Deep Learning in Classifying Structures for Crystal Systems of Pure Metals

抄録

We used two deep learning methods, convolutional neural networks (CNN) and deep neural networks (DNN), to classify three common metal crystal structures (FCC, BCC, and HCP). The training, validation, and test datasets were created by Atomsk and Python scripts, and the data structure was transformed to meet the input requirements of CNN and DNN. To fully train and test CNN and DNN, we constructed four crystal structure datasets using random parameters. The results show that the accuracy of CNN and DNN algorithms on the test set is 100%, indicating that deep learning methods are effective for metal crystal structure classification. Compared with DNN, CNN has fewer parameters, faster training, and faster classification. It lays the foundation for further studying alloy structure detection and phase transition.