Autoencoder-based feature space of XRD peak patterns

Recent advances in experiments enable us to obtain high-throughput X-ray diffraction (XRD) data, but their characterization is still far from automation. This is because there is no procedure for automatically identifying relevant peaks to characterize the corresponding structure, i.e., either all or specific peaks are necessary for the characterization. To address this, we establish an autoencoder-based scheme for identifying the peak relevance. We first construct an autoencoder neural network so as to map each of XRD patterns onto each of 2-dimensional points in the feature space. This encoder is then used to test the peak relevance: For a given XRD pattern, we mask its concerned peak and input the masked XRD pattern into the encoder. If the masked point significantly shifts from the original one in the feature space, then the masked peak can be considered relevant for characterizing the XRD pattern. Interestingly, applying this scheme to magnetic alloys, we found a low relevant peak having a significant intensity, which is useful for characterizing the XRD pattern. This finding was heuristically revealed by machine learning, which cannot be easily interpreted from physical viewpoints such as higher-order reflections etc.