Information-theoretic approach to density-based machine learning

Information-theoretic descriptors compress a low-cost electron density into compact, interpretable features for task-specific prediction of higher-level molecular properties, with transferability and information loss treated as testable limits.