Research highlight

Inverting the Kohn–Sham Equations with Physics-Informed Machine Learning

Physics-informed neural networks and Fourier neural operators are trained to recover exchange–correlation potentials from electron densities, with the combination of both methods emerging as the most promising route to large-scale inversion.

Kohn–Sham DFT maps interacting electrons onto non-interacting particles that share the same density. The exchange–correlation potential is what keeps the two densities equal, and obtaining that potential directly from a target density — the inverse problem — is a route to understanding which approximations work and why. Neural networks offer a way to learn the density-to-potential mapping.

The authors compare two physics-informed strategies: physics-informed neural networks and Fourier neural operators. They test both on a dataset of one-dimensional atomic and molecular models, discuss the strengths of each, and treat the study as a proof of concept. Their main conclusion is that combining the approaches has the greatest potential for inverting the Kohn–Sham equations at scale.