Extended Data Fig. 1: Representational Variability in Coal Plant Modeling via UMAP and Mapper.
From: Strategies to accelerate US coal power phase-out using contextual retirement vulnerabilities
(a) Each panel displays a low-dimensional embedding of coal plants using UMAP, with colors indicating cluster assignments via HDBSCAN. Varying locality parameters (for example, number of neighbors, minimum distance) leads to substantial differences in structure, despite identical input data. (b) Mapper graphs built on a fixed UMAP projection (with nneighbors = 16, min dist = 0.5, seed = 42) vary across cube counts (ncubes ∈ {5, 7, 10}) and percent overlaps (overlap ∈ {0.55, 0.6, 0.65}), illustrating how topology is sensitive to Mapper’s lensing. Nodes are colored by percent of plants with retirement plans. This variability motivates the use of THEMA to evaluate entire distributions of representations and automatically select models with the highest policy relevance.
