Accelerating Discovery of Metal–Organic Frameworks for Methane Adsorption with Hierarchical Screening and Deep Learning

In recent years, machine learning
(ML) methods have made significant
progress, and ML models have been adopted in virtually all aspects
of chemistry. In this study, based on the crystal graph convolutional
neural networks algorithm, an end-to-end deep learning model was developed
for predicting the methane adsorption properties of metal–organic
frameworks (MOFs). High-throughput grand canonical Monte Carlo calculations
were carried out on the computation-ready, experimental MOF database,
which contains approximately 11 000 MOFs, to construct the
data set. An area under the curve of 0.930 for the test set proved
the reliability of the developed deep learning model. To assess the
transferability of the model, we applied it to predict the methane
adsorption volume for some randomly selected covalent organic frameworks
and zeolitic imidazolate framework materials. The results indicated
that the model could also be suitable for other porous materials.
We also applied it to the hierarchical screening of a hypothetical
MOFs database (∼330 000 MOFs). Four hypothetical MOFs
were demonstrated to have the highest performance in methane adsorption.
A calculated maximum working capacity of 145 cm3/cm3 at 5–35 bar and 298 K indicated that the hypothetical
MOF is close to the Department of Energy’s 2015 target of 180
cm3/cm3. Further analyses on all screened out
MOFs established correlations between some structural features with
the working capacity. The successful incorporation of ML and hierarchical
screening can accelerate the discovery of new materials not just for
gas adsorption, but also other areas involving interactions in materials
and molecules.