A machine-learning framework predicts electronic structure across length scales, running up to three orders of magnitude faster than DFT and reaching systems where DFT is infeasible.
Almost every material property depends on how electrons are arranged, and DFT has become the standard way to compute that arrangement. Its cost, however, grows quickly with the number of atoms, which restricts first-principles work to relatively small cells. The authors propose a machine-learning framework for predicting electronic structure across length scales.
On systems small enough for DFT to remain tractable, the framework is up to a thousand times faster; beyond that threshold it still produces predictions, extending electronic-structure modeling to scales that no current first-principles calculation can address. The work frames this as removing a long-standing computational bottleneck, with applications running from semiconductor devices to the interiors of giant planets.