Accelerating first-principles GW calculations using deep learning

The GW method has long been the state-of-the-art in the ab initio calculation of the properties of quasiparticles of many materials; but it is computationally intensive for very large systems. Here, to solve this problem, we integrate the GW method with the DeepH and DeepH-E3 frameworks [1-3], which have been powerful tools for deep learning electronic structure calculations. We develop a deep learning framework that takes the atomic structure as input and can be used to predict GW quasiparticle band structures. The key feature of this framework is the generalizability of the neural network models, which allows them to learn from GW calculation data on small systems and accurately predict on much larger ones, enabling first-principles study of large-scale material structures at the GW level with order-of-magnitudes reduction in computational cost. One particularly useful application of this deep learning framework is to study electron-phonon interactions at the GW level, which can be prohibitive if using the traditional frozen-phonon method with supercells.