To know how a molecule behaves, chemists need its electron density, the cloud of electrons around the nuclei. The standard way to compute it iterates: guess the density, calculate a field, improve the density, repeat until stable. It is accurate but slow.
The researchers trained machine-learning models, inspired by networks originally built for image analysis, to turn a rough initial guess into a refined electron density in one pass. One model uses attention, letting each region look at others; the other uses cheaper fixed filters. A single final calculation step then gives energies, forces and orbitals.
The best model's densities averaged 0.046% error on a standard benchmark, with a median energy error of 0.036 millielectronvolts, far below the threshold normally called chemical accuracy. The predicted forces also drive realistic molecular motion. This could bring quantum-level simulation costs down substantially.