Molecular diffusion enhanced performance evaluation of metal-organic frameworks for CO2 capture Authors: Rodrigo Neumann Barros Ferreira, Yogesh V. Joshi, Felipe Lopes Oliveira, Anantha Sundaram, Ashish B. Mhadeshwar, Jayashree Kalyanaraman, Binquan Luan, Mathias B. Steiner Published: 2026 ABSTRACT Molecular diffusion is a fundamental property that limits the performance of solid sorbents in carbon dioxide capture and separation applications. Unique to each sorbent, gas diffusion is determined by the physical and chemical interactions that occur between the gas molecules and a sorbent’s surface atoms. At the process level where carbon dioxide capture performance is validated, however, simulations are typically carried out using generalized parameters that omit the structure-specific, molecular kinetics occurring in each sorbent. Here, we report process-scale simulations of carbon dioxide capture performance in metal-organic frameworks (MOF) informed by molecular adsorption predictions that represent the unique structural properties of each MOF. By evaluating a total of 10,143 MOFs for post-combustion carbon dioxide capture, we demonstrate how the inclusion of the material-specific, molecular diffusion dynamics alters their simulated, process-level performance. The method could be applied to evaluate a broader class of solid sorbents, including covalent organic frameworks and zeolites. Introduction Point-source capture of carbon dioxide (CO2) is a key feature of net-zero emission strategies1–3. Solid sorbents, such as, for example, activated carbons, zeolites, and covalent organic frameworks have been proposed as potentially high-performing CO2 capture materials for industrial applications4. Among the candidate materials, metal-organic frameworks (MOF) stand out due to their large chemical variability enabled by their building blocks, i.e., organic linkers and metal clusters5. While there exist trillions of MOF descriptions6, merely thousands of those structures have been realized and validated in the lab so far7. Therefore, analyzing and rank-ordering all known MOF structures with regards to their application performance remains a major challenge, even by using advanced computational methods8,9. Computational screening workflows typically include calculations of geometric and adsorption properties of the specific MOF candidates to be evaluated, however, the predictions are often limited to the molecular scale10–14. Wilmer et al.15 performed a large-scale evaluation of MOFs for CH4 capture based solely on molecular-level adsorption uptake predictions. Similarly, Lin et al.16 derived a metric from molecular-level data for assessing and rank zeolites and zeolitic imidazolate frameworks (ZIF) for CO2 capture. While indicative of the potential at process scale, individual molecular-level metrics have yet failed to predict process-level material performance overall17. Nevertheless, the application-specific gas separation performance of any given material must ultimately be assessed at the process level, by taking into consideration realistic figures-of-merits for purity, productivity, recovery and energy consumption18. It is, therefore, necessary to devise a multi-scale screening approach for integrating the molecular-to-process scale performance evaluation19. Recent studies have provided molecular-level simulation data as inputs to macro-scale process models, however, the approach poses challenges20. Specifically, the CO2 uptake of a MOF structure, simulated at molecular scale either by means of Grand Canonical Monte Carlo (GCMC) or Ideal Adsorbed Solution Theory (IAST), is used as input for pressure- or vacuum-swing adsorption (PSA/VSA) models at process scale that comply with the 95% CO2-purity and 90% CO2-recovery targets, respectively, issued by the the U.S. Department of Energy21–26. Burns et al.24 integrated atomistic simulations with a VSA process simulator while Farmahini et al.25 studied the effect of mesoscale materials properties, such as size and porosity of the material constituents, on productivity and energy efficiency of the VSA process. These approaches, however, were based on the assumption that gas adsorption in MOFs does not depend on structure-dependent differences that occur at microscopic scales. Conceptually, this was reflected by assuming a constant mass transfer coefficient for all MOF candidates under investigation. An enhanced performance assessment at process level would require an effective mass transfer coefficient Keff that captures both structure-independent, macro-pore diffusivity, as well as structure-dependent, micro-pore diffusivity27. Indeed, several investigations28,29 have already revealed a dependence of diffusivity on micro-pore, molecular loading. As gas diffusion and adsorption are, ultimately, determined by a MOF’s structural properties, we suggest including micro-pore gas diffusivity, provided for each MOF structure through MD simulations, into the process simulation of CO2 capture performance by means of Keff, as conceptually visualized in Fig. 1.Fig. 1Impact of molecular diffusion in MOF on process performance in post-combustion CO2 capture.Four representative MOF structures are shown on the left. The inclusion of their material-specific, micro-pore diffusivity Dμ in the effective mass transfer coefficient Keff, see Equation (2), leads to significant changes in CO2 capture outcomes at process level: Recovery (higher is preferable), Productivity (higher is preferable) and Energy Requirement (lower is preferable). Impact of molecular diffusion in MOF on process performance in post-combustion CO2 capture. Four representative MOF structures are shown on the left. The inclusion of their material-specific, micro-pore diffusivity Dμ in the effective mass transfer coefficient Keff, see Equation (2), leads to significant changes in CO2 capture outcomes at process level: Recovery (higher is preferable), Productivity (higher is preferable) and Energy Requirement (lower is preferable). In the following, we introduce a multi-scale, computational workflow to evaluate MOFs for CO2 post-combustion capture with temperature-swing absorption (TSA) in natural gas combined cycle (NGCC) power plants. The workflow, see Fig. 2a, uses molecular-level simulations for characterizing MOFs with high CO2 adsorption selectivity, which are shortlisted based on active learning cycles, and rank-ordered based on process simulations informed by molecular-scale, adsorption figures-of-merit. Finally, process optimizations at realistic operation conditions confirm each MOF’s performance at process level. We demonstrate that inclusion of molecular kinetics improves the workflow’s ability to discover false leads, which reduces computational costs significantly. Moreover, based on the top-20 MOF ranking results, we show for representative crystallite diameters how the inclusion of microscopic diffusivity affects process-level, CO2 capture performance predictions.Fig. 2Active learning model for MOF selection.a Model architecture with iterative material selection workflow that targets CO2/N2 selectivity and CO2 capacity. b Evolution of 5 successive model iterations and the associated MOF distributions in terms of CO2 and N2 adsorption free energy, respectively. The highest-ranked material in each iteration exhibits a significant increase of CO2/N2 selectivity with regards to prior iterations. c Pore-limiting diameter distribution of selected MOFs for each of the 5 active-learning iterations. Source data are provided as a Source Data file. Active learning model for MOF selection. a Model architecture with iterative material selection workflow that targets CO2/N2 selectivity and CO2 capacity. b Evolution of 5 successive model iterations and the associated MOF distributions in terms of CO2 and N2 adsorption free energy, respectively. The highest-ranked material in each iteration exhibits a significant increase of CO2/N2 selectivity with regards to prior iterations. c Pore-limiting diameter distribution of selected MOFs for each of the 5 active-learning iterations. Source data are provided as a Source Data file. Results In the following sections, we report the results obtained by applying the computational workflow that integrates structure selection by active-learning, molecular simulation and process-level optimization. As reference database for our study, we have selected the non-disordered, all solvent removed (ASR) portion of CoRE-MOF-2019 (v1.1.2)30,31. From the initial dataset of 10,143 structures, we have removed 3659 spurious MOF structures using mofchecker32 as well as an element-specific, charge-based heuristics. Within the remaining 6484 MOFs, we have identified those structures providing the highest productivity associated with the lowest energy requirement, as defined in Supplementary Note 4, while achieving a purity of at least 95% as required for CO2 capture applications at NGCC power plants. In the following sections, we detail how this evaluation was performed. Active learning-based structure selection We have built an active-learning model for predicting CO2 and N2 adsorption free energy. The training data was extracted by fitting model functions to the GCMC-simulated adsorption isotherms, see Methods section. The active learning workflow uses extensive featurization combining multiple tools. In short, feature selection is performed by combining Lasso regression and variable importance discerned from initial machine-learning (ML) models. Multiple ML models are built using the selected features and free energies of adsorption, see Fig. 2a. The best-performing model with the lowest uncertainty is used for Pareto-searching higher-performing materials for subsequent iterations. The material selection improves CO2 vs. N2 selectivity as well as CO2 capacity, and selects materials with lower SCScore33 as well as lower prediction uncertainty. Given that selectivity has been shown to be a poor descriptor of separation performance9, we have assigned higher priority to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G}_{{{{\rm{CO}}}}_{2}}$$\end{document}ΔGCO2 as it represents both selectivity and capacity of 4% CO2 NGCC flue gas in the Henry regime. Further details are provided in the Methods section. Overall, 982 materials were subjected to GCMC simulations over 5 active learning iterations. The progression of material selectivity is shown in Fig. 2b. The best performing material in each iteration is depicted by solid symbols, showing stepwise selectivity improvements. As a result of the active learning process that targets CO2 selectivity, the distribution successively shifts towards structures with narrower pores, see Fig. 2c. Most materials in iterations 4 and 5, respectively, are MOFs having a pore-limiting diameter close to 2.5 Å. We note that for structures with pore dimensions similar to the size of the gas molecules, mass transfer is controlled by micro-pore diffusion34. In the following, we will complement the analysis of the 982 MOF structures by simulating molecular diffusion of CO2 and N2, respectively, for each individual MOF structure within the refined data set. The diffusion data are then combined with the adsorption data for enabling the process-level CO2 capture evaluation of each MOF. Molecular- to process-level evaluation For all selected MOF structures, we have simulated CO2 and N2 molecular diffusion and adsorption uptake at representative pressure and temperature levels. The self-diffusion coefficient was calculated based on mean squared molecular displacement. The heat of adsorption was extracted from the adsorption isotherms by taking the dominant site of the best fit multi-site Langmuir model which is representative of adsorption affinity. In Fig. 3a, we plot the heat of adsorption for CO2 and N2, colored by the CO2 diffusion coefficient at T = 30 °C. We focus our analysis on the lower right region with regards to the parity line which contains promising capture materials candidates due to their comparatively higher CO2 adsorption affinity. Among those structures, MOFs with a low CO2 heat of adsorption, between 10 and 30 kJ/mol, tend to diffuse faster, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D}_{\mu }^{{{{\rm{CO}}}}_{2}}\gtrsim 1{0}^{-9}\,{{{\rm{m}}}}^{2}/{{\rm{s}}}$$\end{document}DμCO2≳10−9m2/s. With increasing CO2 heat of adsorption, the corresponding diffusion coefficient tends to decrease several orders of magnitude, due to the stronger binding affinity and/or smaller pore size. We note that the N2 diffusion coefficient is usually 2–5 times larger, as shown in Supplementary Fig. 8.Fig. 3Effect of molecular diffusion on CO2 capture productivity.a Distribution of heat of adsorption for CO2 and N2, respectively, for a set of select MOF structures. For each structure, the diffusion coefficient is indicated by color. Note, that MOFs with a high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {H}_{C{O}_{2}}/\Delta {H}_{{N}_{2}}$$\end{document}ΔHCO2/ΔHN2 ratio can have negligible microscopic CO2 diffusion. b Effective mass transfer coefficient Keff as function of the microscopic diffusion coefficient Dμ for three representative crystallite diameters Rμ. Below a diameter-dependent diffusivity threshold, mass transfer is limited by microscopic diffusion. c CO2 capture productivity simulated at process-level with a material-independent, constant Keff (left) as well as with a material-dependent Keff that accounts for the microscopic diffusion coefficient of each MOF (right). Open symbols represent materials for which the process-level simulation did not yield a CO2 purity of at least 95%. Source data are provided as a Source Data file. Effect of molecular diffusion on CO2 capture productivity. a Distribution of heat of adsorption for CO2 and N2, respectively, for a set of select MOF structures. For each structure, the diffusion coefficient is indicated by color. Note, that MOFs with a high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {H}_{C{O}_{2}}/\Delta {H}_{{N}_{2}}$$\end{document}ΔHCO2/ΔHN2 ratio can have negligible microscopic CO2 diffusion. b Effective mass transfer coefficient Keff as function of the microscopic diffusion coefficient Dμ for three representative crystallite diameters Rμ. Below a diameter-dependent diffusivity threshold, mass transfer is limited by microscopic diffusion. c CO2 capture productivity simulated at process-level with a material-independent, constant Keff (left) as well as with a material-dependent Keff that accounts for the microscopic diffusion coefficient of each MOF (right). Open symbols represent materials for which the process-level simulation did not yield a CO2 purity of at least 95%. Source data are provided as a Source Data file. The predominance of MOFs with high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {H}_{C{O}_{2}}/\Delta {H}_{{N}_{2}}$$\end{document}ΔHCO2/ΔHN2 ratio but low diffusivity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D}_{\mu }^{{{{\rm{CO}}}}_{2}}\lesssim 1{0}^{-11}\,{{{\rm{m}}}}^{2}/{{\rm{s}}}$$\end{document}DμCO2≲10−11m2/s) has important implications. Generally, we expect that a high-performing CO2 capture material should provide conditions under which gas adsorbents efficiently diffuse within the pores and bind strongly to the adsorption sites, even though a trade-off exists. For analyzing the impact of molecular-level diffusivity in each MOF structure at the process level, the numerical diffusivity values, Dμ, are propagated via the effective mass transfer coefficient Keff in Equation (2). The Keff vs. Dμ curves in Fig. 3b exhibit two distinct regimes: micro-porosity limited transfer, and macro-porosity limited transfer. Each value of the crystallite size Rμ is associated with a characteristic threshold below which mass transfer is limited by microscopic diffusion. As a rule-of-thumb, we conclude that the larger the crystallite diameter, the stronger the impact of microscopic diffusion. To illustrate the crystallite size dependence of the results, we have chosen three characteristic diameters: 1, 10 and 100 μm. The 1-μm and 10-μm values characterize the lower and upper bounds, respectively, of typical crystal size distributions taken from commercially available samples. The 100-μm value serves as an extrapolation for confirming the sensitivity observed at the 10 μm-level. Nevertheless, from a volume fraction perspective, larger crystallites dominate despite their lower occurrence. We now compare the process-level CO2 capture productivity for two scenarios: a constant, material-independent Keff, equivalent to setting Rμ = 0 in Equation (2), thus neglecting microscopic diffusivity limitations, as well as a structure-dependent Keff accounting for microscopic diffusivity. In both cases, we have assumed a crystallite diameter of Rμ = 10 μm and a CO2 purity of 95%. The resulting productivity values are indicated by color shades in Fig. 3c. For the group of select MOFs, the plot reveals a correlation between CO2 affinity, CO2/N2 selectivity, and capture productivity. Most MOFs in our selection have a low CO2 affinity, as evidenced by a small \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {H}_{C{O}_{2}}-\Delta {H}_{{N}_{2}}$$\end{document}ΔHCO2−ΔHN2 difference. Due to the high concentration of N2 in the feed stream, their separation performance is low. Therefore, we focus our analysis on the right-hand side of the graphs. As a key result of the comparison, 16% of high-productivity MOFs, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gtrsim 1\frac{{{\rm{kg}}}}{{{\rm{day}}}}.{{\rm{L}}}$$\end{document}≳1kgday.L, fail to sustain high productivity once molecular diffusivity is taken into account by means of a structure-dependent Keff. From a computational screening perspective, we consider these structures as “false leads” that can be effectively removed from the performance-based MOF rank-order. Process-level optimization and materials ranking For a given carbon dioxide capture process, the choice of the best capture material, ultimately, depends on the trade-off between productivity and energy requirement. For our select group of MOF structures, we plot in Fig. 4 energy requirement as function of productivity for a representative TSA process. We compare side-by-side the results obtained by assuming a constant and a material-dependent Keff. We note that the distribution of energy requirement spans several orders of magnitude while the productivity distribution is relatively narrow.Fig. 4Effect of molecular diffusion on energy requirement and productivity.a Distribution of energy requirement vs. productivity for a crystallite diameter Rμ = 10 μm, assuming a constant Keff (left) and a material-dependent Keff (right) that accounts for microscopic diffusivity. The respective CO2 diffusion coefficients are indicated by color. Solid squares represent “process-feasible” (PF) materials, i.e., those whose performance achieved 95% purity and 90% recovery, while open squares mark the positions of excluded materials whose performance failed to meet those targets. The dashed rectangle delineates the “economically viable” (EV) region with productivity higher than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\frac{{{\rm{kg}}}}{{{\rm{day}}}.{{\rm{L}}}}$$\end{document}1kgday.L and energy requirement lower than 5 MJ/kg of CO2. b Effect of microscopic diffusivity in Keff on process performance for crystallite size Rμ = 10 μm. Arrows indicate changes undergone by structures for which the inclusion of molecular diffusivity has led to a significant drop in process performance, pushing them out of the EV region. c Percentages of PF and EV materials with respect to the starting set of “promising candidates” as selected by the active-learning model, assuming constant (dashed lines) or material-specific (solid lines) Keff. The percentage of PF materials dropped from 81% (798 out of 982) to 80% (784 out of 982) upon inclusion of microscopic diffusivity. The prevalence of EV materials went from 28% (277 out of 982) to 24% (234 out of 982) of the starting set, or from 35% (277 out of 798) to 30% (234 out of 784) of their respective PF set. This corresponds to a 16% reduction (from 277 to 234) in the original EV population, representing the “false leads''. Source data are provided as a Source Data file. Effect of molecular diffusion on energy requirement and productivity. a Distribution of energy requirement vs. productivity for a crystallite diameter Rμ = 10 μm, assuming a constant Keff (left) and a material-dependent Keff (right) that accounts for microscopic diffusivity. The respective CO2 diffusion coefficients are indicated by color. Solid squares represent “process-feasible” (PF) materials, i.e., those whose performance achieved 95% purity and 90% recovery, while open squares mark the positions of excluded materials whose performance failed to meet those targets. The dashed rectangle delineates the “economically viable” (EV) region with productivity higher than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\frac{{{\rm{kg}}}}{{{\rm{day}}}.{{\rm{L}}}}$$\end{document}1kgday.L and energy requirement lower than 5 MJ/kg of CO2. b Effect of microscopic diffusivity in Keff on process performance for crystallite size Rμ = 10 μm. Arrows indicate changes undergone by structures for which the inclusion of molecular diffusivity has led to a significant drop in process performance, pushing them out of the EV region. c Percentages of PF and EV materials with respect to the starting set of “promising candidates” as selected by the active-learning model, assuming constant (dashed lines) or material-specific (solid lines) Keff. The percentage of PF materials dropped from 81% (798 out of 982) to 80% (784 out of 982) upon inclusion of microscopic diffusivity. The prevalence of EV materials went from 28% (277 out of 982) to 24% (234 out of 982) of the starting set, or from 35% (277 out of 798) to 30% (234 out of 784) of their respective PF set. This corresponds to a 16% reduction (from 277 to 234) in the original EV population, representing the “false leads''. Source data are provided as a Source Data file. We introduce terminology for describing groups of materials at different performance screening levels. We call “promising candidates”, those out of the 6484 initial structures that were preselected by the active learning model to undergo multi-scale simulations. Among those, we term “process-feasible” (PF) those whose simulated process-level performance achieves 95% CO2 purity and 90% CO2 recovery targets. Finally, the subset of PF materials that achieve CO2 productivity higher than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\frac{{{\rm{kg}}}}{{{\rm{day}}}.{{\rm{L}}}}$$\end{document}1kgday.L and energy requirement lower than 5 MJ/kg of CO2 is labeled “economically viable” (EV). In Fig. 4a, left panel, we plot energy requirement as function of productivity assuming a constant, material-independent Keff. In order to guide the reader and allow for comparison, we have assigned colors for indicating the respective CO2 diffusion coefficient of each structure, even though those diffusivity values were not used to calculate Keff in the simulations that originated this plot. We note that many of the materials in the EV region have very small CO2 diffusivity values. Such small diffusivity values could lead to significant changes in CO2 capture performance if accounted for in the process model. For analyzing the change, we have included the microscopic diffusion coefficients in the mass transfer coefficient Keff, see Equation (2), and recalculated energy requirement and productivity for all structures assuming a diffusivity-dependent Keff. The results are shown in the right panel of Fig. 4a. We note that high-performance MOFs that were initially located in the EV region are now excluded by failing to surpass the purity/recovery process feasibility targets. In the plot on the right, their initial positions are marked by open symbols. We find that this trend depends on crystallite size, having a smaller impact for Rμ = 1 μm, see Supplementary Fig. 15, and a larger impact for Rμ = 100 μm, see Supplementary Fig. 16. In Fig. 4b–c, we have identified the structures whose process-level performance was the most affected. We compare their performance coordinates before and after inclusion of molecular diffusivity for Rμ = 10 μm. For reference, we plot open symbols in their original positions taken from the left panel of Fig. 4a. The arrows indicate MOFs that moved out of the EV region once their structure-dependent, molecular diffusivity was included into the calculation of Keff. The percentage of promising candidates that achieved the purity/recovery process feasibility targets was 81% under the constant Keff approximation, but some of them no longer achieved these targets after the inclusion of microscopic diffusivity. The relative reduction in the PF population grows from 1%, to 2%, to 5% with increasing crystallite diameter Rμ = 1, 10 and 100 μm, respectively. We now turn our attention to the subset of EV materials, which correspond to 28% of the promising candidates and 35% of the PF population, under the constant Keff approximation. In this case, the relative reduction of the EV population was 14%, 16% and 20% for Rμ = 1, 10 and 100 μm, respectively. We term these materials as “false leads”, as they would have been wrongly ranked high despite their low performance due to diffusivity constraints. We now discuss how the observed changes affect the CO2 capture performance rank-order. Potentially interesting materials, with high productivity and low energy requirement, may lose their high ranking upon consideration of microscopic diffusivity. Figure 5 depicts this effect whereby low diffusivity materials drop from the upper ranks upon increasing importance of microscopic diffusivity in the overall mass-transfer coefficient. Materials with higher diffusion coefficients are shown with nearly horizontal curves, as the increase in Rμ does not severely affect Keff and hence process based performance. On the other hand, lower diffusivity materials are characterized by drops (rises) in productivity (energy requirement) upon increasing Rμ. These results are in accordance with the behavior of Keff in Fig. 3b.Fig. 5Effect of molecular diffusion on CO2 capture performance ranking.Productivity (left) and energy requirement (right) based on constant Keff, as well as structure-dependent Keff for representative crystallite sizes Rμ = 1, 10 and 100 μm. Each curve represents a specific MOF structure and its color represents its microscopic diffusion coefficient. The higher (lower) the productivity (energy requirement), the higher the performance rank of the candidate material. High diffusion coefficients are displayed with transparency for making productivity drops and energy requirement increases visible in the plots. Data points for structures that fail to surpass the 95%-purity threshold are not shown. Source data are provided as a Source Data file. Effect of molecular diffusion on CO2 capture performance ranking. Productivity (left) and energy requirement (right) based on constant Keff, as well as structure-dependent Keff for representative crystallite sizes Rμ = 1, 10 and 100 μm. Each curve represents a specific MOF structure and its color represents its microscopic diffusion coefficient. The higher (lower) the productivity (energy requirement), the higher the performance rank of the candidate material. High diffusion coefficients are displayed with transparency for making productivity drops and energy requirement increases visible in the plots. Data points for structures that fail to surpass the 95%-purity threshold are not shown. Source data are provided as a Source Data file. As representative examples, we discuss in Supplementary Note 4 the properties of the two high-performing MOFs, MOCHIW and MOCHIW, as well as the two low-performing MOFs, BETGUE and EVOMOR, introduced in Fig. 1. Specifically, we provide for each MOF the adsorption isotherms (Supplementary Fig. 11), diffusion coefficients (Supplementary Fig. 12), breakthrough cycle profiles (Supplementary Fig. 13) and process-level Pareto plots (Supplementary Fig. 14). The plots enable a in-depth analysis of how the inclusion of molecular diffusion affects each step of the multi-scale performance evaluation. Discussion Using a micro- to macro-scale, computational simulation and discovery workflow for MOFs, we have demonstrated how the inclusion of structure-dependent, molecular diffusion dynamics affects post-combustion CO2 capture figures-of-merits, i.e., purity, recovery, productivity, and energy requirement, of a simulated TSA process for application at NGCC power plants. An active-learning, structure selection model has iteratively identified a group of 982 MOFs with high molecular CO2 / N2 selectivity and CO2 capacity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\Delta {G}_{{{{\rm{N}}}}_{2}}-2\times \Delta {G}_{{{{\rm{CO}}}}_{2}})$$\end{document}(ΔGN2−2×ΔGCO2). As the identified structures exhibit narrow pore diameters, of the order of the size of the gas molecules, their CO2 capture performance is limited by microscopic diffusion. For quantifying the effect, we have compared the simulated TSA process performance obtained by using an effective mass transfer coefficient for all structures, as well as by calculating for each of the 982 MOFs in the selection a structure-dependent mass transfer coefficients accounting for molecular-scale diffusion dynamics. While pellet properties of representative MOFs have previously been used as decision variables for optimizing macro-pore mass transfer in vacuum-swing adsorption25, we have focused in our work on microscopic diffusivity in crystallites, mainly due to the pore size distribution of the MOF candidates under study. Overall, the inclusion of molecular diffusion has led to the exclusion of up to 20% of MOF candidates from the initial rank-order. This is because either the structures did not achieve the 95% CO2 purity threshold or because of their performance loss in terms of productivity and energy requirement. In the rank-order obtained by using structure-dependent diffusivity and mass transfer coefficients, 27 MOFs have achieved volumetric productivity larger than 1.5 kg/(day.L) and specific energy consumption lower than 1 MJ/kg. For comparison, reference CO2 solid sorbents such as Zeolite 13X and well-known MOF materials UTSA-16 and Mg-MOF-74, have reported productivity and energy consumption of 1–3 kg/(day.L) and 0.5–1 MJ/kg, respectively, in vacuum-swing adsorption17. More recently, CALF-20 has achieved industrial-scale production with a productivity of 0.53 kg/(day.L) and an energy requirement of 0.72 MJ/kg, respectively, in pressure-vacuum-swing adsorption35,36. In Supplementary Table 10, we provide our top-20 performance list that exemplifies the effect of microscopic diffusivity on rank-order that increases with crystallite diameter. For Rμ = 1, 10 and 100 μm, respectively, a total of 3, 7, and 10 MOFs have dropped off the initial ranking. In case of the best-performing material, BOHGOU, the inclusion of molecular diffusivity in the process simulation has decreased volumetric productivity, from 1.84 to 1.33 kg/(day.L), and increased the energy requirement, from 1.07 to 1.50 MJ/kg of CO2. We note that in the case of Rμ = 100 μm, BOHGOU drops out of the top-20 while the second best material for all other crystallite sizes, MOCHIW, moves to the first position. This highlights the importance of MOF crystallite size on CO2 capture performance. In this work, we have focused on methodologically including microscopic diffusivity in computational material screening. To that end, we had to make reasonable simplifications to other equally challenging dimensions of material performance evaluation. In particular, we have assumed dry NGCC flue gas throughout the results reported in this manuscript. In order to address performance implications related to the presence of water, out of the 982 MOFs under consideration, we have removed those with higher water affinity than CALF-2035 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G}_{{{{\rm{H}}}}_{2}{{\rm{O}}}}\le -8.1\,{{\rm{kJ}}}/{{\rm{mol}}}$$\end{document}ΔGH2O≤−8.1kJ/mol). To further validate of our findings, we performed an additional, structural quality check based on metal oxidation states, by using MOSAEC37. As shown in Fig. 6a, 4 out of the 381 remaining MOFs, referred to as “water-tolerant”, fall within the economically viable region. Note, that these MOFs have CO2, N2 and H2O affinities that are similar to or better than those reported for CALF-20, see Supplementary Note 4.3. We note that one of the structures, namely EXOFIH, was flagged as “bad” by MOSAEC. Interestingly, two of them have one-dimensional square pores similar to CALF-20. We would like to emphasize that, by virtue of including diffusion in the screening workflow, these water-tolerant MOFs would not be kinetically limited. To assess the impact of using EQeq charges, we re-ran the end-to-end simulation workflow for all the water-tolerant, economically-viable MOFs using DDEC06 charges computed with PACMOF38. Their relative rank order with regard to CO2 capture performance remained unchanged. The refined water tolerance screening is provided in Supplementary Note 4.4.Fig. 6Water-tolerant MOF candidates.a Distribution of energy requirement vs. productivity for water-tolerant MOFs, i.e., with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G}_{{{{\rm{H}}}}_{2}{{\rm{O}}}} > -8.1\,{{\rm{kJ}}}/{{\rm{mol}}}$$\end{document}ΔGH2O>−8.1kJ/mol, assuming a crystallite diameter of Rμ = 10 μm. We have identified 4 promising MOF structures in the economically viable region. Empty symbols represent structures flagged as “bad” by MOSAEC37 b Crystal structures of the economically viable, water-tolerant MOF candidates. We marked EXOFIH to represent that it was flagged as “bad” by MOSAEC37. Source data are provided as a Source Data file. Water-tolerant MOF candidates. a Distribution of energy requirement vs. productivity for water-tolerant MOFs, i.e., with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G}_{{{{\rm{H}}}}_{2}{{\rm{O}}}} > -8.1\,{{\rm{kJ}}}/{{\rm{mol}}}$$\end{document}ΔGH2O>−8.1kJ/mol, assuming a crystallite diameter of Rμ = 10 μm. We have identified 4 promising MOF structures in the economically viable region. Empty symbols represent structures flagged as “bad” by MOSAEC37 b Crystal structures of the economically viable, water-tolerant MOF candidates. We marked EXOFIH to represent that it was flagged as “bad” by MOSAEC37. Source data are provided as a Source Data file. We conclude that the materials at the top of the diffusion-enabled MOF ranking are high-performing CO2 sorbents with the potential to challenge, or even surpass, the performance of the best solid sorbents deployed today. Future work should include structural checks with MOFClassifier39 and address stability and synthesizability within the screening protocol. For enabling their validation and exploration, we have made our data publicly available, see Data availability. The computational simulation workflow for molecular diffusivity, which is based on Crystallographic Information File (CIF) format and made available as open-source code, see Code availability, could be similarly applied to the evaluation of broader classes of solid sorbents, including covalent organic frameworks and zeolites. Methods Methodologically, we have integrated machine-learning based structure selection, molecular-level simulation, as well as process-level simulation for optimizing CO2 capture performance. MOF properties derived from the molecular-level simulations are used as input for both the active-learning structure selection as well as the equations of the process-level temperature-swing adsorption model. Spurious structure removal We have used mofchecker32 to eliminate 3246 CoRE-MOF-2019 structures from our data set based on a suite of composition, coordination and consistency checks. An additional set of 413 structures were eliminated by an element-specific, charge-based heuristics. The heuristic aims at the removal of structures whose atoms present partial charges at the top, or bottom, 1% of a particular element charge distribution. The breakdown of elimination criteria by percentage and acceptable charge ranges per element are provided in Supplementary Note 1. For the analysis of water-tolerant MOFs, see Fig. 6, we applied additional, spurious structure removal criteria. White et al. proposed a pipeline that further refines the identification of spurious structures, leading to the removal of 224 structures that were flagged as “bad” by MOSAEC37. Active-learning based structure selection model As shown in Fig. 2a, active learning modeling is done by combining extensive featurization with iterative material selection. The featurization workflow combines multiple independent featurizers such as MAGPIE40, JARVIS41, persistent homology42, CGCNN43, synthesis complexity score (SCScore)33, Zeo++44 and PoreBlazer45. As a fist step of the iterative materials section process, feature selection is carried out to reduce the feature space. This is done using combination of Lasso regression, variable contribution from random forest ensembles and shapely metrics. The reduced feature space is used to predict adsorption free energies ΔG for CO2 and N2. We employ an ensemble of models to predict the adsorption performance. Our ML modeling tool box comprised of GPR, RFE, SVM, PLS, NN with details given in Supplementary Note 2. The best model is selected based on performance on the test dataset, as well as the prediction uncertainty over the entire database. The selected model is used to Pareto search and select the next set of materials to be subjected to detailed GCMC simulations. The Pareto search is performed by optimally trading-off \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G}_{{{{\rm{N}}}}_{2}}$$\end{document}ΔGN2 against \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-2\times \Delta {G}_{{{{\rm{CO}}}}_{2}}$$\end{document}−2×ΔGCO2 to maximize CO2/N2 selectivity and CO2 capacity simultaneously. Within each Pareto front, the selected materials are sorted and filtered based on their SCScore and based on the prediction uncertainty for adsorption free energy. We carried out 5 such active learning iterations, each including: detailed molecular simulation, (re-)building machine learning models to predict adsorption performance and material selection for subsequent iterations. Molecular-level simulations The molecular-level simulation framework consists of a combination of the Grand Canonical Monte Carlo (GCMC), Widom particle insertion and Molecular Dynamics (MD) methods for simulating the adsorption isotherms and diffusion coefficients, respectively. In all cases, we choose RASPA46,47 as our simulation engine. The interaction energies between non-bonded atoms were computed using the Coulomb and Lennard-Jones (LJ) potentials with Lorentz-Berthelot48 mixing rules 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${U}_{ij}({r}_{ij})=4{\varepsilon }_{ij}\left[{\left(\frac{{\sigma }_{ij}}{{r}_{ij}}\right)}^{12}-{\left(\frac{{\sigma }_{ij}}{{r}_{ij}}\right)}^{6}\right]+\frac{1}{4\pi {\epsilon }_{0}}\frac{{Q}_{i}{Q}_{j}}{{r}_{ij}},$$\end{document}Uij(rij)=4εijσijrij12−σijrij6+14πϵ0QiQjrij,where rij is the interatomic distance, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{ij}=\sqrt{{\varepsilon }_{ii}{\varepsilon }_{jj}}$$\end{document}εij=εiiεjj is the well depth and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{ij}=\frac{{\sigma }_{ii}+{\sigma }_{jj}}{2}$$\end{document}σij=σii+σjj2 is the well diameter between atoms i and j. The values of LJ parameters for framework atoms are a hybrid between the Universal Force Field (UFF)49 and DREIDING50 (see Supplementary Table 6), while the LJ parameters for the adsorbate molecules were taken from the TraPPE51 force field (see Supplementary Table 7). Electrostatic energy contributions were calculated using EQeq52 partial atomic charges Qi and the Ewald sum technique. Despite the emergence of structure databases with DFT-derived partial atomic charges53–55, we have chosen EQeq charges which, although approximate and subject to certain limitations, offer certain benefits56,57. The post-processing of GCMC time series and MD trajectories was fully automated. The equilibrated averages of molecular uptake were extracted from the GCMC time series using the Marginal Standard Error Rule (MSER) as implemented in pyMSER58 and shown in Supplementary Fig. 3. Adsorption isotherm curves for CO2, N2 and H2O were built by simulating the molecular uptake at several pressures, for each considered temperature, as shown in Supplementary Fig. 4. Multi-site Langmuir adsorption models (see Equation 1 on Supplementary Note 3.1) were adjusted to the data points, and the best fit was kept. The resulting analytical expressions for the adsorption as a function of pressure and temperature were input to the process-level model equations. The self-diffusion coefficients were extracted from the MD trajectories using an automated routine that locates the normal diffusion regime (〈r2〉 ∝ t, see Supplementary Fig. 6) in the mean squared displacement curve, as shown in Supplementary Fig. 7. Then, the self-diffusion coefficients Dμ calculated at different temperatures were adjusted to an Arrhenius expression (see Equation 4 on Supplementary Note 3.2.4), as shown in Supplementary Fig. 9. Finally, we have used the fitted expression for Dμ(T) as transport diffusivity under low loading approximation for the process modeling. For more information on the approach to molecular-level modeling, please refer to Supplementary Note 3. Process-level simulations The process-level model assumed a four-step temperature-swing adsorption cycle—i.e., adsorption, purge, desorption and cooling—in a fixed bed configuration, as depicted in Supplementary Fig. 10c. gPROMS® Model Builder software was used for the implementation. The packed bed is modeled as one-dimensional, including heat and mass transfer. The effective mass transfer coefficient59 is defined as 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{{{\rm{eff}}}}=\frac{15}{\left(\frac{{R}_{p}^{2}}{{D}_{p}^{*}}+\frac{{R}_{\mu }^{2}}{{D}_{\mu }}\right)},$$\end{document}Keff=15Rp2Dp*+Rμ2Dμ,where the pellet radius Rp = 2 mm and effective pellet diffusivity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D}_{p}^{*}=5\times 1{0}^{-8}\,{{{\rm{m}}}}^{2}/{{\rm{s}}}$$\end{document}Dp*=5×10−8m2/s represent the macropore (or intercrystalline) diffusion ignoring external film resistance. The crystallite radius Rμ and diffusivity Dμ represent the micropore (or intracrystalline) diffusion. Thermal management is achieved using a temperature-controlled water stream flowing through the heat exchange jacket of the bed as well as direct heat exchange through steam or cooler flue gas. The competitive adsorption of CO2 and N2 is modeled as a multi-component, mixed Langmuir isotherm derived from the combination of pure component isotherms during adsorption. The H2O is included in the process modeling as an inert purge gas during regeneration, with the purpose of accounting for the steam requirement within the specific energy evaluation. Mass balance was imposed for the gas and adsorbed phases (Equation 6 on Supplementary Note 4.1). The rate of change in the concentration of the adsorbed phase was modeled as a linear restorative term proportional to Keff (Equation (2)) towards the equilibrium concentration (Equation 1 on Supplementary Note 3.1). The energy balance at the gas and adsorbed phases, outer walls and external heat transfer fluid is described in Equation 7 on Supplementary Note 4.1. The momentum balance is modeled using the Ergun equation (Equation 6 on Supplementary Note 4.1) and the total pressure at each point is determined using the ideal gas law. The desorption temperature and the four cycle step times, namely, adsorption step time, purge step time, desorption step time and cooling step time, are subject to an optimization routine to determine the combination of values that lead to the best performance. The possible choices for these parameters contain inherent performance trade-offs, defining a Pareto front. The point with 95% purity is selected from the Pareto front as representative of the material’s performance metrics. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file Source data Source Data Source Data