Fig. 1: Visualization of the VAE latent space representation.
From: Deep learning for visualization and novelty detection in large X-ray diffraction datasets
a Schematic VAE architecture. XRD patterns are encoded into a low dimensional representation (mean, variance). The red cross marker indicates the latent space position of the encoded XRD pattern with respect to the prior (circles). The latent vector is decoded into a reconstructed XRD pattern. b Latent space embedding with color-coded diffraction angle 2Θ of maximum XRD intensity. c Latent space embedding of an exemplary synthetic dataset. Im\(\overline{3}\)m is clearly separated from Fm\(\overline{3}\)m and P63/mmc. The x-markers show the latent space position of the corresponding XRD patterns in (d). d XRD patterns of the x-marked latent space positions in (c). The latent space embedding in b) clearly shows the main reflection axes of the different crystal structures, i.e. P63/mmc has six main reflection axes in the angular range from 20 to 90°2Θ (for Cu Kα). It further elucidates possible ambiguities between different structures that exhibit a preferred orientation: Fm\(\overline{3}\)m (111) and P63/mmc (002) as well as Im\(\overline{3}\)m (020) and P63/mmc (102) have peaks at a similar diffraction angle. This is important during the pattern matching task, as an experimental XRD pattern could be a result of either structure.
