Practicing deep learning in materials science: An evaluation for predicting the formation energies

Deep learning in recent years has entered the chemistry and materials research arsenal with many successful accomplishments in tasks considered to be intractable using traditional means. However, the widespread application of this data-driven technology is still challenged by the requirement of large training data, poor model interpretability, and hard-to-detect errors that undermine the soundness of conclusion. Here, we performed a systematic study for the modeling of the formation energies of inorganic compounds using deep learning. Our results proved the advantage of deep learning methods over several non-deep learning methods in this specific task and demonstrated the abstraction of knowledge using deep learning, which was a unique ability compared to non-deep learning methods. Several aspects that critically affected the conclusion were also highlighted, including the importance to rigorously compare model performance with the same dataset, the design of input representation, and the careful selection of model architecture. Findings from the current study demonstrate the capabilities of deep learning solving complicated problems in materials research and serve as new guidelines for future practicing of deep learning in this field. Published under license by AIP Publishing.