Efficient Computation of Off-Diagonal Elements of the Full-Frequency GW Self Energy in BerkeleyGW

The GW approximation is a workhorse of many-body perturbation techniques, and correctly captures dispersion and non-bonding interactions that are missed by DFT. These interactions may be particularly important for understanding the transition states and energetics of reactions on solvated surfaces where specific catalytic configurations contribute significantly to reactivity. Generally the off-diagonal matrix elements in the self energy are ignored and only the diagonal quasiparticle energy shifts are considered, due to the significant additional computational expense involved. These off-diagonal matrix elements, however, allow for the calculation of molecular forces and the use of self-consistent techniques at the GW level of theory.

The BerkeleyGW software suite implements GW calculations using a plane wave basis in a massively parallel, scalable, and portable manner, but lacks an efficient implementation to calculate off-diagonal elements of the self-energy, epseically within the full-frequency GW formalism. Here we report on advances in fully utilizing modern HPC resources, especially the capabilities of GPU-accelerated systems, to make calculating off-diagonal elements of the self-energy within BerkeleyGW feasible. We further explore how this paves the way to advance computational material science beyond the level of DFT.