Abstract— Zeolites are industrially important catalysts and adsorbents, typically synthesized using specific molecules known as organic structure-directing agents (OSDAs). The templating effect of the OSDAs is pivotal in determining the zeolite polymorph formed and its physicochemical properties. However, de novo design of selective OSDAs is challenging because of the diversity and size of the zeolite–OSDA chemical space. Here we present ZeoBind, a computational workflow powered by machine learning that enables an exhaustive exploration of the OSDA space. We design predictive tasks that capture zeolite–molecule matching, train predictive models for these tasks on hundreds of thousands of datapoints and curate a library of 2.3 million synthetically accessible, hypothetical OSDA-like molecules enumerated from commercially available precursors. We use ZeoBind to screen nearly 500 million zeolite–molecule pairs and identified and experimentally validated two new OSDAs that template zeolites with novel compositions. The scale of the OSDA library, along with the open-access tools and data, has the potential to accelerate OSDA design for zeolite synthesis.
Introduction— Zeolites are microporous, crystalline materials that are used in industrial separations1 and catalysis2. While some occur naturally as minerals, the majority are synthesized using hydrothermal processes, often requiring the presence of a highly specific organic structure-directing agent (OSDA) molecule to facilitate the crystallization of specific zeolite frameworks3,4. OSDAs are crucial for selectively driving the synthesis of a particular framework against other polymorphs, and also for influencing the density and distribution of heteroatoms in the framework, such as the density and placement of aluminum in aluminosilicate zeolites5,6, which act as Brønsted acid catalytic sites. The heteroatom distribution then impacts the reactivity, selectivity and stability of zeolites in industrial and environment applications5,7.
 |
Rafael Gomez-Bombarelli
|