Automated Wannierizations

Maximally localized Wannier functions (MLWFs) are widely used in computational condensed matter physics. The standard approach to construct MLWFs often requires initial guesses which are based on chemical intuition and some measure of trial and error. Here, we first introduce an algorithm based on "projectability disentanglement" that provides reliably and automatically atom-centered Wannier functions describing both occupied and empty states. Then, we show how to mix these again into target subspaces; e.g., to describe valence and conduction bands separately. We test these algorithms on 200 representative materials (77 insulators), showing that the final MLWFs are very well localized and can accurately interpolate band structures at the meV scale. Such approaches enable automated Wannierizations for both metals and insulators, promoting further applications of MLWFs in physical applications and high-throughput calculations.