Two-component $GW$ calculations: Cubic scaling implementation and comparison of vertex corrected and partially self-consistent $GW$ variants

We report an all-electron, atomic orbital (AO) based, two-component (2C) implementation of the $GW$ approximation (GWA) for closed-shell molecules. Our algorithm is based on the space-time formulation of the GWA and uses analytical continuation of the self-energy, and pair-atomic density fitting (PADF) to switch between AO and auxiliary basis. By calculating the dynamical contribution to the $GW$ self-energy at a quasi-one-component level, our 2C $GW$ algorithm is only about a factor of two to three slower than in the scalar relativistic case. Additionally, we present a 2C implementation of the simplest vertex correction to the self-energy, the statically screened $G3W2$ correction. Comparison of first ionization potentials of a set of 67 molecules with heavy elements (a subset of the SOC81 set) calculated with our implementation against results from the WEST code reveals mean absolute deviations of around 70 meV for $G_0W_0$@PBE and $G_0W_0$@PBE0. These are most likely due to technical differences in both implementations, most notably the use of different basis sets, pseudopotential approximations, different treatment of the frequency dependency of the self-energy and the choice of the 2C-Hamiltonian. Finally, we assess the performance of some (partially self-consistent) variants of the GWA for the calculation of first IPs by comparison to vertical experimental reference values. $G_0W_0$PBE0 (25 \% exact exchange) and $G_0W_0$BHLYP (50 \% exact exchange) perform best with mean absolute deviations (MAD) of about 200 meV. Eigenvalue-only self-consistent $GW$ (ev$GW$) and quasi-particle self-consistent $GW$ (qs$GW$) significantly overestimate the IPs. Perturbative $G3W2$ corrections improve the agreement with experiment in cases where $G_0W_0$ alone underestimates the IPs. With a MAD of only 140 meV, 2C-$G_0W_0$PBE0 + $G3W2$ is in best agreement with the experimental reference values.
Version with recalculated results due to a bug in the code with which the original results where produced