Paola Gori-Giorgi
Microsoft Research AI for Science
Abstract
Machine learning offers a new route to one of the central challenges of density functional theory: constructing accurate and broadly applicable exchange-correlation functionals. In this lecture, I will introduce Skala, our machine-learned exchange-correlation functional. I will begin with a pedagogical overview of why learning an XC functional is a particularly challenging problem, from representing the electronic density and incorporating non-local information to satisfying the physical and numerical requirements needed for self-consistent DFT calculations. I will then take a deeper look at the solutions we developed in Skala to address these challenges, including the choices of architecture, training data, physical constraints, and training strategy. The aim is to give both an accessible introduction to machine-learned XC functionals and a practical view of what it takes to turn them into accurate and robust tools for electronic-structure calculations.