Information-theoretic approach to density-based machine learning.

Density-based machine learning encompasses several distinct tasks: predicting an electron density or Hamiltonian, learning Kohn-Sham (KS) exchange-correlation or orbital-free kinetic-energy functionals, and using a computed density to predict a higher-level molecular property. This Perspective focuses on the last route and asks when quantities from the information-theoretic approach (ITA) provide a useful, physically interpretable compression of the electron density. We present the established Shannon- and Fisher-type density functionals and relative information measures adopted in the literature, clarify their electron-density normalization and associated conventions, and place ITA models within a broader taxonomy of electronic-structure ML. Employing the correlation energy and molecular polarizability as illustrative targets, we examine when ITA descriptors may retain useful density information and identify the baselines and out-of-domain tests needed to evaluate their transferability. We then compare ITA density-to-property learning with direct density and Hamiltonian prediction, learned density functionals, and geometry-based machine-learning interatomic potentials. The proposed ITA-ML workflow is neither a replacement for density functional theory (DFT) nor a universal density functional. It is a task-specific strategy that exchanges the cost of a low-level density calculation for compact features sensitive to charge, spin, and electronic state. Its principal open questions concern information loss, transferability, total computational cost, uncertainty, and compatibility with variational constraints when information-theoretic quantities are incorporated into self-consistent models.