Atom table convolutional neural networks for an accurate prediction of compounds properties

Abstract

Abstract Machine learning techniques are widely used in materials science. However, most of the machine learning models require a lot of prior knowledge to manually construct feature vectors. Here, we develop an atom table convolutional neural networks that only requires the component information to directly learn the experimental properties from the features constructed by itself. For band gap and formation energy prediction, the accuracy of our model exceeds the standard DFT calculations. Besides, through data-enhanced technology, our model not only accurately predicts superconducting transition temperatures, but also distinguishes superconductors and non-superconductors. Utilizing the trained model, we have screened 20 compounds that are potential superconductors with high superconducting transition temperature from the existing database. In addition, from the learned features, we extract the properties of the elements and reproduce the chemical trends. This framework is valuable for high throughput screening and helpful to understand the underlying physics.