Preface

There will never be a better video game than Pac-Man.1 Well, yes, this is a matter of opinion, but it suggests a level of creativity that might never be reached by AI (Artificial Intelligence). But, if AI could ever reach a similar level, would we be happy with an AI equaling human's creativity? Limits to AI are necessary. AI is becoming more and more integrated into an ever-increasing number of aspects of human life and ethical concerns must be discussed along with or even prior to these developments. The ethics applied to AI will mirror our own ethical beliefs, hence concerns about AI are also concerns about ourselves.2 Justice and transparency are perennial needs of our society and will be implemented in future developments of AI.3 But power also attracts particular interests, with the risk of a globally biased AI.4

With our task in this book being rather modest with respect to the big trends outlined above, it is nevertheless our duty not to produce useless data as well as to follow recommendations and protocols that will make information and results understandable, reproducible, and widely available.5 AI has made an enormous impact recently in the field of materials chemistry, and in particular nanoporous materials. Books may not contribute actively to impact, but they do provide an invaluable service to the scientific community by presenting a more careful approach to the foundations of our knowledge. This is our task with this book, with the help of the invaluable commitment of all the authors, who analyze how AI in its various flavors is pervading our way to do research and, hopefully, gain knowledge.

1 Molecular Modeling and New Zeolite Structures

The book opens with an account of the expertise collected by Stacey Zones and coworkers at the Chevron zeolite group in the development of new zeolites and the role therein of OSDAs (organic structure directing agents) in combination with other synthesis parameters. For more than 40 years this world leading group have been pioneers which, from the industry, have played an outstanding role in the elucidation and interpretation of the mechanisms of zeolite synthesis. The chapter describes in particular how molecular modeling techniques have provided insights into the working mechanism of OSDAs, and how these insights are becoming instrumental in the search for new zeolites. The role of high-throughput and automation of collecting, validating, and analyzing data in order to further develop machine learning and AI tools is discussed. The ultimate goal is to gain a deeper insight into the zeolite synthesis, as well as to use these tools to open up the large structural space of hypothetical zeolites, paving the way for future prediction of on-demand synthesis of zeolite materials for target applications.

2 De Novo Design of OSDAs

Given the progress in molecular modeling and machine learning techniques to accurately model the templating effect of OSDAs (organic structure directing agents), the second chapter by Frits Daeyaert and Michael Deem describes de novo design of OSDAs as an effective approach to exploit this ability. De novo design originated in drug design as an automated tool to generate molecules interacting with a biological target. In the zeolite-OSDA setting, the aim is to design molecules that closely interact with a target zeolite structure. It is discussed how issues like the synthesizability of the designed compounds and the simultaneous optimization of multiple additional properties and constraints of OSDAs can be addressed. The application to several OSDA design projects of a de novo algorithm developed by the authors is described.

3 ML Search for SDAs Using Molecular Topology and MC Techniques

In Chapter 3, María Gálvez-Llompart and German Sastre present the application of 3D-QSAR algorithms to the modeling and design of SDAs for zeolite synthesis. 3D-QSAR algorithms have proven their merit in drug design. They rely on the principle that the activity of a drug molecule relies on its binding energy with a biological target, and that this binding energy can be modeled using a combination of well-chosen molecular descriptors and appropriate machine learning algorithms. This situation resembles zeolite synthesis with the aid of SDAs, where the interaction energy between an SDA and its target zeolite is a determining factor in the outcome of a zeolite synthesis. As in drug design, accurate prediction of this interaction energy using atomistic modeling is very computationally expensive. It is shown how the use of molecular topology descriptors in combination with appropriate statistical and machine learning techniques allows an orders of magnitude speed-up in predicting zeolite-SDA interaction energies. This in turn enables virtual screening of large numbers of available molecules for use as OSDAs for zeolite synthesis. The method is demonstrated for BEA zeolite and, importantly, uses carefully selected experimental data.

4 Big Data in Zeolite Simulations

Application of data science methods to zeolite discovery has been hindered by the diversity of synthesis routes and conditions and by the absence of machine-readable experimental data. In Chapter 4, Daniel Schwalbe-Koda and Rafael Gómez-Bombarelli discuss how combining simulations and data-driven methods lead to the understanding and prediction of the role of OSDAs in zeolite synthesis. The concept of the FAIR, findable-accessible-interoperable-reusable, principle in the context of computational material databases is discussed. An overview is given of existing zeolite databases of experimental and predicted structures, and of efforts to create curated datasets of zeolite syntheses based upon the FAIR principle. The creation of a web-based platform to interactively explore OSDA-templated synthesis routes by simultaneously mining both experimental and computational data is described. Details of how the computational data have been compiled are discussed. These include efficient algorithms for OSDA-zeolite docking and binding energy calculations, the set-up and analysis of high-throughput computational screening, and consideration of additional OSDA descriptors such as synthetic accessibility, volume, and shape and charge descriptors.

5 Co-Templating for Small-Pore Zeolite Catalysts

Small-pore zeolites, which are of importance for their use in catalytic reactions and selective adsorption of small molecules, often contain more than one type of cavity. Thus, improved synthesis of these materials might be envisaged by the use of more than one template. This is discussed in Chapter 5 by Alessandro Turrina and coworkers forming a solid cluster of collaborators, including several UK universities and the leadership of Johnson Matthey who has worked on zeolite synthesis for more than three decades since the pioneering work by Paul Wright, John Casci, and Paul Cox. Synthesis of nanoporous solids using co-templating, aided by computational simulations, has been described for SAPO zeotypes. For Al-Si zeolites, the situation is complicated by the presence of metal cations that interact with both the crystallizing frameworks and OSDAs that are present in reaction mixtures. In the case of mixed inorganic/organic templating, it is discussed how computational modeling is applied in both synthesis design and structure elucidation. Co-templating has been successfully exploited in the synthesis and structural control of disordered intergrowth zeolite structures containing more than one cavity type.

6 Computer Generation of Hypothetical Zeolites

In Chapter 6, German Sastre and coworkers discuss the algorithmically closely related fields of computer generation of hypothetical zeolite structure prediction (ZSP) and determination (ZSD). Both ZSP and ZSD present complex, combinatorial problems, and one approach to solve these is the use of genetic algorithms. Thus, a separate paragraph is devoted to this subject. An overview of ZSP and ZSD algorithms and their software implementation is given and one particular method, the zeoGAsolver program, is discussed in detail. ZSP and ZSD are very computationally expensive, and therefore can benefit from the advent of high-speed GPUs, requiring efficient parallelization of existing algorithms, as has been implemented by Laurent Baumes and coworkers in the code Parallel Genetic Hybrid Algorithm for Zeolites. Also, the groundbreaking algorithms included in the SCIBS (Symmetry-Constrained Intersite Bonding Search) software are described in detail by Michael Treacy. This method enumerates all possible four-valent networks within each space group given the number of unique tetrahedral atoms. Using this program, in collaboration with other authors, the first database of hypothetical zeolites was created and has been freely available since 2004.

7 Numerical Representations of Chemical Data for Machine Learning

In Chapter 7, Gyoung Na discusses how molecular structures are represented in a way that can be input into machine learning algorithms. In chemistry, a natural representation of molecules is as molecular graphs, which can be input into graph neural networks. The general architecture of graph neural networks is presented, and practical implementations including graph convolutional networks, graph attention networks, message passing neural networks, and crystal graph convolutional neural networks are summarized. Many problems in chemistry involve interactions between multiple molecules, and therefore the adaptation of graph neural networks to these problems is also discussed. While graph-based...