Insights from the reciprocal space revealed by a convolutional neural network and transfer learning

Convolutional Neural Networks (CNNs) have achieved significant success due to the integrated feature engineering within their architecture, but they are often perceived as black boxes. In this letter, we provide a perspective of the transformation process of powder diffraction patterns by a Convolutional Neural Network (CNNs), particularly in the context of estimating lattice parameters from powder diffraction patterns of organic materials. We propose that the convolutional layers resample the data points to segment the diffraction pattern in a set of components defined by the feature maps. We identified that the first convolutional layers were focused on the high-angle regions to remove dissimilarities among diffraction patterns with the expanded segments. In contrast, the final features were more influence by the low-angle regions. We show that the engineered features produced by the convolutional layers are sufficient for effective transfer the learning to inorganic materials, with accuracy enhanced by incorporating crystal system codification.