Machine learning material properties from the periodic table using convolutional neural networks

Abstract

In recent years, convolutional neural networks (CNNs) have achieved great success in image recognition and shown powerful feature extraction ability. Here we show that CNNs can learn the inner structure and chemical information in the periodic table. Using the periodic table as representation, and full-Heusler compounds in the Open Quantum Materials Database (OQMD) as training and test samples, a multi-task CNN was trained to output the lattice parameter and enthalpy of formation simultaneously. The mean prediction errors were within DFT precision, and the results were much better than those obtained using only Mendeleev numbers or a random-element-positioning table, indicating that the two-dimensional inner structure of the periodic table was learned by the CNN as useful chemical information. Transfer learning was then utilized by fine-tuning the weights previously initialized on the OQMD training set. Using compounds with formula X2YZ in the Inorganic Crystal Structure Database (ICSD) as a second training set, the stability of full-Heusler compounds was predicted by using the fine-tuned CNN, and tungsten containing compounds were identified as rarely reported but potentially stable compounds.