Machine Learning Prediction on Properties of Nanoporous Materials Utilizing Pore Geometry Barcodes

In this work, we propose a computational framework for machine learning prediction on structural and performance properties of nanoporous materials for methane storage application. For our machine learning prediction, two descriptors based on pore geometry barcodes were developed; one descriptor is a set of distances from a structure to the most diverse set in barcode space, and the second descriptor extracts and uses the most important features from the barcodes. First, to identify the optimal condition for machine learning prediction, the effects of training set preparation method, training set size, and machine learning models were investigated. Our analysis showed that kernel ridge regression provides the highest prediction accuracy, and randomly selected 5% structures of the entire set would work well as a training set. Our results showed that both descriptors accurately predicted performance and even structural properties of zeolites. Furthermore, we demonstrated that our approach predicts accurately properties of metal-organic frameworks, which might indicate the possibility of this approach to be easily applied to predict the properties of other types of nanoporous materials.