Fig. 5
From: Massive Atomic Diversity: a compact universal dataset for atomistic machine learning
Two-dimensional projections of 1,000 representative MAD structures, selected via farthest-point sampling, based on high-dimensional features from the PET-MAD model. (A) PCA, (B) UMAP, (C) t-SNE, (D) Sketch-map projection, and (E) MLP-predicted sketch-map projection, learned to map high-dimensional descriptors to 2D space. The axes labels indicate the method used to determine the low-dimensional latent space in each map.
