2025 AIChE Annual Meeting

Construction of Chemistry-Independent Representations for Gas Adsorption Prediction in Metal-Organic Frameworks

Metal-organic frameworks (MOFs), crystalline networks of metal ions and organic linkers, have garnered significant attention due to their customizability and vast structural diversity. Because MOFs can be highly selective for specific gases, MOF adsorption presents a wide variety of potential uses in industrial chemical processes such as gas separation and storage. As the interactions that determine adsorption behavior are complex and vary widely across frameworks, screening MOFs for adsorption properties requires resource-intensive computational simulation for each new adsorbate-MOF pairing. Machine learning (ML) prediction methods offer more cost-effective alternatives to screen the vast computational space of adsorbate-MOF combinations. However, current ML methods rely on intricate structural representations and chemistry-based descriptors to train ML models that yield accurate predictions, which can lead to a lack of generalizability for frameworks outside of the studied set of structures. Thus, developing representations based on MOF-independent physical descriptors has the potential to greatly increase the generalizability of ML models.

In this study, we aim to evaluate the prediction of gas-MOF interaction properties using a representation built on non-bonding interactions to describe a given MOF’s structure. Our proposed MOF representation is constructed by superimposing a high-resolution grid on a given MOF and locating atomic species that are closest to each grid point. The physical parameters for the non-bonded interactions of these atoms are then binned by proximity and compressed into a compact array. We evaluate the performance of this representation by reducing 2,938 MOF structures to our compressed array format and feeding the resulting data into various ML models. These models are trained to predict adsorption loadings for 2,3-DMB, CO2, H2, and NH3 at different industrially relevant thermodynamic conditions. Representation performance is evaluated using test predictions from the resulting trained models. We also train an ML model to predict multiple adsorption loadings simultaneously with the goal of evaluating the representation’s generalizability to other adsorbates and conditions.

Molecule adsorption loadings for the MOF structures are predicted with acceptable accuracy for most targets. For charged adsorbates such as CO2, the models exhibited somewhat reduced prediction accuracy. Models trained on multiple adsorbates and conditions had prediction accuracies comparable to those of models trained for single adsorbates. The developed ML models demonstrate potential for high performance and generalizability with low resource investment, which is important for the development of MOFs in industrial gas adsorption processes. Our work presents an easily accessible and robust pipeline for processing and evaluating MOF adsorption capabilities. In addition, the representation we have developed provides insight into how a MOF’s physical structure affects its interaction with adsorbing molecules. We hope our representation and models will prove useful in future efforts to screen and select MOFs for industrial separation and storage processes.