Predicting Thermal Conductivity of Metal-Organic Frameworks using Graph Convolutional Neural Networks

Metal-Organic Frameworks (MOFs) have gained prominence as revolutionary materials for gas storage and separation applications due to their high porosity, large surface areas, tunable pore geometry and chemistry. However, the practical usefulness of MOFs, especially in gas storage applications, depends on how rapidly they can disperse the tremendous amount of thermal energy generated during the typically exothermic adsorption process. Although adsorption-related properties of MOFs have been extensively investigated, comparatively limited studies have focused so far on understanding their thermal transport properties. To provide a data-driven perspective into this problem, we performed the first high-throughput computational screening of thermal transport properties in MOFs by performing classical molecular dynamics (MD) simulations on a diverse set of hypothetical MOFs. In addition to establishing universal structure-thermal conductivity relationships, we trained a crystal graph convolutional neural network on the thermal conductivities of 5353 hypothetical MOFs that were computationally generated with the ToBaCCo-3.0 code. Our model predicts the thermal conductivity of MOFs in a separate test set with a mean absolute error of 0.072 W m^-1 K^-1. This model only requires MOF atomic and bonding information, which can help accelerate MOF discovery with tailored thermal transport properties by replacing expensive atomistic simulations that require interatomic potentials.