Preprint reports molecular prediction gains from electron-density constraints

Researchers report in an arXiv preprint that their molecular prediction method surpasses state-of-the-art performance on six out of seven extensive properties in the QM9 dataset. The method uses ground-state electron density as an input and requires that density to minimize a learned mathematical function. The approach also improves generalization to larger molecules in QM40. Across all extensive properties in both datasets, it outperforms density-informed models without those constraints. The method still requires ground-state electron density as an input.