Representation Learning of Porous Materials for Zeolite Synthesis

Representation of periodic, porous materials for machine learning remains a challenge due to specificity of prediction tasks and open questions about the type and use of information embedded in the representation. Having a good representation produces better predictions, insight into the underlying application, and better generalizability to new chemical space, but it is not always possible <i>a priori</i> to determine the appropriate representation. Zeolites are a class of such nanoporous materials used heavily in the chemicals industry for catalysis and separations. They are synthesized using molecules (termed organic structure directing agents or OSDAs), which stabilize the crystallization of the zeolite from solution during hydrothermal synthesis.<br/><br/>Previous work has looked at how high throughput docking and computation combined with machine learning can aid screening of a large library of molecules to find suitable OSDAs for targeted zeolites, but the models fail to extend to unseen zeolites. In this work, we investigate the use of Voronoi nodes to characterize the void space within zeolites. We utilize message passing on both the crystal atoms and Voronoi nodes to learn zeolite embeddings, and showcase the models' ability to generalize to unseen zeolites. We also show preliminary work on active learning for intelligently acquiring new samples to update our models in the vast, unseen chemical space of hypothetical zeolite - hypothetical OSDA pairs. These tools hold promise for unlocking otherwise inaccessible synthesis routes to a large pool of hypothetical zeolites and therefore novel chemistry.