Eigenvalue decomposition of spectral features in density of states curves

The purpose of this paper is to show a computational strategy based on pattern recognition methods to extract features from density of states (DOS) curves that can be traced back to crystal chemistry and structure. We show how the knowledge base developed by eigenvalue decomposition of DOS curves derived from accurate DFT calculations for a few simple metals can serve as a training set for computing DOS spectra of other simple metals without requiring a separate set of DFT calculations. The comparison of these derived DOS curves are shown to agree with DFT calculations. The implications of this data-driven approach to model DOS curves of new materialsas a way to generateapproximations of DOS spectra prior to full scale DFT calculations in a high throughput fashion are also discussed.