In recent years, GW methods gained popularity in the simulation of charged excitations in molecular systems with applications to, for instance, dye-sensitized solar cells or band gaps of materials. This rising popularity is due to GW being more accurate than Density Functional Theory while being more efficient than Wavefunction methods. At the same time, AI has seen disruptive development in the natural sciences with a rapidly increasing number of sophisticated machine learning models being crafted for and applied to molecular modelling. With the recent release of massive databases containing a vast array of molecules and their associated quantum chemical properties, the field is on the verge of groundbreaking innovations. In line with those developments, I will demonstrate how in our research we merge those two worlds. I will show how we computed a GW database of unprecedented data quantity and quality and how we leverage it to develop machine learning models with the ultimate goal of enabling highly efficient GW calculations. I will discuss how the huge potential of massive pretraining seen in large language modelling translates to machine learning for GW even across levels of theory, how it helps to obtain more robust and more generalizable models, and potential challenges
Dario Baum