Machine Learning Using Combined Structural and Chemical\nDescriptors for Prediction of Methane Adsorption Performance of Metal\nOrganic Frameworks (MOFs)

Abstract

Using molecular simulation for adsorbent\nscreening is computationally\nexpensive and thus prohibitive to materials discovery. Machine learning\n(ML) algorithms trained on fundamental material properties can potentially\nprovide quick and accurate methods for screening purposes. Prior efforts\nhave focused on structural descriptors for use with ML. In this work,\nthe use of chemical descriptors, in addition to structural descriptors,\nwas introduced for adsorption analysis. Evaluation of structural and\nchemical descriptors coupled with various ML algorithms, including\ndecision tree, Poisson regression, support vector machine and random\nforest, were carried out to predict methane uptake on hypothetical\nmetal organic frameworks. To highlight their predictive capabilities,\nML models were trained on 8% of a data set consisting of 130,398 MOFs\nand then tested on the remaining 92% to predict methane adsorption\ncapacities. When structural and chemical descriptors were jointly\nused as ML input, the random forest model with 10-fold cross validation\nproved to be superior to the other ML approaches, with an <i>R</i>² of 0.98 and a mean absolute percent error\nof about 7%. The training and prediction using the random forest algorithm\nfor adsorption capacity estimation of all 130,398 MOFs took approximately\n2 h on a single personal computer, several orders of magnitude faster\nthan actual molecular simulations on high-performance computing clusters.