Automated discovery of porous molecular materials facilitated by characterization of molecular porosity
- Autores: Ismael Gómez García
- Directores de la Tesis: Maciej Haranczyk (dir. tes.)
- Lectura: En la Universidad Carlos III de Madrid ( España ) en 2021
- Idioma: español
- Tribunal Calificador de la Tesis: German Ignacio Sastre Navarro (presid.), Javier Carrasco Rodríguez (secret.), Andreas Mavrantonakis (voc.)
- Programa de doctorado: Programa de Doctorado en Ciencia e Ingeniería de Materiales por la Universidad Carlos III de Madrid
- Materias:
- Enlaces
- Tesis en acceso abierto en: e-Archivo
- Resumen
Introduction Porous materials are critical to many industrial sectors, including petrochemicals, energy and water. Traditional porous polymers and zeolites are currently most widely employed within membranes, as adsorbents for separations and storage, and as heterogeneous catalysts. The emerging advanced porous materials, e.g. extended framework materials and molecular porous materials, can boost performance and energy-efficiency of the current technologies because of the unprecedented level of control of their structure and function. The enormous possibilities for tuning these materials by changing their building blocks mean that, in principle, optimally performing materials for a variety of applications can be systematically designed. However, the process of finding a set of optimal structures for a given application could take decades using the traditional materials development approaches. This is a substantial payoff for developing tools and approaches that can accelerate this process. Among advanced porous materials, porous molecular materials are one of the most recent members though they have already attracted significant interest. Porous molecular materials are built from discrete molecules interacting with to each other through non-covalent bonds. Their porosity emerges either through inefficient packing of the atoms of the molecule itself (i.e. intrinsic molecular porosity), or through inefficient packing of molecules in the solid, or both. Among molecules supporting porous molecular materials, porous molecules are of especial interest. Some authors refer to two different subclasses of porous molecules, i.e. belt molecules (i.e. porous molecules with two ways to access, or windows) and cage molecules (i.e., porous molecules which shape roughly reminds to that of a cage). The exact definition of a porous molecule is not yet uniformly settled in the literature.
Porous molecular materials exhibit interesting properties that are not inherent to porous frameworks, which may make them more suitable to some applications than framework materials. The major advantages and otherwise unique features of porous molecular materials and/or of their molecular building blocks, as discussed in the literature, are: • Non-covalent synthesis: In contrast porous frameworks, formation of porous solid structures in porous molecular materials (either crystal or amorphous structures) does not require formation of new bonds. This may facilitate the process of growing larger porous solids (for instance, large porous molecular crystals can be grown by solvent evaporation, whereas in frameworks erroneous side reactions can negatively affect framework construction). The possibility of non-covalent material synthesis can be exploited to obtain solids with different properties: o Amorphous porous solids: Porous molecules do not need to be grown into crystals to harness their properties. It is possible to synthesize amorphous molecular solids that retain porosity at the material level.
o Structural flexibility: Since porous molecular materials do not have covalent bonds among their molecules, large rearrangements of the solid state are possible. This has led to porous crystals capable of switching their porosity (either "on/off” or changing selectivity). This property arises naturally in porous molecular materials and is expected not to happen in framework materials. However, some similar behavior has been noticed in some MOFs, and is being investigated.
o Solubility: Many porous molecules can be desolvated in organic solvents, allowing them to be solution processed into thin films and membranes. This property applies to molecular cages, but also to PIMs.
• Porous liquids: One particularly interesting feature of porous molecules is that they can form solutions in which porosity is maintained. These exploit intrinsic molecular porosity to produce materials that can retain pores even in liquid phase, which is obtained by adequate temperatures or solvents.
Materials and Methods In this PhD dissertation, we develop novel computational methodologies that accelerate the discovery of porous molecular materials with properties tailored to specific applications. The computational design of novel materials involves prediction, enumeration and characterization of both cage molecules and their crystalline and amorphous solid-state phases as well as reliable high-throughput prediction of the material performance ahead of synthesis. In particular, a schematic discovery pipeline is organized as follows: (1) First, molecules are proposed, either by a chemist who envisions them, or by systematic screening or enumeration of chemical databases. (2) Second, material structure prediction is performed, producing a computer model of the structural arrangement of the molecules in space. (3) Third, from the material structure, material properties are computationally predicted, including properties that are related with performance of the material in a given application. In this dissertation, we develop methodology that can be combined with the existing structure prediction and characterization methodologies to enable such computational design of porous molecular materials by exploiting material informatics and big data. We demonstrate this methodology by direct application to material discovery, introducing novel porous molecular materials in different applications at a computational level.
The current state of the art methodologies allows for execution of the key steps in the above discovery pipeline. Namely, computational materials science methods, e.g. molecular simulations and crystal structure prediction approaches allow building solid materials models starting from a molecular structure as well as predicting properties of the resulting solid materials. The recent literature demonstrates a few applications of these techniques in characterization of limited number of porous molecular materials. However, the material design process typically requires investigation of many candidate structures through either sequential improvement of the candidate or high-throughput screening of large sets of candidate materials. In this dissertation, we develop methodology that can be combined with the existing structure prediction and characterization methodologies to enable such computational design of porous molecular materials by exploiting material informatics and big data.
A representative part of this methodology development is the effort to characterize molecular porosity, which enables the first step of the discovery pipeline as a key aspect to identify porous molecules that can lead to novel porous molecular materials. Determination of molecular porosity is a task that so far has required a trained chemist to visually inspect molecular structure. Instead, we aim to provide the community with definitions and tools to determine whether a molecule is intrinsically porous, removing ambiguity and speeding up the analysis process, therefore enabling high-throughput studies of large sets of molecules. In this PhD work, we developed several algorithms to characterize molecular porosity. These algorithms address some key aspects of our problem, e.g.:
• Detection of internal voids in a provided molecule, and their characterization with a pore exposure ratio descriptor, which measures the exposure of internal voids to the surroundings of the molecule. This is required to classify voids as molecular pores.
• Detection of molecular windows that connect the internal space of the molecule with its surroundings; as the result, the number of windows present in the molecule is provided.
• Detection of paths of maximal distance to atoms that connect the surroundings of the molecule to its internal voids, obtaining two molecular descriptors: (1) the number and (2) size of entry paths can be calculated.
• Calculations of probe sizes that could access/occupy pores; as the result of solving this problem, we obtain two molecular descriptors: (1) the molecular largest cavity diameter and (2) pore limiting diameter.
• Computation of internal surface of a molecule.
Addressing these problems required the combination of computational geometry techniques to analyze relative atom positions, the usage of graph algorithms for the determination of topological properties of both molecules and the empty space, and the utilization of machine learning techniques to deal with structural similarities and noise.
Beyond molecular characterization, we aim to introduce tools that help on the overall acceleration of the discussed discovery pipeline. In this sense, we leverage the potential of machine learning tools, combined with the availability of chemistry databases, to develop models that can predict material properties based on the molecular descriptors introduced in this work. In particular, we introduce several supervised learning models, such as random forest models, alongside unsupervised learning approaches to cluster together porous molecules with similar properties (e.g. clustering, PCA), to improve discovery either via candidate selection for crystal structure prediction or by identification of similar molecular candidates that lead to porous materials with shared properties.
Results and discussion The work developed in this thesis demonstrated that automated molecular porosity analysis aids several key aspects of the porous molecular materials discovery pipeline. For example, by screening of chemical databases (such as PubChem or the Cambridge Structural Database) we discovered porous molecules that had never been considered in this context. Another application of molecular porosity algorithms is the study of molecular interlocking in solid phase (e.g. crystal structures involving several molecules in a periodic box), facilitating CSP and amorphous structure prediction by decreasing the number of possible initial configurations. Moreover, with help of newly trained random forest models, our pipeline shall overcome the need of performing CSP, or, at the very least, bring data-driven guidelines to select the most promising candidates to be either studied via this method or experimentally synthesized, saving time and resources for the scientists.
This PhD thesis contributes with the community of porous molecular materials researchers in three forms. First, it introduces definitions and algorithms to study molecular properties (mainly porosity, but also molecular shape), bringing cheminformatics tools for automatic calculation of molecular descriptors. Second, it presents novel materials aided by the tools and definitions previously introduced, also demonstrating their applicability. More particularly, we demonstrate the discovery of several novel porous molecular materials throughout different studies. Third, it introduces a computational toolbox and online resources readily available for their use by researchers that aim to take advantage of its results.
We introduce the following contributions in form of applications in the field of porous molecular materials:
• Twenty novel porous molecular materials: We introduced a total of 20 computationally predicted novel PMMs, 6 of which came from direct screening, 10 more obtained via in silico automated design combined with candidate analysis, another 3 coming from a combined screening/rational modification of candidates and one more based on a rationally designed, large imine cage which is also demonstrated to serve for water desalination purposes.
• Two repositories of porous molecules: Our efforts allowed us to introduce the largest existing repository of porous molecules, consisting on 6020 molecules extracted from PubChem plus another 17832 molecules extracted from CSD.
Finally, we introduce the following online resources, available to be used by the scientific community. In particular:
• A chemoinformatics C/C++ software tool, named Molipor, available online (http://www.nanoporousmaterials.org/programs/) ready to be used for characterization of molecular porosity in various contexts.
• A family of random forest models, available online, for improved prediction of material properties based on molecular descriptors (computable with Molipor). These are provided alongside the scripts to compute them with R.
• A library of functions to compute shape descriptors for belt-like molecules, available online.
On the molecular characterization side, we introduce seven molecular descriptors for molecular porosity, namely:
• Pore Exposure Ratio: This novel definition provides a measure of how exposed a given point is to its surroundings. The algorithm studies relative position of atoms and bonds, assigning a value between 0 and 1 (the closer to 0, the more surrounded it is). With the help of this value and a threshold (empirically selected), we can decide if a point is inside the molecule.
• Molecular cavity characterization: Via an algorithm that combines PER and Voronoi tessellation, we provide a list of relevant molecular cavities, along with their size (i.e., distance from the center of the cavity to closest atom).
• Molecular windows definition: Definition of what a chemical window is, understood as a set of atoms and bonds that separate inner and outer space. Along with the definition, an algorithm to discover chemical windows is presented.
• Entry paths: Trajectory of a probe molecule that enters or leaves a cavity of the molecule. An algorithm, based on Voronoi tessellation and cavity detection, is provided along with the definition. Entry paths have associated the distance to closest atom, which determines the “entry size”. The largest among the entry sizes defines the Maximum Window Size descriptor. The Number of Entry Paths is another descriptor obtained here.
• Internal surface area: Based on the PER definition, provides information on the amount of molecular surface that is oriented towards the inside. A Monte Carlo approach is provided to compute this value.
Characterization of molecular porosity is a challenging task. In our model, atoms are solid spheres, and bonds are segments connecting atoms’ centers. Thus a molecule is considered a 3-dimensional object that may have voids between the regions of space occupied by atoms. It is precisely the presence of these voids (and how are they surrounded) what determines whether a molecule can is porous, and thus, identification of such regions is our main interest. The space within the molecule would be empty, but somehow surrounded by bonds and atoms. The founding of this PhD work is on detecting this scenario, introducing definitions (and corresponding algorithms) that formalize this concept and resolve boundary cases. Classical mathematical theories have a lot of tools to deal with the idea of a point being interior to an object. Such tools can operate in very arbitrary spaces, but don't adjust well to this problem. The reason for this is that, even though we look for 'internal space', that space is, in the set-theory sense, 'outside' the object of our interest, as it does not intersect with atoms. This constitutes the main challenge in this study: finding a criterion that tells us which part of the space, not being part of the molecule, is trapped by it. To turn this idea into a working method, we exploit atom positions and bonding information, analyzing how would the surroundings of the point would look if observed from that position. From an actual cavity within the molecule, this vision should be restricted to one or several “patches” of space, which should not be too large. On the other hand, from outside the molecule, the surroundings should be easily visible. This intuition is what led us to the definition of Pore Exposure Ratio. This is a completely novel technique, which addresses two difficult challenges as finding the space trapped by a solid object and detecting windows to that space. This technique could be extended, potentially, to many other structure-related problems, ubiquitous in chemistry and biochemistry. We hope to explore that in the future.
Conclusions This PhD work has been developed with the goal of constructing a pipeline for the discovery of porous molecular materials. The working hypothesis is the idea that molecular property (i.e. molecular porosity in all its variants) plays a major role in material properties in terms of porosity, and thus the discovery of novel porous molecular materials can be guided solely by the characterization of their building blocks. Throughout this work, we’ve demonstrated this hypothesis to be true, while introducing novel materials obtained via different ways, and establishing computational tools such as the Molipor software, which is available online, and a variety of random forest models that allow for the identification of material properties based on those of the molecules that conform them. We expect this work to serve as the foundation of valuable findings in the field of materials science, both leading the way in the discovery of novel porous molecular materials, and serving as a reference for similar strategies in other similar areas of interest in materials science.
