Data-driven design of metal–organic frameworks for wet flue gas CO₂ capture

Abstract

Limiting the increase of CO₂ in the atmosphere is one of the largest challenges of our generation¹. Because carbon capture and storage is one of the few viable technologies that can mitigate current CO₂ emissions², much effort is focused on developing solid adsorbents that can efficiently capture CO₂ from flue gases emitted from anthropogenic sources³. One class of materials that has attracted considerable interest in this context is metal–organic frameworks (MOFs), in which the careful combination of organic ligands with metal-ion nodes can, in principle, give rise to innumerable structurally and chemically distinct nanoporous MOFs. However, many MOFs that are optimized for the separation of CO₂ from nitrogen^4,5,6,7 do not perform well when using realistic flue gas that contains water, because water competes with CO₂ for the same adsorption sites and thereby causes the materials to lose their selectivity. Although flue gases can be dried, this renders the capture process prohibitively expensive^8,9. Here we show that data mining of a computational screening library of over 300,000 MOFs can identify different classes of strong CO₂-binding sites—which we term ‘adsorbaphores’—that endow MOFs with CO₂/N₂ selectivity that persists in wet flue gases. We subsequently synthesized two water-stable MOFs containing the most hydrophobic adsorbaphore, and found that their carbon-capture performance is not affected by water and outperforms that of some commercial materials. Testing the performance of these MOFs in an industrial setting and consideration of the full capture process—including the targeted CO₂ sink, such as geological storage or serving as a carbon source for the chemical industry—will be necessary to identify the optimal separation material.

Main

Different strategies have been developed to mitigate the negative effects of water on the CO₂/N₂ separation selectivity in MOF materials. For example, some MOFs have open metal sites at which amines can be attached, taking advantage of the specific amine chemistry that is also used in conventional amine scrubbing^10,11,12,13. A previous screening study¹⁴ investigated whether MOFs could adsorb CO₂ in the presence of water, and the results suggested that such MOFs could be de novo designed. In this work, we develop a systematic strategy for the design and preparation of custom-made MOFs that can capture carbon from wet flue gases. Our design methodology is inspired by the rational design of drug molecules, in which organic molecules that fit well into the binding pocket of a protein are mined from databases of known chemicals^15,16. The difference in our case is that the ‘drug molecule’ is known (that is, CO₂), but the substrate that binds it optimally (the MOF) is not. We therefore generated a library of 325,000 hypothetical MOFs, and screened each material for its CO₂/N₂ selectivity and its CO₂ working capacity. The chemical building blocks used in the generation of these materials are shown in Extended Data Figs. 1 and 2. Figure 1a shows that 8,325 hypothetical materials possess a working capacity for CO₂ greater than 2 mmol g⁻¹ and a CO₂/N₂ selectivity greater than 50—performance that surpasses that of zeolite 13X under dry conditions¹⁷.

Fig. 1: Computational screening of MOFs for strong CO₂ adsorption and selectivity.

a, Results from the screening of 325,000 hypothetical MOFs under conditions that mimic post-combustion capture (adsorption at ambient temperature and 1 bar with a molar ratio of 15:85 CO₂/N₂ and regeneration at 363 K and 0.1 bar). The materials in the green box were selected for more refined screening and adsorbaphore identification; the colour-coding represents the number of MOFs according to the colour bar on the right. b, The H₂O affinity of the top-performing materials, as characterized by a plot of the Henry coefficients (K_H) for H₂O against those of CO₂. The colours represent the three different adsorbaphores found in the top-performing structures: A1, parallel aromatic rings; A2, metal–oxygen bridges; A3, open metal sites. Some materials have both A1 and A2 sites. c, The adsorbaphore containing parallel aromatic rings (A1), which was discovered using the feature-recognition algorithm.

Source data

A key part of drug design is to analyse the optimally binding molecules for a common feature or spatial arrangement of atoms at the binding site, which is known as the pharmacophore¹⁵. In analogy with this, we coin the term ‘adsorbaphore’ to describe the common pore shape and chemistry of a binding site in a MOF that provides optimal interactions to preferentially bind to a particular guest molecule, in this case CO₂. From our top-ranked 8,325 materials, we identified 106,680 such CO₂-binding sites (see Extended Data Fig. 3 for examples). A similarity analysis of these binding sites revealed three main classes of adsorbaphore: A1, two parallel aromatic rings with interatomic spacings of approximately 7 Å (31% of all binding sites); A2, metal–oxygen–metal bridges (32%); and A3, open metal sites (21%) (see Supplementary Information for details). Subsequently, we screened the materials that possessed these adsorbaphores for their affinity for water. Figure 1b shows the Henry coefficient for water in these high-performing materials. Analysis of the data shows that the materials with the parallel aromatic rings (A1) have a low Henry coefficient for H₂O, whereas those with metal–oxygen bridges (A2) and open metal sites (A3) tend to have higher Henry coefficients (Fig. 1b). A graphical representation of the different adsorbaphores is presented in Extended Data Fig. 4. Comparison of the binding energies—computed using density functional theory—for the adsorbaphore shown in Fig. 1c indicates a preference for CO₂ (−10.2 kcal mol⁻¹) over N₂ and H₂O by 2.7 and 1.5 kcal mol⁻¹, respectively (see Extended Data Table 1). The parallel aromatic rings provide a near-optimum interaction with all three atoms of CO₂, whereas for H₂O the lack of hydrogen-bonding sites limits its binding energy.

The next step was to identify a subclass of MOFs in our library that contains the preferred adsorbaphore. From an experimental point of view, MOFs with the frz topology—characterized by tetra-carboxylated organic ligands coordinated to one-dimensional metal–oxygen rods—are an attractive starting point. One such example has been synthesized with indium as a metal node, resulting in a structurally stable, non-breathing MOF¹⁸. In this topology, the metal rods provide an ideal scaffolding to which we can attach our adsorbaphore. By varying the metal ion we have some flexibility to tune the distance between the aromatic rings. Our calculations predict that the ideal adsorbaphore distance of 6.5–7.0 Å—which was determined by adjusting the spacing of the aromatic rings incrementally (Extended Data Fig. 5)—can be approached if In(iii) is replaced by Al(iii) (Extended Data Table 2). In addition, aluminium is an attractive choice because it is an abundant metal and it ensures a strong bond with the carboxylate O-atoms of the ligands¹⁹; this considerably improves the thermal and hydrolytic stability of a MOF^20,21.

We generated a library of 35 isoreticular materials using our MOF-generation algorithm²², and from the mixture isotherms we computed the CO₂/N₂ selectivity of the materials in dry and wet flue gases (Extended Data Figs. 6 and 7). Our calculations show that all of our predicted materials maintain an excellent selectivity at low pressures, and in about 75% of these materials the selectivity was not influenced by the presence of water under flue-gas conditions. The concept of an adsorbaphore focuses on the design of an adsorption site that optimizes selectivities at low pressure. At higher partial pressures of water, its adsorption is dominated by the energetics of hydrogen-bond formation. Further analyses showed that, for the materials that maintain a high CO₂ uptake at high humidity, it is the pore shape that frustrates the formation of these hydrogen bonds. This is illustrated in Fig. 2a, b, which compares the effect of water on the CO₂ uptake of two materials that have the same adsorbaphore but different pore structures (hypothetical MOFs m8o67 and m8o71). Figure 2a shows that m8o67 is resistant to H₂O flooding: even at a relative humidity of approximately 85%, we find that H₂O has only a small effect on CO₂ capacity. Conversely, m8o71 completely loses its CO₂ capacity at 60% relative humidity (Fig. 2b). In Fig. 2c, d we visualize the hydrogen-bond network that is formed at 100% relative humidity in both materials. For m8o71 we see a complete hydrogen-bonding network (Fig. 2d), whereas for m8o67 (Fig. 2c) we observe a less extensive network; the benzoate groups that separate the adsorbaphores frustrate the formation of a complete hydrogen-bonding network.

Fig. 2: The effects of water on different MOFs with the same adsorbaphore.

a, b, Simulated adsorption of a ternary mixture of CO₂/N₂/H₂O at 313 K by hypothetical MOFs m8o67 (a) and m8o71 (b). The partial pressure of CO₂ was held at 0.15 bar as the relative humidity was increased. The N₂ uptake was negligible and is not shown. c, d, Visualization of the water loading at 100% relative humidity in m8o67 (c) and m8o71 (d). The benzoate groups are represented by grey sticks; water is shown using red and white space-filling atoms. In m8o67, the benzoate groups perpendicular to the plane of the figure prevent hydrogen-bond formation across the adsorbaphores.

Source data

On the basis of these predictions, we synthesized two frz-based MOFs using organic ligands that possess the water-frustrating properties reported above: Al-PMOF¹⁹ (m8o66) and Al-PyrMOF (m8o67). These MOFs are based on one-dimensional rods of Al(iii) linked by TCPP (tetrakis(4-carboxyphenyl)porphyrin) and TBAPy (1,3,6,8-tetrakis(p-benzoic acid)pyrene) ligands, respectively (Fig. 3a, b). Figure 3c, d shows no loss of crystallinity upon activation as well as upon exposure to different harsh conditions, including immersion in water for 7 days.

Fig. 3: Structural representation and stability of [Al-PMOF] and [Al-PyrMOF].

a, b, Ball-and-stick representation of the structures of [Al-PMOF] (a) and [Al-PyrMOF] (b). The orientation of the tetracarboxylate ligands around the Al(iii) rods results in the generation of the three-dimensional non-interpenetrated structures containing the adsorbaphore (red box). Atom colour code: pink, Al; grey, C; blue, N; red, O; pale yellow, H. c, d, Laboratory powder X-ray diffraction patterns of [Al-PMOF] (c) and [Al-PyrMOF] (d). Black, simulated; red, as-synthesized material; blue, acetone-exchanged material; green, activated material; sky blue, activated material immersed in liquid water for 7 days; pink, activated material exposed to a controlled atmosphere of nitric acid vapour for 3 h.

Source data

By identifying adsorbaphores in these hypothetical materials, we assume that our in silico screening method can correctly predict the structure of a MOF, its adsorption properties, and the nature of the binding sites of CO₂ and H₂O. We are able to test these assumptions for Al-PMOF and Al-PyrMOF. In Fig. 4a we show that the experimental and predicted CO₂ and N₂ adsorption isotherms are in good agreement. The CO₂ binding positions in the adsorbaphore and the effect of H₂O are more challenging to observe experimentally. The siting of CO₂ was studied using in situ CO₂-loading powder X-ray diffraction. Upon loading, we observed a considerable change in the intensity and peak position of the Bragg reflections (see Supplementary Fig. 2.1). Subsequent Rietveld refinement and Fourier analysis²³ revealed the preferred locations of CO₂ in the pores of Al-PMOF, as shown in Fig. 4b. These results confirm that CO₂ preferentially adsorbs in the adsorbaphore.

Fig. 4: CO₂ adsorption, ¹³C cross-polarization MAS NMR and breakthrough experiments for [Al-PMOF] and [Al-PyrMOF].

a, Experimental (filled) and computational (open) single-component adsorption isotherms for CO₂ (squares) and N₂ (circles) adsorption collected on activated [Al-PMOF] (red) and [Al-PyrMOF] (blue) at 313 K. b, Rietveld refinement of the X-ray diffraction data (Supplementary Information) revealed that CO₂ binding in [Al-PMOF] occurs between the porphyrin cores—that is, in the adsorbaphore. c, The linewidth of each carbon peak of the TBAPy ligand in [Al-PyrMOF] in the ¹³C cross-polarization MAS spectrum, plotted as a function of relative humidity. Each carbon atom in the ligand is labelled in the inset. d, Linewidths extracted from the ¹³C static NMR spectra of ¹³CO₂ loaded in [Al-PyrMOF], plotted against relative humidity (RH). e, CO₂ capture capacity profiles for [Al-PyrMOF] and [Al-PMOF] during breakthrough experiments under dry and humid (85% relative humidity) conditions, with 85/15 v/v of N₂/CO₂ (313 K and 1 bar). f, Benchmarking the CO₂ working capacity of [Al-PyrMOF] and [Al-PMOF] against UiO-66-NH₂, activated carbon and zeolite 13X under dry and humid (85% relative humidity) conditions, with 85/15 v/v of N₂/CO₂ (313 K and 1 bar). For wet flue gases, we studied the performance stability after 3 cycles for reference materials, and after 10 cycles for [Al-PyrMOF] and [Al-PMOF].

Source data

The effect of water on the siting of CO₂ has been further addressed by solid-state nuclear magnetic resonance (NMR) analysis. Under magic-angle spinning (MAS), high-resolution ¹³C NMR chemical shifts are very sensitive to changes in the chemical environment. The ¹³C NMR spectra of Al-PyrMOF and Al-PMOF are provided in Extended Data Fig. 8, which also shows the assignment of the peaks to specific atoms on the MOF. The chemical shifts associated with the atoms of the adsorbophore (inset) are shown in Fig. 4c as a function of the water concentration. At low levels the adsorbophore atoms experience no change in chemical environment with water loading, whereas at the highest water loadings there are modest changes in the carbon-13 chemical shifts of only those atoms that are close to the aluminium-coordinated carboxylate groups next to the adsorbaphore (carbons B and F in Fig. 4c). This broadening is consistent with dipolar broadening from proximate water molecules, thus confirming that the adsorbaphore itself is not a preferential adsorption site for H₂O.

Our simulations predict that CO₂ adsorbed in the adsorbaphore is insulated from the adsorption of water. Because the ¹³C NMR spectrum of adsorbed ¹³CO₂ is extremely sensitive to the proximity of water molecules in terms of chemical shift and line broadening, any disruption of the chemical environment of adsorbed CO₂ by water should be immediately apparent. Figure 4d shows that the chemical shift of the adsorbed ¹³CO₂ is independent of water content, although a broadening of the ¹³C NMR peak is observed with increasing humidity. If this broadening is due to the proximity of the protons in water, it should disappear if the experiments are repeated using D₂O; however, it does not (Fig. 4d). This observation corroborates our simulation results (Fig. 2c), confirming that water has only a limited effect on CO₂ adsorption in Al-PyrMOF.

The ability of these materials to capture CO₂ from wet flue gases is of important practical concern. We therefore used a breakthrough experiment to determine the capture capacity of both Al-PMOF and Al-PyrMOF for a mixture of CO₂/N₂ under dry- and humid-conditions²⁴ (Fig. 4e). These results confirm the predictions of the simulations (Extended Data Fig. 7): humidity in the flue gases has only a minimal influence on the capture capacity of Al-PMOF, whereas for Al-PyrMOF the capture capacity is in fact enhanced. Furthermore, repeated cycling²⁵ (Fig. 4f) does not result in degradation of the material or a change in separation performance. It is instructive to compare the performance of our materials with that of a set of reference materials, including those that are commercially available—such as zeolite 13X and activated carbon—and a water stable, amino-functionalized MOF, UiO-66-NH₂. The capture capacity of these reference materials lies between that of Al-PyrMOF and Al-PMOF in dry flue gases; however, unlike our MOFs, their performance reduces considerably in humid flue gases. Although our materials do not have the highest reported working capacities¹⁴, it is encouraging to see that, in wet flue gases, Al-PMOF outperforms commercial materials such as zeolite 13X and activated carbon.

Large-scale screening of databases of hypothetical MOFs for various gas separation and storage applications has been reported previously^26,27,28,29; however, here we have focused on identifying binding pockets—or structural motifs termed adsorbaphores—as synthetic targets, rather than whole materials. This enhances the synthetic viability of the approach, as demonstrated by the identification of one new material with the targeted adsorbaphore that was synthesized and shown to adsorb CO₂ as predicted. The concept of linking computational screening with the synthesis of the corresponding materials through such adsorbaphores should be applicable to other gas separations of increasing complexity.