Semi-automated estimation of hydrogenic initial states for localized Wannier functions
Tatsuki Oikawa, Kota Ido, Takahiro Misawa, Takashi Koretsune, Kazuyoshi Yoshimi
TL;DR
This work tackles the challenge of generating meaningful initial guesses for Wannier function construction by introducing a semi-automated pipeline that derives hydrogenic projections from Bloch functions at the $\Gamma$-point using a neural-network encoder to predict spherical-harmonic and radial coefficients. The method also employs a clustering-based determination of projection centers, enabling automatic treatment of interstitial and molecular states. Applied to SrVO$_{3}$, FeSe, Si, Na$_8$Al$_6$Si$_6$O$_{24}$, and $(\text{TMTTF})_2\text{PF}_6$, the approach yields Wannier functions with comparable or improved localization and interpretability relative to SCDM, while maintaining accurate band reproduction and reasonable gauge-invariant spreads. Integrated with cif2qewan, this framework supports efficient, interpretable, and potentially high-throughput Wannierization workflows for complex materials.
Abstract
We present a semi-automated method for obtaining an initial estimate of Wannier functions, designed to facilitate the construction of Wannier functions for describing low-energy effective models of solids, particularly those relevant to strongly correlated electron systems. Our approach automatically determines the hydrogenic projections orbitals and the center of the Wannier functions from information on Bloch wavefunctions at the $Γ$ point. This method is integrated into cif2qewan, enabling seamless generation of input files for Quantum ESPRESSO and Wannier90. We validate our method through applications to both inorganic and organic compounds, such as Si, SrVO$_3$, FeSe, Na$_8$Al$_6$Si$_6$O$_{24}$, and (TMTTF)$_2$PF$_6$. The obtained results demonstrate that our semi-automated projections give a good initial estimate of the Wannier functions. We also show the comparisons with other methods for estimating the initial states of the Wannier functions, such as the Selected Columns of the Density Matrix (SCDM). Our methodology shows an efficient way to construct Wannier functions, paving the way for high-throughput calculations in the study of complex materials.
