Metal-organic frameworks (MOFs) are crystalline materials, made from metal clusters tethered together by organic linking molecules, that exhibit nano-sized pores connected by channels. Gas molecules selectively adsorb into these sponge-like materials allowing MOFs to efficiently separate or purify mixtures of gases encountered in industry. Choosing different metals or linking molecules gives hundreds of thousands of possible MOFs combinations. How does one find the right match between a desired molecular separation and the type of MOF to do the job? This project aims to develop a computational approach that uses images of the 3D pore structures of hundreds of thousands of MOFs to quantitatively describe the pores inside MOFs. Serving as a fingerprint of pore shape, these descriptions are useful for (i) grouping together MOFs that have similar pore structures and (ii) predicting the ability of a MOF to separate chemically similar molecules effectively. By computationally finding the right MOF for the job instead of experimentally testing them all (a time- and resource-intensive process), this proposed research is expected to accelerate the deployment of MOFs for gas separations, thereby saving energy, making goods cheaper to produce, and reducing pollution. The investigators will also mentor high school and undergraduate minority students in summer research programs at Oregon State University to encourage these students in computational science.
Nanoporous materials, such as metal-organic frameworks (MOFs), have the ability to selectively adsorb gas molecules which make them industrially relevant to gas separations. Owing to their modular synthesis, MOFs are highly tunable and offer a vast chemical space in which to optimize gas separation properties. Because the pore geometry within a MOF strongly influences its adsorptive selectivity, mathematical characterization will greatly improve efficiency in investigating MOF structures. This project will design a deep convolutional autoencoder to encode the pore structures of MOFs into low-dimensional, information-rich, latent vector representations, i.e. fingerprints. Three-dimensional, computational images of the porosity of over 200,000 MOFs will serve as training data for the autoencoder. This research will determine the optimal architecture, regularization scheme, and loss function to coax the autoencoder to extract the salient features of pore structures in MOFs in a translation- and rotation-invariant manner. The machine-learned fingerprints of MOFs will be practically useful for (i) building a classification or regression model of adsorptive selectivity and (ii) by defining a meaningful distance metric, identifying candidates exhibiting the most similar pore structures to a promising but suboptimal "lead" material to fine-tune its adsorptive selectivity in a "lead optimization" strategy for materials discovery. To demonstrate the practical utility of the latent vector representations, the predictiveness of simulated adsorptive selectivity for several industrially-relevant separations will be assessed. Deployment of the trained autoencoder to encode the salient features of pore structures of MOFs into fingerprints will accelerate the discovery and deployment of nanoporous materials that display the optimal pore structure to target the adsorption of a particular gas species, thereby enabling separation.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
