Learning Hidden Chemistry with Deep Neural Networks

Abstract: We demonstrate a machine learning approach designed to extract hidden chemistry/physics to facilitate new materials discovery. In particular, we propose a novel method for learning latent knowledge from material structure data in which machine learning models are developed to present the possibility that an atom can be paired with a chemical environment in an observed materials. For this purpose, we trained deep neural networks acquiring information from the atom of interest and its environment to estimate the possibility. The models were then used to establish recommendation systems, which can suggest a list of atoms for an environment within a structure. The center atom of that environment was then replaced with the various recommended atoms to generate new structures. Based on these recommendations, we also propose a method of dissimilarity measurement between the atoms and, through hierarchical cluster analysis and visualization using the multidimensional scaling algorithm, illustrate that this dissimilarity can reflect the chemistry of the elements. Finally, our models were applied to the discovery of new structures in the well-known magnetic material Nd$2$Fe${14}$B. Our models propose 108 new structures, 71 of which are confirmed to converge to local-minimum-energy structures with formation energy less than 0.1 eV by first-principles calculations.