» Back to overview
Proposing Institution

Laboratory of Molecular Simualtions, Institute of Chemical Science and Engineering, Ecole Polytechnique Fédérale de Lausanne
Project Manager

Prof. Berend Smit
Rue de l'Industrie 17
CH-1951 Sion
Metal-organic frameworks (MOFs) have in the last decade received much attention for green energy related applications, for example, capturing carbon dioxide gas from coal-fired power plant exhaust. However, the behavior of water in a MOF is often conclusive for any given application it has been designed for. Consequently, understanding, characterizing and predicting how water is adsorbed on the diverse chemical and topological pore spaces typifying MOFs is the key challenge. This knowledge will be essential when identifying and/or designing a MOF for a certain application where humidity is a critical factor.We take on this challenge by proposing to develop a Materials Genomic approach to address the key question at study: Which chemical and topological properties in MOF pores diminish water uptake and which enhance it? How can we combine the metals and linkers to obstruct a material’s affinity to water or steer it to a certain level?However, with the nearly limitless potential for unique MOF materials, and the millions of hypothetical materials already at our disposal, an enormous amount of CPU time would be necessary to perform a brute force computational screening of simulated water in these structures. A new strategy must therefore be employed.Here, we propose a Machine-learning approach to screening the Materials genome. This novel approach will allow us to uncover essential characteristics that govern MOF-water interactions in the context of structure-activity relationships (SAR). State of the art in-house descriptors that uniquely classify porous materials, that, when trained with a deep neural network will enable the prediction of complex water behavior within these materials. It is anticipated that this will unlock a number of hitherto undiscovered similarities between structures, and will be a key parameter in our machine-learning program. The resulting model, trained from water data computed in tens of thousands of structures, will then be validated and applied to the millions of structures in the Materials Genome database. This will be at only a fraction of the cost associated with brute-force screening.Not only will this approach allow us to screen millions of MOFs for performance descriptors, it will also reveal important SARs greatly adding to our fundamental understanding of water adsorption in porous media. Such knowledge will allow us to postulate much needed guidelines on how to tailor a MOF with hydrophobic(/-philic) properties optimized for any given application.

Impressum, Conny Wendler