Machine learning accelerated high-throughput screening of zeolites for the selective adsorption of xylene isomers.

Abstract

The production of widely used polymers such as polyester currently relies upon the chemical separation of and transformation of xylene isomers. The least valuable but most prevalent isomer is meta-xylene which can be selectively transformed into the more useful and expensive para-xylene isomer using a zeolite catalyst but at a high energy cost. In this work, high-throughput screening of existing and hypothetical zeolite databases containing more than two million structures was performed, using a combination of classical simulation and deep neural network methods to identify promising materials for selective adsorption of meta-xylene. Novel anomaly detection techniques were applied to the heavily biased classification task of identifying structures with a selectivity greater than that of the best performing existing zeolite, ZSM-5 (MFI topology). Eight hypothetical zeolite topologies are found to be several orders of magnitude more selective towards meta-xylene than ZSM-5 which may provide an impetus for synthetic efforts to realise these promising materials. Moreover, the leading hypothetical frameworks identified from the screening procedure require a markedly lower operating temperature to achieve the diffusion seen in existing materials, suggesting significant energetic savings if the frameworks can be realised.