Learnable features for predicting properties of metal-organic frameworks with deep neural networks

Materials science is being rapidly transformed by machine learning tools. This paper introduces a machine learning approach for predicting energy and other derived properties in metal-organic frameworks (MOFs). Using neural networks, our approach generates embedding characteristics for both local atomic structures and the overall MOF system by extracting hidden representations of pairwise interactions among atoms inside MOFs. These networks are trained using total energies derived from density functional theory computations, and they are shared for all paired terms. The model performs better than others in terms of total energy prediction, with a mean absolute error of about 0.09 eV/atom. Furthermore, we demonstrate the transferability of the learned features to accurately predict band gaps. t-Distributed stochastic neighbor embedding is utilized to gain insights into the meaningful patterns within the MOF space, while a K-means clustering model is carried out to detect distinct groups of MOFs.