Speaker
Description
Electronic structure calculations in solids are commonly performed in large plane wave or real space basis sets. These basis sets are compressed and localised around atoms for advanced postprocessing, such as calculations of the Berry curvature or optical properties. For this purpose, most commonly one constructs Maximally Localised Wannier Functions (MLWFs). MLWFs are constructed to preserve the band structure for some low-energy bands exactly, but may not yield accurate results for properties requiring the explicit calculation of the corresponding wave function.
We instead propose a method for a wave function based scheme for Hamiltonian compression. We started from optimised Atomic Orbitals(AOs) and generate optimal Local Orbitals (LOs) by optimising a projectability operator variationally, using algorithmic differentiation. We disallowed mixing between orbitals of different atoms for the optimisation to preserve the localised nature of our orbitals. We included an iterative improvement of a projectability-based disentanglement. This approach requires little information about the system and is therefore compatible with high throughput calculations. It also allows for systematic improvement by increasing the number of LOs.
We also used schemes to improve the interpolation of the band structure in local orbital space, such as a Loewdin correction to the Hamiltonian or a k point dependent mixing of the LOs based on the principle of Intrinsic Atomic Orbitals.
Finally, we showcase the potential of using such a compressed basis set in calculating optical properties using the example of dipole operator calculations.