Title:
Rethinking Exchange-Correlation in Density Functional Theory: From Limited Data to Transferable Functionals
Abstract:
Density functional theory (DFT) relies critically on exchange–correlation (XC) functionals, whose exact form is unknown. In practice, approximations are constructed within a hierarchical framework—Jacob’s ladder—designed to incorporate known physical constraints and progressively improve accuracy. In this context, machine learning offers new opportunities, but also raises a central question: how can we learn functionals that respect physical structure while remaining reliable beyond the data used to train them?In this talk, I will present a machine-learning framework that approaches XC construction as the learning ofphysical functionals under limited and structured data. I will describe a “cloning” strategy, in which neural networks reproduce established functionals, providing physically grounded initializations that ensure stable self-consistent behavior. Building on this, I will discuss the role of constraint enforcement and its impact on generalization across systems, from molecules to solids, and across different electronic structure codes.Finally, I will highlight recent developments incorporating more nonlocal descriptors and self-attention mechanisms to capture richer electronic environments. These results point toward a path for constructing transferable XC functionals that combine data-driven flexibility with physical constraints.