Identification of optimal metal-organic frameworks by machine learning: Structure decomposition, feature integration, and predictive modeling

Abstract

A novel integrated machine learning (ML) framework, consisting of structure decomposition, feature integration and predictive modeling, is proposed to correlate MOF structures with gas adsorption capacities. First, metal nodes, organic linkers, and underlying topologies are identified from MOF structures. Numerical features of the organic linker are generated by molecular graph convolution. Later, the metal node is embedded and integrated with organic linker's features to capture the MOF chemical information. In addition, embedded MOF topology and geometric descriptors are considered as additional structure-level features. Finally, using all the chemical and geometric features as inputs, ML models are trained to predict MOF adsorption capacities. Through two case studies on hydrogen storage and ethylene/ethane separation, the proposed ML framework is demonstrated to be reliable and efficient for MOF property modeling. The resulting ML models can provide a fast and reliable prediction of MOF performance indicators, which thereby significantly accelerate the discovery of novel MOFs. The major novelty of the present work is the incorporation of full chemical information of MOF for ML modeling, which largely increases the prediction accuracy of adsorption performance and meanwhile facilitates the subsequent model-based optimal MOF discovery.