Density functional theory is the workhorse of computational materials science, and its cost is the main brake on how many candidate materials a lab can screen. A paper in Nature Machine Intelligence by Xiwen Li, Yang Zhong, Xingao Gong and colleagues proposes a way round much of that cost: predict a material's electron charge density directly with a neural network, then use it to compute the electronic structure without running the slow self-consistent loop.
The model, the Equivariant Density Graph Neural Network or EdenGNN, predicts both the charge density and the augmentation occupancies needed by the projector augmented-wave method, which is what lets it reach the accuracy of standard plane-wave calculations. Two design choices drive the speed. It expands the density in a basis with fully trainable radial functions, and it uses delta learning: rather than predicting the whole density, it learns the difference between the real density and a simple superposition of atomic densities, which mostly captures how charge moves between atoms. The authors report a speed-up of more than an order of magnitude over earlier general-purpose charge density models.
The headline result is a universal model, EdenGNN-Uni, trained on the Materials Project database. It covers non-magnetic structures across the periodic table, and the authors say it predicts band structures accurately for most materials it has not seen, including test structures drawn from Google DeepMind's GNoME database, and more robustly than universal models that predict the Hamiltonian instead. Code, weights and data are public on GitHub, Hugging Face and Zenodo, with patches for OpenMX and ABACUS and instructions for VASP.
This sits beside the better-known machine-learned interatomic potentials, which predict energies and forces rather than the electron density itself. Charge density is a more fundamental quantity, so a good density model can feed many downstream properties at once.
The limits are real. The universal model is restricted to non-magnetic materials, which rules out a large share of systems of practical interest. The authors also warn that a trained model is tied to the pseudopotentials, energy cutoff and other settings used to make its training data, so results only carry over to workflows that match them. We could read only the abstract and supporting material, not the full paper, so the size of the accuracy gains on specific material classes remains to be checked.