Fig. 3: VAE reconstruction error of phases and phase mixtures.
From: Deep learning for visualization and novelty detection in large X-ray diffraction datasets
a Violin plot showing the statistics of the reconstruction error for pure phases, two-phase mixtures and an unknown pure structure. The median of phase mixtures is significantly higher than for pure phases. The median of the unknown phase shows a one order of magnitude increase in reconstruction error with respect to pure phases. Pure phases show a multimodal distribution, phase mixtures a bimodal distribution and the unknown phase a normal distribution. A broader distribution is observed for phase mixtures and the unknown phase. b Reconstruction error versus mixing ratio of binary mixtures. The reconstruction error shows a maximum at appr. 50% mixture for which the VAE has the highest uncertainty. We suggest that the reconstruction error be used as a metric for uncertainty that indicates XRD patterns that are either mixtures of phases or new phases that are not contained in the training data. The reconstruction error could be used to guide a data-driven acquisition function in the search for single phase regions in large chemical composition spaces and the discovery of new materials.
