A universal model for accurately predicting the formation energy of inorganic compounds

Harnessing recent advances in data science and materials engineering, it is feasible today to build reliable models for predicting materials properties. Here we employ a comprehensive dataset of 170,714 inorganic crystalline compounds obtained from high-throughput accurate quantum mechanics calculations, to train a machine learning model for the precise prediction of the formation energy of inorganic compounds. Distinct from previous studies, our model can be universally applied to a large phase space of inorganic materials as all the data is utilized for the training, and the model reaches a fairly good predictive ability (R-2 = 0.982 and mean absolute error = 0.072 eV atom(-1), DenseNet model). The improvement comes from several effective structure-dependent descriptors, which are carefully designed to take into account the information of the electronegativity difference between neighboring atoms and local atomic structure. This model provides a useful tool to predict the energy landscape of the compound systems in a fast and cost-effective manner.