Accelerating The Selection of High Uptake Covalent Organic Frameworks with Supervised Machine Learning for Methane Storage

Covalent organic frameworks (COFs) have been identified as ideal candidates for natural gas storage and catalysis applications primarily because of their tunable microstructure as well as the vast design space for their chemical makeup. COFs with high gas uptake values at workable pressures are highly desired for applications such as gas storage, carbon capture and catalysis. Methane (CH₄) storage is highly sought after because of the low carbon-hydrogen ratio, high calorific value and natural abundancy. However, the vast design space for COFs makes the screening of high CH₄ uptake COFs very difficult as it requires time-intensive grand-canonical Monte Carlo (GCMC) simulations to calculate the uptake values. To overcome this challenge, we employ machine learning to develop a model that captures the structure-property relationship in COFs using a large in-silico database of hypothetical COFs. A machine learned regression model based on powerful random forest algorithm is developed that maps the structural and chemical features of the 5000 COFs used in training to their uptake values. The validated model is then used to find the uptakes of COFs in the experimental database eliminating the need of resource intensive GCMC. Thousands of diverse COFs are selected from the database and featurized using MATMINER, a powerful inbuilt python library used in material’s informatics. The features are used to uniquely represent each COF which is mapped to their respective CH₄ uptake values during training of the ML model. The feature space includes the chemical features like elemental property features and structural features like radial distribution functions and density-based structural features. Various dimensionality reduction techniques like feature selection using co-relation analysis and feature extraction using principal component analysis are employed to make an accurate and efficient ML model. We employed this model to predict the CH₄ uptake values for an experimental database of ~ 600 COFs and screened top performers with CH₄ uptake values as high as 1260 molecules/unitcell.