Mapping chemical composition space onto phase-diagram-like geometries via principal component analysis

In materials science, the means to survey the vast body of previously reported experimental data across an entire field have remained limited. Here, using tens of thousands of thermoelectric material entries collected in Starrydata, we establish a method for visualizing chemical composition by principal component analysis (PCA). When PCA is applied to composition vectors normalized so that their components sum to unity, the distribution of compositions appears as geometric shapes resembling phase diagrams and equilibrium diagrams, and these shapes change systematically into triangles, quadrilaterals, prisms, and related forms according to the degrees of freedom permitted in site substitution. We developed Starrydata Auto-Explorer, a program that generates these composition maps together with property scatter plots for many material systems in a single batch, enabling the entire field to be surveyed as images. In addition, analysis of the large-scale dataset revealed a broad empirical tendency across chemically diverse thermoelectric materials, in which high-ZT samples are predominantly found around Seebeck coefficient S ≈±200–300 µV/K, and electrical conductivity σ ≈104–105 S/m.