Crystal synthesis prediction using deep learning

Predicting the synthesizability of new materials has proven to be challenging due to the wide range of parameters that control the synthesis of materials. In this work, we introduce a high throughput deep learning method utilizing Convolutional Neural Networks to predict synthesizable materials as plausible candidates for various applications. To tackle the difficult task of anomaly selection, those hypothetical materials that are less likely to be synthesized, we used a data-driven approach that identifies unobserved crystal structures of the most studied materials in literature as anomalies. We encode the materials from Crystallography Open Database (COD) into three-dimensional images and feed them to a Convolutional Autoencoder in order to extract the hidden information of atomic arrangements. Different binary classifiers are trained on synthesized crystals in COD alongside with the anomalies according to the extracted features in order to provide the best synthesizability predictor. We accomplish a high precision without losing generality by lessening human intervention in extracting the features of synthesizability. Moreover, the capability of our model to extract features out of the three-dimensional digitized crystal images can be borrowed individually as an atomic descriptor in other prospective works.