Application of deep transfer learning to predicting crystal structures of inorganic substances

Convolutional neural network (CNN) consists of shallow learning machine and automatic feature extractor. The feature extractor of a well-trained CNN on a big dataset can be reused in related tasks with small datasets. This technique is called deep transfer learning which not only bypasses manual feature engineering but also improves the generalization of new models. In this study, we attempted to predict crystal structures of inorganic substances, a challenge for material science, with CNN and transfer learning. CNNs were trained on a big dataset of 228 k compounds from open quantum materials database (OQMD). The feature extractors of the well-trained CNNs were reused for extracting features on a phase prototypes dataset (containing 17 k inorganic substances and involving 170 crystal structures) and two high-entropy alloy datasets. The extracted features were then fed into random forest classifier as input. High classification accuracy (above 0.9) was achieved in three datasets. The visualization of the extracted features proved the effectiveness of the transferable feature extractors. This method can be easily adopted in quickly building machine learning models of good performance without resorting to time-consuming manual feature engineering routes.