DOI: 10.1002/cphc.201200554

Similarity-Driven Discovery of Zeolite Materials for Adsorption-BasedSeparations

Richard L. Martin,[a] Thomas F. Willems,[a] Li-Chiang Lin,[b] Jihan Kim,[c] Joseph A. Swisher,[b] Berend Smit,[b, c] andMaciej Haranczyk*[a]

Crystalline porous materials such as zeolites and metal-organicframeworks (MOFs) have applications as materials for separa-tions, catalysis and gas storage.[1, 2] In particular, they are prom-ising candidates for use as selective CO2 adsorbents for carboncapture.[3] As their importance in energy-related applicationsgrows, so does the need for new materials to meet engineer-ing and economic constraints. Large databases of computa-tionally predicted zeolites[4] and MOFs[5] are being generated,in principle enabling researchers to screen for outstanding ma-terials. However, screening large databases of materials posesa serious challenge. Established computational techniques—in-cluding electronic structure calculations and molecular simula-tions—allow for accurate prediction of guest-related propertiesof a single material.[1a] However, the computational cost ofthese calculations prohibits the characterization of a databaseof millions of structures, which would be required to performbrute-force screening. Moreover, as structures that exhibit de-sired properties constitute only a small fraction of predictedmaterials, a brute-force screening approach would involvemany wasteful calculations.

Recent studies focused on screening porous materials forgas separations and storage have applied pre-screens to limitthe number of structures to be characterized. These pre-screens have been based on either geometrical parameters[6]

or approximated physical properties.[5] Furthermore, our grouphas developed tools to increase the number of structures thatcan be characterized in a brute-force approach to about100000 materials by exploiting the parallel computing powerof clusters of graphics processing units (GPUs),[7] incorporatingautomatic structure analysis algorithms.[8]

Still, these developments do not allow efficient screening ofa database of millions of materials. Herein, we present a novelapproach to materials discovery which further reduces thecomputational cost of identifying promising candidate materi-

als by at least an order of magnitude. Our method is based onthe similar property principle[9]—successfully employed in drugdiscovery[10]—which states that similar chemical structureshave similar properties, a simple yet powerful concept thatleads to efficient chemical space exploration. Our databasescreening approach has two steps. First, we characterize di-verse materials to identify substructural features which domi-nate physical properties of interest; second, the database is an-alyzed using a computationally inexpensive substructure com-parison method to identify materials which exhibit similarproperty-determining substructures. As a result, we can effi-ciently discover a majority of the highest-affinity adsorptionstructures by characterizing only a fraction of a database. Wedemonstrate this procedure by screening a previously unexa-mined database of millions of zeolites[4] for structures for theadsorption-based separation of CO2/N2, a critical separation forpost-combustion carbon capture.

Adsorption-based separation processes are based on differ-ences in adsorption of two or more to-be-separated species.Two physical properties that influence adsorption behavior ofa material are the Henry coefficient (KH), which determines theaffinity of a guest for adsorption in the low-pressure regime,and the heat of adsorption (HA), the latent heat given off bya guest adsorbing to the framework. These can be calculatedin Monte Carlo simulations[11] (all simulations discussed hereinwere performed at 300 K using the force fields of ref. [12] andref. [13]). In the case of CO2/N2 separation, since the KH and HA

of N2 are comparatively low, a good separation material willhave a high KH and HA for CO2.[14]

KH and HA are physical properties of an entire structure, butthey can be dominated by the presence of strong bindingsites for the guest molecule due to the exponential contribu-tion of guest–host interaction energies.[11] A structure that ex-hibits strong binding sites—“sweet spots”—will have signifi-cant KH and HA (Figure 1). Moreover, structures with similarbinding sites will be similar in their adsorption properties. Weexploit this behavior in order to efficiently discover materialsfor optimum adsorption through geometry-based structurecomparison.

The first step in our process is to characterize a small, di-verse sample of structures. In a large dataset of porous materi-als there exist many different void space features: large andsmall cages; wide and narrow, straight and curved channels ;junctions, and networks of different dimensionality. Throughdissimilarity-based selection of structures we can efficientlysample this shape diversity with only a few tens of materials.[15]

The strength of our approach is the exploitation of the effi-cient sampling provided by diversity selection to obtain repre-

[a] Dr. R. L. Martin, T. F. Willems, Dr. M. HaranczykComputational Research DivisionLawrence Berkeley National LaboratoryBerkeley, CA 94720 (USA)E-mail : mharanczyk@lbl.gov

[b] L.-C. Lin, J. A. Swisher, Prof. Dr. B. SmitDepartment of Chemical and Biomolecular EngineeringUniversity of California, BerkeleyBerkeley, CA 94720 (USA)

[c] Dr. J. Kim, Prof. Dr. B. SmitMaterials Science DivisionLawrence Berkeley National LaboratoryBerkeley, CA 94720 (USA)

Supporting information for this article is available on the WWW underhttp://dx.doi.org/10.1002/cphc.201200554.

ChemPhysChem 2012, 13, 3595 – 3597 � 2012 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 3595

sentative geometric search queries from characterizing onlya small training set of materials. By characterizing the diverseset we observe that the most energetically favorable bindingsites for the examined guest molecules comprise frameworkatoms arranged so as to approximately enclose the molecule.Moreover, eight silicon atoms of the framework are positionedin close proximity to each of the two extremes of the linearmolecule, for example, the oxygen atoms for CO2. Thus, foreach guest molecule, we can screen for materials which exhibittwo such eight-silicon features in close proximity, indicatingthe potential existence of a guest-molecule-enveloping sweetspot (Figure 2).

To identify binding sites within a zeolite framework we firstdetermine the guest-accessible regions of void space.[16] By fo-cusing on the surface of guest-accessible cavities, we can have

confidence that the guest canindeed diffuse to identified bind-ing sites. We then performshape comparison followingref. [17] to detect geometricmatches between substructuresof the framework and query fea-tures. Finally, the interactionenergy of the guest molecule ina binding site can be estimatedby docking the guest betweenthe two features, and estimatingthe molecule’s minimum-energyorientation and position (Fig-ure 2 a). The guest–host interac-tion energy can then be calculat-ed using a force field.[12, 13] Wenote that for chemically hetero-geneous materials such as MOFs,a simpler technique such aspharmacophore matching[18]

may be sufficient to identify binding sites.We validated our screening approach on the PCOD set of

about 140000 zeolites previously characterized using GPU-based molecular simulations.[14] We rank each structure in theset by its predicted minimum guest–host interaction energy,our hypothesis being that a strongly favorable energy, correct-ly predicted, will dominate adsorption properties. The perfor-mance of our method is quantified by how this ranking posi-tions the best adsorbing materials—those within the top 1 %of KH or HA—near the top of the ranked list. We measure theenrichment in the first 500 structures in the ranked list, that is,the ratio of structures retrieved therein with the desired prop-erty, versus the number that would be retrieved at random.

Accurate retrieval of structures with outstanding adsorptionproperties can be achieved by our technique. At the cost ofcharacterizing only 500 zeolites, we retrieve 327 materialswithin the top 1 % of CO2 KH (i.e.>1.301 � 10�3 mol kg�1 Pa�1) inthe PCOD dataset, that is, a greater than 65-fold enrichmentover random selection or brute-force characterization. A great-er than 72-fold enrichment is similarly observed for materialswithin the top 1 % of CO2 HA (i.e. <�49.730 kJmol�1).

Furthermore, to demonstrate the applicability of ourmethod on a database one order of magnitude larger, wescreened the PZEF set[4] of about one million hypothetical zeo-lites that have not previously been characterized. Within thisset we detected about 30000 new materials that exhibited theknown features. This set of candidate materials representsa quantity that can be examined in detail with more sophisti-cated methods. The single best CO2-adsorbing material discov-ered in the PZEF set is illustrated in Figure 3.

In conclusion, our method enables the efficient discovery ofzeolite materials with optimum adsorption properties. Asa result we can screen material databases at a fraction of thecost of brute-force characterization, enabling the analysis ofdatasets at least an order of magnitude larger than those ex-amined by current state-of-the-art methods. Our technique is

Figure 1. Three hypothetical zeolites and their energy contours for CO2; CO2-accessible space is green, and low-energy (favorable) positions blue. a) PCOD8286959, with CO2 KH 2.1 � 10�3 mol kg�1 Pa�1. b) PCOD8290342, the 5thmost similar by overall shape to (a); however without sweet spots the KH is only 5.7 � 10�6. c) PCOD8278310, the124647th most similar to (a) ; due to the presence of sweet spots, a similarly high KH of 2.3 � 10�3, despite littleshape similarity. Similarity of void space shape calculated using methodology of ref. [15]

Figure 2. a) Two features (black, purple) are detected in a substructure ofzeolite PCOD8207615. A vector between features (blue) enables estimationof the CO2 binding position. It can be seen that the minimum-energy posi-tion for CO2 obtained from simulation lies approximately along this vector ;b)–g) all six CO2 binding site features identified in the diverse set of hypo-thetical zeolites.

3596 www.chemphyschem.org � 2012 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim ChemPhysChem 2012, 13, 3595 – 3597

based on material informatics: diversity selection, similaritysearching and docking, and it selects promising candidate ma-terials by identifying substructural features that dominatephysical properties of interest and screening for them efficient-ly within a database. Our screening approach is novel in that itis based on local structural features, allowing us to identifyoutstanding adsorption materials based on their specificguest-binding sites. We have demonstrated this computation-ally inexpensive method by screening a database of about onemillion zeolites for structures for the adsorption-based separa-tion of CO2/N2 mixtures, and have shown that many of thestrongest CO2 adsorbing structures in a dataset can be identi-fied over 65 times more efficiently than through brute-forcemethods. Our approach is also appropriate for other gas sepa-rations, with results for the separation of C2H4/C2H6 describedin the Supporting Information.

Acknowledgements

This work was supported by the Director, Office of Science, Ad-vanced Scientific Computing Research, of the U.S. Department ofEnergy under Contract No. DE-AC02-05CH11231. In addition, itwas supported jointly by DOE Office of Basic Energy Sciences andthe Office of Advanced Scientific Computing Research throughSciDAC project #CSNEW918 entitled “Knowledge-Guided Screen-ing Tools for Identification of Porous Materials for CO2 Separa-tions”, and as part of the Center for Gas Separations Relevant toClean Energy Technologies, an Energy Frontier Research Center

funded by the U.S. Department of Energy, Office of Science, Officeof Basic Energy Sciences under Award Number DE-SC0001015.This research used resources of the National Energy Research Sci-entific Computing Center (NERSC), which is supported by theOffice of Science of the U.S. Department of Energy under Con-tract No. DEAC02-05CH11231.

Keywords: adsorption · carbon dioxide · microporousmaterials · virtual screening · zeolites

[1] a) B. Smit, T. L. M. Maesen, Chem. Rev. 2008, 108, 4125 – 4184; b) R. Krish-na, J. M. van Baten, Chem. Eng. J. 2007, 133, 121 – 131.

[2] a) J.-R. Li, J. Sculley, H.-C. Zhou, Chem. Rev. 2012, 112, 869 – 932; b) R. B.Getman, Y.-S. Bae, C. E. Wilmer, R. Q. Snurr, Chem. Rev. 2012, 112, 703 –723.

[3] a) K. Sumida, D. L. Rogow, J. A. Mason, T. M. McDonald, E. D. Bloch, Z. R.Herm, T.-H. Bae, J. R. Long, Chem. Rev. 2012, 112, 724 – 781; b) D. M.D’Alessandro, B. Smit, J. R. Long, Angew. Chem. 2010, 122, 6194 – 6219;Angew. Chem. Int. Ed. 2010, 49, 6058 – 6082.

[4] a) M. W. Deem, R. Pophale, P. A. Cheeseman, D. J. Earl, J. Phys. Chem. C2009, 113, 21353 – 21360; b) R. Pophale, M. W. Deem, P. A. Cheeseman,Phys. Chem. Chem. Phys. 2011, 13, 12407 – 12412.

[5] C. E. Wilmer, M. Leaf, C. Y. Lee, O. K. Farha, B. G. Hauser, J. T. Hupp, R. Q.Snurr, Nat. Chem. 2012, 4, 83 – 89.

[6] E. Haldoupis, S. Nair, D. S. Sholl, Phys. Chem. Chem. Phys. 2011, 13,5053 – 5060.

[7] J. Kim, R. L. Martin, O. R�bel, M. Haranczyk, B. Smit, J. Chem. TheoryComput. 2012, 8, 1684 – 1693.

[8] a) M. Haranczyk, J. A. Sethian, Proc. Natl. Acad. Sci. USA 2009, 106,21472 – 21477; b) M. Haranczyk, J. A. Sethian, J. Chem. Theory Comput.2010, 6, 3472 – 3480; c) R. L. Martin, Prabhat, D. D. Donofrio, J. A. Sethi-an, M. Haranczyk, Int. J. High Perform. C. , 10.1177/1094342011431591.

[9] M. A. Johnson, G. M. Maggiora, Concepts and Applications of MolecularSimilarity, John Wiley & Sons, New York, 1990.

[10] a) V. J. Gillet, WIREs Comput. Mol. Sci. 2011, 1, 580 – 589; b) P. Willett,WIREs Data Mining Knowl. Discov. 2011, 1, 241 – 251.

[11] D. Frenkel, B. Smit, Understanding Molecular Simulation, Academic Press,San Diego, 2002.

[12] E. Garcia-P�rez, J. B. Parra, C. O. Aria, A. Garc�a-S�nchez, J. M. van Baten,R. Krishna, D. Dubbeldam, S. Calero, Adsorption 2007, 13, 469 – 476.

[13] B. Liu, B. Smit, F. Rey, S. Valencia, S. Calero, J. Phys. Chem. C 2008, 112,2492 – 2498.

[14] L.-C. Lin, A. Berger, R. L. Martin, J. Kim, J. Swisher, K. Jariwala, C. H. Ry-croft, A. Bhown, M. W. Deem, M. Haranczyk, B. Smit, Nat. Mater. 2012,11, 633 – 641.

[15] R. L. Martin, B. Smit, M. Haranczyk, J. Chem. Inf. Comput. Sci. 2012, 52,308 – 318.

[16] T. F. Willems, C. H. Rycroft, M. Kazi, J. C. Meza, M. Haranczyk, MicroporousMesoporous Mater. 2012, 149, 134 – 141.

[17] R. L. Martin, E. J. Gardiner, S. Senger, V. J. Gillet, J. Chem. Inf. Comput. Sci.2012, 52, 757 – 769.

[18] R. P. Sheridan, S. K. Kearsley, Drug Discovery Today 2002, 1, 903 – 911.

Received: July 6, 2012Published online on August 22, 2012

Figure 3. Structure and energy contour of PZEF1264075, the material withthe highest affinity for CO2 adsorption discovered to date in the PZEF hypo-thetical zeolite set, which exhibits a CO2 KH of 6.1 � 10�2 mol kg�1 Pa�1.

ChemPhysChem 2012, 13, 3595 – 3597 � 2012 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.chemphyschem.org 3597