3) exhibits systematically higher deviations from the direct simulations in the high-permeability regime. Since MOF/polymer MMMs have been evaluated based on the trade-off between gas permeability and selectivity, we further assessed the selectivity predictions obtained from the three ML approaches. Figure S3 compares the ML-predicted and simulated selectivities of MOF/polymer MMMs for three industrially important gas separations, CO2/CH4, CO2/N2, and H2/CO2. For each gas pair, selectivity predictions were evaluated over the corresponding set of MOF/polymer MMMs shared by the two gases, resulting in datasets of 73,251 MMMs for CO2/CH4, 78,147 MMMs for CO2/N2, and 97,536 MMMs for H2/CO2 separations. Simulated CO2/CH4 selectivities were between 1.9 and 60. Approaches (1) and (2) successfully reproduced this range with predicted selectivities of 2–60, Approach (3) slightly overestimated the selectivities up to 63 as shown in Fig. S3a–c. In Fig. S3d–f, the simulated selectivities between 4.6 and 46 for CO2/N2 separation were well predicted by Approaches (1) and (2), yielding selectivities of 4.7–46 and 4.9–46, respectively, while Approach (3) again slightly overestimated values, reaching 47. For H2/CO2 separation shown in Fig. S3g–i, the simulated selectivities ranging from 0.2 to 101 were in good agreement with the predictions of all three approaches. Figure 8 compares the ML-predicted gas permeabilities of various MOF/polymer MMMs obtained from the three different ML approaches we introduced with the experimentally reported values available in the literature25–35. This comparison allows a direct assessment of the predictive accuracy of the proposed ML approaches across different MMM compositions and gas species. Figure 8a shows that Approaches (1) and (2) consistently provide closer agreement with experimentally reported permeabilities whereas Approach (3) has larger deviations. This trend is particularly evident for highly permeable MMMs such as Mg-MOF-74/PIM-1 and UiO-66/PIM-1 MMMs whereas Approach (3) underestimates gas permeabilities. We note that the comparison in Fig. 8a was restricted to MOF/polymer MMMs that were present in the training set of ML models for Approach (3), since this approach was trained using data generated at a fixed filler volume fraction of 20 wt.%. Nevertheless, the lower accuracy observed for Approach (3) indicates its limited reliability. Validation of three ML approaches against simulated and experimentally reported gas permeabilities of MOF/polymer MMMs. Comparison of Approach (1), (2), and (3) predictions, simulation results, and experimentally reported gas permeabilities of various types of MOF/polymer MMMs. a Comparison of three ML approaches with experimentally reported permeabilities for MMMs at 20 wt.% filler loading. b, c Comparison of Approach (1) and (2) with experimentally reported permeabilities for MMMs with different polymers and MOF loadings. The experimental data were collected from the following refs. 25–35. Figure 8b, c further compare Approaches (1) and (2) with experimentally reported permeabilities for MMMs having different polymers and MOF loadings. The MMMs considered were composed of representative MOFs (Mg-MOF-74, CuBTC, MOF-5, MIL-53, MIL-53(ht), and UiO-66) embedded in various polymers (PIM-1, PDMS, PI, Matrimid, PSF, 6FDA-ODA, and PU) over a range of filler loadings. All MOFs included in this comparison were part of the training sets of ML models in Approaches (1) and (2), enabling a consistent assessment of their predictive performance. Both approaches exhibit comparable agreement with experimental data. At this point, we note that a key distinction between two approaches lies in their modeling complexity. Approach (1) relies on separately trained ML models to predict gas uptakes and diffusivities of MOFs, where uptake and diffusivity models employ 13 and nine input features, respectively. In contrast, Approach (2) employs a single ML model to directly predict gas-specific MOF permeabilities using only eight input features, a simpler and more efficient workflow. Therefore, we provided additional validation of Approach (2) against experimentally reported MMM permeabilities in Fig. S4, extending its validity to additional 17 different MMMs. After demonstrating the strong predictive power of Approach (2), we used it to assess CO2/CH4, CO2/N2, and H2/CO2 separation potentials of 73,251, 78,147, and 97,536 MOF/polymer MMMs in Fig. 9. For CO2/CH4 separation shown in Fig. 9(a), 15,729 MMMs exceed the 2008 upper bound, while none surpass the 2019 upper bound. MMMs having low-permeability polymers such as Matrimid, CA, TADATO/DSDA-DDBT, PI-5, PDPA-3f show enhanced CO2 permeability upon the MOF addition, but MOF fillers do not influence the selectivity. On the other hand, MOF identity significantly affects the performance of the MMMs having highly permeable polymers (PIM-1, PIM-TMN-Trip, PTMGP, and PTMSP). The addition of MOFs to these polymers can increase, decrease, or have a negligible effect on selectivity depending on the MOF type. For example, for PIM-1, 4,694 types of hypothetical MOFs and 3,379 types of synthesized MOFs were found to yield MMMs that can exceed the Robeson upper bound. Although MOF incorporation improves the CO2 permeability, selectivity improvement is not observed. For instance, PIM-1 exhibits a CO2/CH4 selectivity of 18 in its pristine form, which varies between 12 and 18 after MOF incorporation, indicating that upper-bound exceeding behavior is mainly driven by the increase in permeability. Gas separation performance of MOF/polymer MMMs evaluated using Approach (2). Gas separation performance of MOF/polymer MMMs for (a) CO2/CH4, (b) CO2/N2, and (c) H2/CO2 separations. In Fig. 9b, 26,846 MMMs exceed the 2008 upper bound and 481 MMMs further surpass the 2019 upper bound for CO2/N2 separation. MMMs consisting of low-permeable polymers mainly show permeability improvements with limited changes in selectivity. MMMs incorporating highly permeable polymers more frequently exceed both upper bounds. For CO2/N2 separation, PTMSP-based MMMs show an increase in selectivity from 11 to 13, while PTMGP-based MMMs exhibit an increase from 14 to 16, and both exceed the upper bound together with enhanced permeability. For PTMSP and PTMGP, 5,124 types of hypothetical MOFs and 4,945 types of synthesized CoRE MOFs were found to yield MMMs that can exceed the upper bound. Overall, this trend shows that MOF fillers most effectively complement highly permeable polymer matrices for CO2/N2 separation. For H2/CO2 separation shown in Fig. 9c, 8533 MMMs exceed the 2008 upper bound. Most MMMs show increased H2 permeability with limited changes in selectivity, particularly for low and medium-permeable polymers. In contrast, MMMs incorporating highly permeable polymers (PIM-1, PIM-TMN-Trip, PTMSP, and PTMGP) show increases in both permeability and selectivity, with selectivity reductions observed only in a limited number of MMMs. To better examine the effect of MOF identity on the MMM performance, we specifically focused on PTMSP polymer. Figure 10 represents the relation between pore sizes of MOFs and the selectivity and permeability performance of 8139, 8683, and 8128 MOF/PTMSP MMMs for CO2/CH4, CO2/N2, and H2/CO2 separations, respectively. Figure 10a shows that the incorporation of MOFs with large pores (> 10 Å) into polymers results in increased permeability (up to 1.72 times that of the pristine polymer), but at the cost of decreased selectivity. Figure 10b indicates that when the pore size of the MOF is very small (below 4 Å), a subset of MMMs exhibit enhancements mostly in selectivity, while permeability also increases moderately (up to 1.5 times), enabling them to surpass the upper bound. Another group of MOFs, with pore sizes between 4 and 54 Å, leads to MMMs whose selectivity remains close to, but slightly lower than, that of the pristine polymer, while permeability shows a significant increase. Thanks to this improvement in permeability, these MMMs were still able to exceed the upper bound. In Fig. 10c, we observe that the incorporation of MOFs increases the permeability of all polymers and selectivity of a large fraction of them is also improved, in some cases by up to a factor of two. However, these improvements are not sufficient to surpass the upper bound, since the pristine polymer is already located well below the bound. We also examined the effects of other structural features of MOFs on the performance of MOF/PTMSP MMMs in Fig. S5. Clear differences were observed in the pore size-related descriptors. MOFs in upper bound exceeding MMMs predominantly have smaller LCD values, mostly in the 5–15 Å range. Similarly, ϕ distributions indicate that approximately 70% of the MOFs in these MMMs fall within a moderate porosity range, between 0.6 and 0.8. Consistent with these trends, 93% of the MOFs in such MMMs have PV values below 2 cm3/g. In contrast, the metal% and halogen% exhibit broad distributions, suggesting that elemental composition is less influential than pore size and porosity. MOF structural features governing the separation performance of MOF/PTMSP MMMs. Structure-performance relationships of MOF/PTMSP MMMs for (a) CO2/CH4, (b) CO2/N2, and (c) H2/CO2 separations. Stars indicate MMMs exceeding the upper bound. While the results demonstrate the predictive capability of all three ML approaches, it is also necessary to discuss the main limitations of our methodology. In this study, the ML models for all three approaches were developed using the gas uptake and diffusivity data of MOFs obtained from GCMC and MD simulations. Therefore, the predictive accuracy of our ML models is inherently linked to the assumptions adopted in these simulations. As commonly used in large-scale screening studies, MOFs were modeled as rigid and defect-free crystalline structures1. Such representations may lead to higher permeability results compared to experimentally reported permeabilities, where crystal defects, partial pore blockage, interfacial voids, and non-ideal polymer-filler interactions may be present9. These deviations were expected to be more pronounced at low MOF loadings, where gas transport is increasingly governed by the polymer matrix rather than the filler phase, and the assumptions of ideal filler dispersion and interfacial contact become less valid36. Using classical force fields and approximate charge assignment methods in molecular simulations are necessary for large-scale screening of MOFs, and they also introduce deviations from experimental results. Here, it is important to note that experimentally reported MMM data are inherently heterogeneous since they originate from different studies employing different membrane preparation protocols (differences in synthesis, defects, filler dispersion etc.) and measurement conditions (set-up used, etc.). Therefore, comparison with experiments should be interpreted in terms of overall trends and performance ranges, rather than as exact agreement with individual data points. Despite these assumptions, the good agreement observed between ML predictions, molecular simulations, and available experimental MMM permeability data that we discussed above indicates that the ML models can reliably reproduce the main trends of MMM permeability. Conclusions We introduced a data-driven framework that integrates molecular simulations and ML to study over 100,000 distinct types of MOF/polymer MMMs derived from thousands of synthesized and hypothetical MOFs. Three ML approaches with different target data were introduced, compared, and validated against simulation results and experimentally reported MMM permeabilities for CO2, CH4, N2, and H2 gases. Among these approaches, direct prediction of MOF permeability followed by integration with the Maxwell model was identified as a reliable approach giving the closest results to both experimental and simulated data. Using this approach, we examined MOF/polymer MMMs for CO2/CH4, CO2/N2, and H2/CO2 separations and showed that MOF incorporation mainly increases gas permeability of polymers. For example, for CO2/CH4 separation, the CO2 permeability increases from 52,800 Barrer (for pristine PIM-TMN-Trip) to a maximum of 89,559 Barrer upon MOF incorporation, while selectivity remains nearly unchanged, allowing the MMMs to surpass the upper bound. For CO2/N2 separation, the CO2 permeability increases from 2300 Barrer (for pristine PIM-1) to a maximum of 4019 Barrer upon MOF incorporation, with again minor changes in selectivity. For H2/CO2 separation, MOF incorporation improves both gas permeability and selectivity: H2 permeability increases from 18,800 Barrer for the pristine PIM-TMN-Trip to 32,322 Barrer, and selectivities increase up to 2.1 times. MOF/polymers MMMs exceeding the upper bounds were further examined to identify structure-performance relationships, and a focused analysis on PTMSP-based MMMs showed that pore size of MOFs is a key descriptor governing permeability and selectivity trends. Overall, this study establishes a practical and accessible framework for the quantitative screening of MOF/polymer combinations prior to experimental synthesis. By eliminating the need for computationally intensive molecular simulations, our publicly available ML models enable predictive, performance-guided design of next-generation MMMs by researchers from diverse backgrounds. Methods MOFs The dataset constructed for this study integrates two distinct classes of porous materials: the experimentally synthesized MOF database (CoRE MOF 2019)37 and a hypothetical MOF database (hMOF)38. We utilized the “computation-ready” subset of the recently updated CoRE MOF database39. The geometric properties of all MOFs, specifically the pore limiting diameter (PLD), the largest cavity diameter (LCD), accessible surface area (ASA), pore volume (PV), density (ρ), and porosity (ϕ), were computed using the Zeo + + package40 (version 0.3). Geometric screening was applied to eliminate materials with pore apertures too narrow to accommodate the target gas molecules. We established specific PLD cut-offs based on the kinetic diameters of the adsorbates: MOFs were retained only if they exhibited PLD > 3.64 Å for CO2 (3.30 Å) and N2 (3.64 Å), and PLD > 3.80 Å for CH4 (3.80 Å) and H2 (2.96 Å) molecules. Following this filtration, partial charges were assigned to the framework atoms via the charge equilibration (Qeq) method41 to accurately model electrostatic interactions, particularly for CO2 and N2. To maintain dataset quality, structures yielding physically unrealistic charges (exceeding ±5e) or exhibiting negligible uptakes at 1 bar (< 1 molecule/unit cell) were discarded. The final screened materials comprised 8683 MOFs for CO2, 8139 for CH4, 8683 for N2, and 8128 for H2. Polymers We examined 12 polymers that have been widely studied to make MMMs in literature. These polymers were selected to ensure a representative set across low, medium, and high permeability classes, enabling the ML models to capture structure-property relations across diverse polymer chemistries: Polymer of Intrinsic Microporosity-1 (PIM-1), Polymer of Intrinsic Microporosity based on triptycene units with tetramethyl nitrile functional groups (PIM-TMN-Trip), poly(trimethylsilylpropyne) (PTMSP), poly(trimethylgermylpropyne) (PTMGP), BTDA-DAPI polyimide (Matrimid 5218), 6FDA-terphenylene polyimide derived from diamine 5 (PI-5), a copolyimide derived from TADATO and DSDA dianhydrides and DDBT diamine (TADATO/DSDA-DDBT), cellulose acetate (CA), emeraldine salt form of poly(aniline) (redoped polyaniline), thermotropic liquid-crystalline copolyester (HBA/HNA-30/70), polybenzimidazole cross-linked by oxalic acid (PBI-OA), and poly(diphenylacetylene) bearing fluorine substituents and hydroxyl groups (PDPA-3f). The physical properties of these polymers such as density (ρpolymer) and fractional free volume (FFV) in addition to their gas permeabilities were collected from the literature and reported in Tables S4, S5 in the Supporting Information (SI). Adsorption simulations Molecular simulations were conducted using the RASPA software42 (version 2.0.5) to determine the adsorption and diffusion characteristics of the gases in MOFs. All simulations were performed at 1 bar, 298 K. The molecular models treated H2 and CH4 as single-site spherical molecules43,44, whereas CO2 was represented as a rigid, linear three-site model45 and N2 as a three-site model featuring a dummy center of mass site46. Framework-gas dispersive interactions were defined by the Lennard-Jones (LJ) 12-6 potential, utilizing parameters from the Universal Force Field (UFF)47. Cross-interaction parameters were derived via Lorentz-Berthelot mixing rules. Electrostatic interactions were computed using the Coulomb potential with Ewald summation utilized for long-range corrections48. A cut-off radius of 13 Å was used for van der Waals interactions, and simulation cells were expanded to ensure a minimum dimension of 26 Å in all directions under periodic boundary conditions. Gas uptakes of MOFs at 1 bar, 298 K, were obtained from the GCMC simulations consisting of 104 initialization cycles and 2 × 104 production cycles. For the calculations related to infinite dilution conditions, the Henry’s coefficients (K0,i) of CO2, CH4, N2 and H2 gases as well as the isosteric heat of adsorption (Qst) at infinite dilution, were computed at 298 K using the Widom particle insertion method49. Diffusion Simulations Gas transport properties of MOFs were evaluated using MD simulations in the NVT ensemble at 298 K, controlled by a Nose-Hoover thermostat50. These simulations were performed to compute the infinite dilution diffusivity (D0,i) using 30 molecules to exclude gas-gas interactions, and the self-diffusivity (Di) using the gas loadings obtained from GCMC simulations performed at 1 bar. The MD simulations involved a 5 ns production run (5 × 106), following 103 initialization, and 104 equilibration cycles, with a time step of 1 fs. Diffusion coefficients were extracted from the linear regime of the Mean Squared Displacement (MSD) profiles using the Einstein relation51. Trajectories were excluded if their linear-fit R2 was below 0.9 or if diffusivities differed by more than three orders of magnitude across directions. MOFs with gas diffusivities below the reliable detection limit of MD (10−8 cm2/s) were excluded. All frameworks were treated as rigid structures to save significant computational time. Permeability calculations The intrinsic gas permeability of the MOFs (PMOF,i) was calculated using the simulated gas adsorption and diffusivity data utilizing where Ni, Di, and fi represent the adsorbed loading of gas i at 1 bar (obtained from GCMC simulations as described above), its corresponding self-diffusivity (obtained from MD simulations as described above), and the feed-side pressure of gas i (1 bar in this work), respectively. We note that this approach based on molecular simulations has been shown to accurately reproduce experimentally measured gas permeabilities of CO2, CH4, N2, and H2 in a variety of pure MOF membranes including Cu-BTC, Ni-MOF-74, MOF-5, MIL-53(Al), ZIF-8, ZIF-90, and ZIF-95 in our previous studies11,52–54. We also computed infinite-dilution gas permeability using where D0,i denotes the infinite dilution diffusivity computed from the MD simulations and K0,i is the Henry’s coefficient of gases derived from the MC simulations. The Maxwell model55 was employed to predict the gas permeability of the MOF/polymer MMMs (PMMM,i). This model has a proven accuracy in reproducing experimental data for MOF/polymer MMMs and it has been very widely used in the literature56. The gas permeability of the MMM was calculated as where ϕ is the filler volume fraction, 0.2 for our study, while PMMM,i, PP,i, and PMOF,i correspond to the permeability of the MMM, the permeability of the pure polymer collected from the experimental literature, and the permeability of the MOF computed as described above, respectively. Finally, the ideal selectivity of the MMMs (SMMM,i/j) was determined as the ratio of the permeabilities of the corresponding gas pairs: With this approach, we computed the gas permeabilities of 104,196 MOF/polymer MMMs for CO2, 73,251 MMMs for CH4, 78,147 MMMs for N2, and 97,536 MMMs for H2 generating a large and diverse dataset. Machine learning We examined three different approaches in our work to predict CO2, CH4, N2, and H2 permeabilities of 104,196 different types of MOF/polymer MMMs using ML models: Approach (1) CO2, CH4, N2, and H2 adsorption (Ni) and diffusivity (Di) data of MOFs were predicted using gas-specific ML models. This approach consisted of two independent ML models for each gas. Four gas-specific ML models were trained to predict Ni and four gas-specific ML models were trained to predict logDi. The target data of the adsorption models were Ni values (obtained from GCMC simulations), and 13 input features consisted of energetic, structural, and compositional descriptors were used: Qst, PLD, LCD, ϕ, PV, ASA, ρ, the elemental compositions (C%, O%, H%, metal%, halogen%, N%). The target data of the diffusion models were logDi values (obtained from MD simulations), and nine input features were used: logD0,i, Ni, Qst, LCD/PLD, PLD, LCD, ϕ, PV, ASA. ML-predicted uptake and diffusivity values of MOFs were then used to calculate PMOF,i values using Eq. (1), which were subsequently inserted into the Maxwell model to compute PMMM,i using Eq. (3). Approach (2) CO2, CH4, N2, and H2 permeabilities of MOFs were predicted by using four gas-specific ML models. The target data of permeability models were logPMOF,i (computed using Eq. (1)) whereas eight input features were used: Qst, PLD, LCD, ϕ, PV, ASA, ρ, and logPMOF,i0 (computed using Eq. (2)). ML-predicted PMOF,i values were subsequently inserted into the Maxwell model to compute PMMM,i using Eq. (3). Approach (3) CO2, CH4, N2, and H2 permeabilities of MOF/polymer MMMs were directly predicted by using four gas-specific ML models. The target data of these MMM models were logPMMM,i and eight input features were used: ρpolymer, FFV, and six MOF descriptors; Qst, PLD, LCD, ϕ, PV, and ASA. All ML models were developed using the Tree-Based Pipeline Optimization Tool (TPOT)57 (version 0.12.2), which automatically selected and optimized regression algorithms implemented in the scikit-learn package58. To maintain consistent feature distributions between the training and test sets, the data was divided using stratified sampling into an 80% training set and a 20% test set. Five-fold cross-validation was used during TPOT optimization to reduce risk of overfitting. The predictive performance of the ML models was evaluated using the coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and standard deviation (SD) as summarized in Table S6. Final ML models and their optimized hyperparameters are given in Tables S7, S8, S9, S10. Supplementary information Supplementary File Acknowledgements 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. Author contributions S.K. conceptualized the study, F.N.Y. performed computations and analysis. Both authors contributed to the discussion and manuscript drafting/editing. Peer review Peer review information Communications Materials thanks the anonymous reviewers for their contribution to the peer review of this work. Data availability The machine learning models developed for all three approaches, along with the corresponding input CSV files used for model training and prediction, are publicly available on GitHub at https://github.com/feridenevayungul/ML_MOF_polymer_MMMs.git. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s43246-026-01207-9. References Associated Data Supplementary Materials Supplementary File Data Availability Statement The machine learning models developed for all three approaches, along with the corresponding input CSV files used for model training and prediction, are publicly available on GitHub at https://github.com/feridenevayungul/ML_MOF_polymer_MMMs.git.