Poster 4: Similarity Searching and Screening for Porous MaterialsRichard Luis Martin1, Maciej Haranczyk1
|1Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, USA|
|We present recently developed techniques for high-throughput analysis and discovery of porous materials, inspired by established methods in chemoinformatics. We describe how concepts from similarity searching and screening of small molecule databases can be extended to overcome the unique challenges presented by periodic, porous crystal structures and chemically homogeneous systems.|
Crystalline porous materials have been exploited in industrial applications for many years, for instance as catalysts for oil refinement, water softeners, and membranes for separations. They are also promising candidates for economically competitive application to emerging energy challenges, such as vehicular natural gas storage, and carbon dioxide capture. Porous materials exhibit complex networks of internal space, which permit the diffusion of guest chemical species. The size, shape and connectivity of a material's pore structure determine the interactions which occur between the guest and the host material.
Identifying materials with the specific pore structure advantageous for a particular application is a great challenge. Zeolites, the most well-known class of porous materials, consist solely (in their simplest, all-siliceous form) of tetrahedral arrangements of silicon and oxygen; nevertheless, approximately 200 unique zeolite topologies are known to exist, and over two million hypothetically achievable zeolite topologies have been computationally enumerated. With such a vast search space of possible materials, it is clear that exhaustive synthesis is not a feasible strategy for materials discovery, and so there is a pressing need for computational methods for high-throughput analysis of porous materials.
In this contribution we present recently developed techniques for the representation, comparison and screening of very large sets of porous materials. We describe Voronoi holograms, pore descriptors based on the Voronoi decomposition, and a similarity function tailored for their comparison. We illustrate similarity searching and dissimilarity-based selection for porous materials using this approach, which has enabled the automated discovery of similar structures and construction of diverse training sets of materials. Finally, we describe how these advancements have facilitated a high-throughput database screening technique which has identified thousands of previously unexplored promising candidate materials for carbon dioxide capture.
The Voronoi hologram abstracts the pore network inside a material as a vector, wherein each entry encodes the existence or multiplicity of a particular local cavity shape within the pore network. In a similar fashion to molecular fingerprints, such an abstraction enables high-speed comparison of materials' pore networks, according to an appropriate similarity function. By the similar property principle, similar materials exhibit similar Voronoi holograms, and we demonstrate that this concept can be utilized to automatically identify families of related material structures. Furthermore, we explore how the choice of similarity function has a large impact on the calculation of material similarity, and in particular, on dissimilarity-based selection.
Local cavity shapes within a porous material have a significant influence on its chemical behavior, for instance, how a diffusing guest molecule may adsorb to (desorb from) the internal surface. These properties of porous materials make them promising candidates as reusable adsorbents for gas separations. For instance, in carbon dioxide capture, adsorbent materials are desired where the strength of interaction between carbon dioxide and the material is such that adsorption of carbon dioxide is energetically preferable to that of other, less harmful flue gases. Molecular simulation is an established computational technique for predicting these properties of a material; however, even with the latest advanced hardware and algorithms, its computational expense prohibits its use as a tool to analyze databases of millions of materials.
In this presentation we illustrate that the similar property principle can be applied to the discovery of materials with promising properties for gas separations from within large databases. By simulating a diverse subset of zeolite materials, it can be observed that even among shape-diverse materials, the best adsorbents exhibit similar local pore shapes: favorable binding sites for a guest molecule. This structure-property relationship can be exploited to efficiently screen very large material databases based on structural geometry rather than simulation. We describe such a binding site screening method utilizing clique detection, and show that promising carbon dioxide capture materials can be discovered approximately two orders of magnitude faster than through exhaustive simulation.