Designing a dataset for convolutional neural networks to predict space groups consistent with extinction laws

A dataset of one-dimensional powder diffraction patterns has been designed with a new strategy to train convolutional neural networks for predicting space groups. The diffraction patterns were calculated from lattice parameters and extinction laws, instead of the traditional approach of generating them from a crystallographic database. The new strategy is shown to be more effective than the conventional method. As a result, the model trained on the cubic and tetragonal training sets from the newly designed dataset achieves a prediction accuracy that matches the theoretical maximum calculated on the basis of extinction laws. These results demonstrate that machine-learning-based prediction can be both physically reasonable and reliable. The model trained on this newly designed dataset shows excellent generalization capability, much better than for one trained on a traditionally designed dataset.